University of Southern California Dissertations and Theses

Essays on monetary policy and international spillovers

(USC Thesis Other)

Essays on monetary policy and international spillovers

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Essays on Monetary Policy and International Spillovers by Rashad Ahmed A Dissertation Presented to the FACULTY OF THE USC GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulllment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (ECONOMICS) May 2021 Copyright 2021 Rashad Ahmed Acknowledgments I'm indebted to my advisors Joshua Aizenman, Hashem Pesaran and David Zeke for their support and patience throughout my doctoral studies. Caroline Betts, Matt Kahn, Jerey Nugent, Romain Ranciere, Alessandro Rebucci, and Jahangir Sultan also provided great encouragement and invaluable guidance throughout the process. I'm grateful to Rodney Ramcharan for several helpful conversations and for sitting as a committee member on my qualifying exam and dissertation defense. I'm also very thankful for the constant support from Irma Alfaro, Alexander Karnazes, Young Miller, Morgan Ponder, and more generally the USC Economics Department administration. Spending a Summer at the Bank for Inter- national Settlements played an essential role in the development of this thesis, where I was fortunate to work closely with Piti Disyatat and Boris Hofmann. I'm grateful to my family for their unconditional love and support. It's helped me push through many challenges faced over the last ve years. My mom, dad, Inaya, nanu, nana, Khalu, Khala, my cousins, Isa. The unwavering support of my friends old and new, made this experience richer both academically and personally. ii Contents Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii Executive Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v 1 Global Flights-to-Safety and Macroeconomic Adjustment in Emerging Mar- kets 1 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Global Flights-to-Safety: A Cross-Market Approach . . . . . . . . . . . . . . 7 1.3 The Impact of Global FTS on Emerging Markets . . . . . . . . . . . . . . . 22 1.4 Cross-Country Heterogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 1.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 2 Monetary Policy Spillovers under Intermediate Exchange Rate Regimes 44 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 2.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 2.3 De-Facto Peg Intensities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 2.4 Trends in Exchange Rate Policy . . . . . . . . . . . . . . . . . . . . . . . . . 55 2.5 Testing the Trilemma: Empirical Strategy . . . . . . . . . . . . . . . . . . . 61 2.6 Baseline Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 2.7 Non-linear Trilemma Trade-os . . . . . . . . . . . . . . . . . . . . . . . . . 75 2.8 What Induces Non-Linear Monetary Spillovers? . . . . . . . . . . . . . . . . 86 2.9 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 2.10 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98 3 Global Demand Spillovers and Financial Stability Near the Zero Lower Bound 105 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 3.2 Estimating Global Demand Pressure from World Commodity Prices . . . . . 111 3.3 Are Global Demand Spillovers Larger Near the ZLB? . . . . . . . . . . . . . 113 3.4 Macroeconomic Adjustment Near the ZLB . . . . . . . . . . . . . . . . . . . 124 3.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 iii Bibliography 136 Appendix 148 A Chapter 1 148 A.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 A.2 Additional Results and Robustness . . . . . . . . . . . . . . . . . . . . . . . 154 A.3 Flight-to-Safety, Excess Risk Sentiment, and Global Demand . . . . . . . . . 170 B Chapter 2 180 B.1 Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181 B.2 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 C Chapter 3 195 C.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 C.2 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 iv Executive Summary This thesis brings together three research papers which investigate empirically the measure- ment and impact of global nancial, commodity, and monetary shocks across advanced and emerging economies. I put particular emphasis on how macroeconomic policies { domestic and international { shape the transmission of these shocks to nancial markets and the real economy. The rst paper studies the international dimension of global nancial shocks known as ` ights-to-safety'. Financial market imperfections point toward large macroeconomic costs associated with ights-to-safety in the absence of policy intervention. I investigate this implication empirically by developing a measure of global ights-to-safety and modeling their impact on emerging markets. Dened as joint tail realizations across developed market risky and safe asset returns, large ights-to-safety map to unexpected tail events and shape future world commodity prices, interest rates and U.S. Dollar uctuations. In emerging markets, a global ight-to-safety induces a sharp rise in sovereign risk and exchange market pressure followed by a protracted drop in economic activity. These eects are substantially larger than those of U.S. monetary policy shocks and domestic nancial shocks. Heterogeneity in adjustment patterns across countries suggest nancial disruption as a key transmission channel but also a role for policy intervention: The impact of ights-to-safety on economic activity is amplied in countries realizing sharper adjustment in nancial conditions, four times larger in emerging markets with U.S. exchange traded funds, and mitigated through `leaning against the wind' with international reserves. The second paper explores whether intermediate exchange rate regimes such as managed oats grant the degree of monetary independence implied by the international Trilemma. Testing the international Trilemma traditionally relies on discretely classied exchange rate regimes. This simplication limits the implications drawn for middle-ground policies like managed oats or basket pegs, and inhibits inference on the empirical shape of the exchange rate stability { monetary autonomy trade-o. To address these issues, this paper proposes a continuous measure of exchange rate exibility for estimating monetary policy spillovers along the entire spectrum of peg intensities. Monetary spillovers generally increase with exchange rate stability, even within middle ground policies, and basket pegs diversify such spillovers. I then estimate the empirical shape of the trade-o using machine learning tech- niques, nding that the relationship between monetary autonomy and exchange rate stability is signicantly non-linear in both advanced economies and emerging markets. Specically, partially targeting the exchange rate translates to disproportionately smaller or larger mon- etary spillovers along middle-ground exchange rate regimes. For emerging markets in par- ticular, active reserves management is a key mechanism associated with these non-linearities. v The third and nal paper is preliminary work revisiting the Zero Lower Bound (ZLB) irrelevance hypothesis, which states that the ZLB constraint on monetary policy does not amplify the macroeconomic adjustment to shocks. This contradicts a wide class of macroe- conomic models predicting that shocks are amplied at the ZLB because the eects of these shocks cannot be oset by lowering interest rates, thereby causing real interest rates to increase rather than decrease { potentially triggering a de ationary spiral. Much of the literature empirically testing this hypothesis has found that the ZLB is irrelevant: that macroeconomic adjustment to recessionary shocks are not amplied near the ZLB. I argue, however, that these studies are subject to the Lucas Critique because the size of a domes- tic shock is endogenously determined in part by the available monetary policy space. To overcome this issue, I evaluate the ZLB irrelevance hypothesis from a multi-country perspec- tive, investigating country-specic adjustment to world shocks across a panel of 17 advanced economies from 1979 to 2019. Contrasting the prevailing literature, preliminary results point toward a rejection of the ZLB irrelevance hypothesis. This work will set the foundation for a full research agenda over the coming years. In addition to the three thesis chapters, I've published several additional research papers during the PhD. Below are short descriptions of each: \Pandemic Shocks and Fiscal-Monetary Policies in the Eurozone: COVID-19 Dominance During January - June 2020", with Joshua Aizenman, Yothin Jinjarak, Sameer Nair-Desai, Weining Xin (forthcoming), Oxford Economic Papers. Abstract: We compare the importance of market factors against that of COVID-19 dynamics and policy responses in explaining Eurozone sovereign spreads. First, we estimate a multifactor model for changes in credit default swap (CDS) spreads over January 2014 - June 2019. Then, we apply a synthetic control-type procedure to extrapolate model-implied changes in the CDS. The factor model does very well over the rest of 2019 but breaks down during the pandemic, especially during March 2020 when there is a large divergence between the actual and model-implied CDS changes. We nd that the March 2020 divergence is well accounted for by COVID-specic risks and associated policies, mortality outcomes, and policy announcements, rather than traditional determinants. Daily CDS widening ceased almost immediately after the ECB announced the PEPP, but the divergence between ac- tual and model-implied changes persisted. This points to COVID-19 Dominance: widening spreads during the pandemic has led to unconventional monetary policies that primarily aim to mitigate short-run fears, temporarily pushing away concerns over scal risk. \In ation and Exchange Rate Targeting Challenges under Fiscal Dominance", with Joshua Aizenman and Yothin Jinjarak (2021), Journal of Macroeconomics. Abstract: Countries have signicantly increased their public-sector borrowing since the Global Financial Crisis. As a consequence, monetary authorities may face pressure to de- viate from their policy targets in ways designed to ease the debt burden. In view of this consideration, we test for greater scal dominance over 2000-2017 under In ation Targeting vi (IT) and non-IT regimes. We nd that evidence of scal dominance varies across coun- tries and debt congurations. Higher ratios of public debt-to-GDP may appear associated with lower policy interest rates in advanced economies. However, a declining natural rate of interest largely explains the pattern of lower rates and higher debt in these countries. The most robust evidence of scal dominance lies among emerging markets under non-IT regimes, composed mostly of exchange rate targeters. For these countries, policy interest rates are non-linearly associated with public debt levels, depending on both the level of hard- currency public debt-to-GDP and the currency composition of public debt. We also show that emerging market economies with greater exchange rate volatility, in ation volatility, and underlying commodity exposure exhibit stronger associations between public debt and policy interest rates. \Commodity Currencies and Causality: Some High-Frequency Evidence" (2020), Economics Letters. Abstract: I investigate the link between economic fundamentals and exchange rate ad- justment to commodity price uctuations. I overcome the traditional issue of simultaneity by exploiting the September 14, 2019 drone attack on two Saudi Arabian reneries as a natural experiment. This unanticipated event caused the largest 1-day global crude oil price shock in over a decade. Using high-frequency exchange rate data for 30 countries, I link the cross- section of exchange rate movements around the event to country-specic fundamentals and currency risk factors. Crude export and import intensities were associated with appreciation (depreciation). Additionally, countries with higher policy interest rates and weaker nancial positions experienced greater currency depreciation while safe haven currencies appreciated, consistent with `risk-o' sentiment triggering carry trades to unwind. Estimated (pre-event) crude oil and VIX betas can also explain the cross-section of exchange rate adjustment, and these currency risk factors are tightly associated with oil-related and nancial fundamentals, respectively. \Meteor Showers and Global Asset Allocation", with Mohammad S. Hasan and Jahangir Sultan (2020), European Journal of Finance. Abstract: Cross-market linkages allow transmission of shocks among markets. Previ- ous measures of such spillovers are based on broader stock market indexes, which cannot identify the industries that are the principal drivers of spillovers and the industries that are most exposed to the spillovers. Using investable equity indexes, we show that basic materials, nancials, industrials, technologies, and telecommunication equity sectors were the primary exporters of volatility from the U.S. and that the magnitude of the spillovers increased primarily during and post-2008 nancial crisis. There is evidence that Canada was most vulnerable to spillovers, while China's exposure was the lowest among the countries in the sample. Based on the minimum variance portfolio optimization, we nd that investing in foreign industries with low exposure to spillovers from the U.S. generates high Sharpe ratios for U.S. portfolio managers, especially during the nancial crisis. \Accounting for Global COVID-19 Diusion Patterns, January-April 2020", with Joshua vii Aizenman, Yothin Jinjarak, Sameer Nair-Desai, Weining Xin (2020), Economics of Disas- ters and Climate Change. Abstract: Key factors in modeling a pandemic and guiding policy-making include mor- tality rates associated with infections; the ability of government policies, medical systems, and society to adapt to the changing dynamics of a pandemic; and institutional and demo- graphic characteristics aecting citizens' perceptions and behavioral responses to stringent policies. This paper traces the cross-country associations between COVID-19 mortality, pol- icy interventions aimed at limiting social contact, and their interactions with institutional and demographic characteristics. We document that, with a lag, more stringent pandemic policies were associated with lower mortality growth rates. The association between stricter pandemic policies and lower future mortality growth is more pronounced in countries with a greater proportion of the elderly population and urban population, greater democratic freedoms, and larger international travel ows. Countries with greater policy stringency in place prior to the rst death realized lower peak mortality rates and exhibited lower du- rations to the rst mortality peak. In contrast, countries with higher initial mobility saw higher peak mortality rates in the rst phase of the pandemic, and countries with a larger elderly population, a greater share of employees in vulnerable occupations, and a higher level of democracy took longer to reach their peak mortalities. Our results suggest that policy interventions are eective at slowing the geometric pattern of mortality growth, reducing the peak mortality, and shortening the duration to the rst peak. We also shed light on the importance of institutional and demographic characteristics in guiding policy-making for future waves of the pandemic. \Regional Heterogeneity and U.S. Presidential Elections", with M. Hashem Pesaran, (Revi- sion Requested), International Journal of Forecasting. Abstract: This paper develops a recursive model of voter turnout and voting outcomes at U.S. county level to investigate the socioeconomic determinants of recent U.S. presidential elections. It is shown that the relationship between many socioeconomic variables and voting outcomes is not uniform across U.S. regions. By allowing for regional heterogeneity and using high-dimensional variable selection algorithms, we can explain and correctly predict the unex- pected 2016 Republican victory. Key factors explaining voting outcomes include incumbency eects, voter turnout, local economic performance, unemployment, poverty, educational at- tainment, house price changes, urban-rural scores, and international competitiveness. Our results corroborate evidence of `short-memory' among voters: economic uctuations realized a few months prior to the election are indeed powerful predictors of voting outcomes as compared to their longer- term analogues. The paper then presents real time forecasts for the 2020 U.S. Presidential Election based on data available at the end of July 2020 which are then updated based on data available as of mid-October. viii Chapter 1 Global Flights-to-Safety and Macroeconomic Adjustment in Emerging Markets 1.1 Introduction Macroeconomic vulnerabilities to sharp swings in global nancial conditions were once more highlighted by the COVID-19 pandemic. Concerns over a global public health crisis left emerging markets indiscriminately exposed, inducing large and volatile capital out ows, currency depreciation, and sharply wider borrowing costs as presented in Figure 1.1. Despite the uniqueness of the pandemic shock, it shares the signatures of many unanticipated left- tail economic events: a ` ight-to-safety' or alternatively, `risk-o'. These refer to abrupt, violent swings across nancial markets in the form of falling risky asset prices and rotation into safe assets associated with aggressive portfolio rebalancing by global investors. Flights- to-safety directly shape the evolution of the global nancial cycle, re ecting both changing risk appetite and expectations over global demand. Flights-to-safety have also increased in 1 severity in the last decade amid an era of unprecedented global liquidity. 1 Figure 1.1: COVID-19, Flight-to-Safety, and Emerging Markets -120 -80 -40 0 0 20 40 60 Business Days USD (In Billions) COVID-19 Taper Tantrum 2008 GFC Non-Resident Portfolio Flows to EMEs 85 90 95 100 Jan Feb Mar Apr May Jun Dec 16 2019 = 100 Asia Latin America Other EME/USD Exchange Rates 300 400 500 600 Jan Feb Mar Apr May Jun Basis Points LC Spread FC Spread Sovereign Bond Spreads for EMEs LHS: COVID-19 (Feb 19 2020), Taper Tantrum (May 22 2013), 2008 GFC-Lehman Bankruptcy (September 15 2008). Center: Lower values imply depreciation vis-a-vis the USD. RHS: Local Currency (LC) and Foreign Currency (FC) Spreads. Data Source: 2020 BIS Annual Economic Report. In this paper, I present a new measure of global shocks intended to capture the intensity of ights-to-safety, dierentiating them from other adverse shocks that shape nancial mar- kets. These ights-to-safety re ect more primitive shifts in risk appetite or global demand, often both. Specically, large shocks are measured as joint tail realizations across risky and safe assets identied through sign restrictions. This way, I distinguish shocks which trigger a ight-to-safety from other adverse shocks which similarly eect global nancial condi- tions but do not induce the same ight-to-safety behavior. I then investigate how global ights-to-safety shape economic dynamics in emerging markets, shedding light on potential transmission mechanisms consistent with the theoretical literature. My proposed methodology to identify ight-to-safety shocks is transparent, easily gen- eralized and exible. Global ights-to-safety are correlated with benchmark measures of nancial conditions such as the VIX index, global realized stock market volatility (Cesa- 1 See Figure 1.3. Note that the 2020 COVID-19 shock at the onset in late February exhibited textbook ight-to-safety features, but by mid-March the indiscriminate selling of both risky and safe assets suggested that it turned to a ight-to-liquidity as it progressed. 2 Bianchi et al. [2020a]) an the global nancial cycle (Miranda-Agrippino and Rey [2020]), yet imperfectly so because they isolate the component of aggregate nancial uctuations driven by shocks that specically trigger a ight-to-safety. These ights-to-safety are informative of future commodity prices, interest rates, in ation expectations and U.S. Dollar uctuations, and map to historically disruptive events. While global ights-to-safety have become a widely studied nancial phenomena, the literature has focused on the nancial market consequences { how asset prices, capital ows, and nancial conditions behave. Meanwhile, there is little evidence linking them to macroeconomic uctuations despite a strong link suggested by the- oretical macro nance models. I show that global ight-to-safety shocks signicantly aect measures of economic activity in both the United States and across emerging markets, and on average, the impact is substantially larger than both the eect of U.S. monetary policy shocks and home-grown domestic nancial shocks. On a country-by-country basis, however, the extent of these eects are highly uneven. By exploiting this heterogeneity, I shed light on multiple channels through which global ight-to-safety shocks drive macroeconomic uctua- tions. Specically, I show that global ight-to-safety shocks transmit through their eect on domestic nancial conditions, are amplied in countries oering U.S. exchange traded funds, and have a substantially weaker impact on economic activity when central banks expend international reserves to `lean against the wind' during such risk-o episodes. These features are supportive of risk-centric macroeconomic models where asset price volatility aects ag- gregate demand through shocks to risk premia or by constraining nancial intermediaries, and macroprudential central bank policy has the ability to moderate such shocks. Ear- lier work includes Bernanke et al. [1999] and Mendoza [2010] in closed and open economy settings, respectively, where nancial frictions amplify the transmission of shocks. Mean- while Caballero and Kamber [2019], Caballero and Simsek [2020c] and Caballero and Simsek [2020a] argue that shocks to risk premia, aecting asset prices, can directly cause demand recessions, rather than acting only as an amplication mechanism. In an international set- ting closely related to this paper, Miranda-Agrippino and Rey [2020], Jeanne and Sandri 3 [2020], and Davis et al. [2020] further show that nancial market imperfections lead to real eects. Nearly all of these models share two features in common: They imply a signicant relationship between nancial market conditions and real activity, along with a buering role for macroeconomic policy. The results corroborate both implications. This paper makes two main contributions to the literature. First, there is little consensus on how to systematically measure ight-to-safety or risk-o phenomena. Both regime based (Beber et al. [2014], Baele et al. [2019]) and intensity based (Datta et al. [2017], Chari et al. [2020]) measures of ight-to-safety or `risk-on/risk-o' have been proposed. Regime-based measures aim to classify periods of extreme safe-risky asset (or currency) price correlations, while intensity based measures provide continuous values which more closely resemble shocks. Other studies rely on o-the-shelf measures of nancial stress like the VIX index (De Bock and de Carvalho Filho [2015a], Caballero and Kamber [2019]). I present a new intensity- based approach to measure global ights-to-safety which starts with the key ingredient many of the prevailing measures share: extreme co-movement between safe and risky asset mar- ket prices. I then incorporate information from multiple markets while emphasizing tail realizations to more sharply identify ights-to-safety. Second, I build a multi-country structural VAR with country specic heterogeneity to investigate the nancial and macroeconomic implications associated with global ights-to- safety. Focusing on emerging markets which tend to take these shocks as exogenous, I provide new evidence on the transmission of ights-to-safety to macroeconomic uctuations. More generally, this relates to the broad literature on using panel VARs to evaluate the impact of external global shocks on emerging markets (Uribe and Yue [2006], Akinci [2013], Shousha [2016], Aizenman et al. [2016], Fernandez et al. [2017], Caballero et al. [2019], Obstfeld et al. [2019b], Cesa-Bianchi et al. [2020a]). Key departures from this literature entail 1) disentangling and focusing on global ights-to-safety specically over broader measures of global nancial stress, 2) while also allowing for country-specic heterogeneity, as in Cesa- Bianchi et al. [2020a] to help shed light on the potential transmission mechanisms driving 4 dierential macroeconomic adjustment. I start by presenting a method to recover a daily index of global ight-to-safety intensity based on nancial market tail realizations and sign restrictions. First, I recover daily as- set price innovations within a asymmetric-GARCH (Generalized Autoregressive Conditional Heteroscedasticity) model of conditional volatility. I take a cross-market approach, applying this procedure across six indices representing major nancial asset classes: equities, volatil- ity, exchange rates, government interest rates, and credit. In a second stage, I aggregate these asset-specic price innovations while imposing a sign restriction such that their daily co-movement satises the covariance structure observed during a ight-to-safety. I dene this as: rising volatility, rising safe asset prices, widening risky credit spreads, and appreciating safe-haven currencies along with depreciating risky assets and risky currencies. This sign-restriction approach implies that global ights-to-safety are disentangled from more general variation in global nancial conditions driven by other types of adverse shocks. Similar conceptually is Jaroci nski and Karadi [2020], where the authors disentangle types of monetary information shocks by considering the co-movement of equity markets with monetary surprises. The overall aim here is to estimate an otherwise unobservable shock using nancial market prices and the co-movement restrictions consistent with ight-to- safety. I show that this measure of global ight-to-safety is signicantly associated with both daily and monthly frequency U.S. Dollar appreciations, and the relationship persists after controlling for uctuations in the VIX index. Moreover, global ights-to-safety are signicantly informative of future movements in world commodity prices, interest rates and in ation expectations. I then model their impact on emerging markets in a multi-country structural VAR. Un- like more traditional panel VAR approaches which assume homogeneous slope coecients and pool information across countries, I allow for country-specic slope heterogeneity, in- corporating interdependencies between emerging markets, while controlling for spillovers from advanced economies, namely the United States. In response to a global ight-to- 5 safety shock, emerging market sovereign spreads sharply widen, exchange market pressure rises (both as currency depreciation and reserves depletion), and a signicant contraction in economic activity follows. On average, industrial production contracts by 0.625 standard deviations, or four percent over an 18-month window following a 1-standard deviation global ight-to-safety shock. These results also hold under impulse responses estimated using lo- cal projection methods instead of a structural VAR, when using an alternative, model-free measure of global ights-to-safety, and when considering variation in ights-to-safety that are uncorrelated with changes in the VIX. The eects are also asymmetric: the impact of positive ight-to-safety shocks, or risk-o shocks are substantially larger than those of neg- ative shocks, or risk-on shocks. The emerging market response to a 1-standard deviation ight-to-safety is also larger in size than the response to a comparably sized U.S. monetary policy shock or a domestic country-specic nancial shock. The heterogeneity admitted by the modeling approach reveals that macroeconomic ad- justment from a global ight-to-safety is far from uniform across countries, and cross-country patterns suggest nancial disruption as a key transmission channel, but also a signicant role for policy intervention { both key implications of the theoretical models with nancial channels. When global ight-to-safety shocks pass through as tighter domestic nancial conditions, the subsequent impact on economic activity is much larger. I also show that the impact of global ight-to-safety on economic activity is signicantly amplied { roughly by a factor of 4 { in countries which have substantial presence in U.S. traded ETFs, even after accounting for nancial openness. This is consistent with specic vulnerabilities arising due to U.S. nancial integration, particularly as Converse et al. [2020] shows through ETFs which amplify the global nancial cycle in emerging markets. Meanwhile, when monetary authorities more aggressively run down international reserves in response to a ight-to-safety, the following economic contraction is much weaker. This policy of leaning against the wind is most eective when the exchange rate is successfully stabilized, supporting reserves accu- 6 mulation and management as a macroprudential policy tool. 2 The overall ndings are consistent with theoretical models pointing toward large macroe- conomic costs associated with global nancial ights-to-safety in the absence of policy in- tervention. Specically, my results suggest a potent nancial channel in the propagation of these shocks to emerging market economies, but also an important role for domestic policies, namely the accumulation and use of international reserves to `lean against the wind' during periods of nancial turmoil. The remainder of the paper is structured as follows: Section 1.2 describes the construc- tion of global ight-to-safety shocks and documents stylized facts. Section 1.3 investigates how global ights-to-safety shape macroeconomic dynamics in emerging markets. Section 1.4 explores the heterogeneity in macroeconomic adjustment across countries to shed light on the transmission mechanism. Section 1.5 concludes. The Online Supplement provides additional details regarding robustness, with Section A.3 specically investigating the role of risk sentiment and global demand components of global ights-to-safety. 1.2 Global Flights-to-Safety: A Cross-Market Approach I estimate an index which captures the intensity of global ights-to-safety by 1) pooling information from key international markets spanning major nancial asset classes and 2) requiring a particular set of co-movements across these markets to be realized. I specically consider six markets due to their international presence: The Wilshire 5000 equity index; 10-year U.S. Treasury yields; 10-year German Bund yields; FX Carry index (long the New Zealand Dollar and Australian Dollar while short the Japanese Yen and Swiss Franc); U.S. corporate high yield spreads; the CBOE VIX index. These indices are considered for two main reasons: For broad international coverage across advanced economies, and for coverage across asset classes. The index, therefore, will have representation from major nancial asset 2 The macroprudential use of international reserves has also been studied in Aizenman and Lee [2007], Jeanne and Ranciere [2011], Dominguez et al. [2012], Ghosh et al. [2016], Jeanne and Sandri [2020], Davis et al. [2020], Ahmed [2020], Ahmed [2021]. 7 markets: Equities, volatility, government bonds, corporate credit, and currencies. The Wilshire 5000 index represents the broad U.S. stock market, while 10-year Treasuries and Bund yields are some of the worlds most recognized safe investments. The FX Carry index captures the relative value of risky, high interest rate, procyclical currencies against safe, low interest rate currencies. The Japanese yen and Swiss Franc act famously as safe havens, appreciating amid turmoil while the Australian and New Zealand Dollar returns tend to be highly procyclical. The U.S. corporate high yield spread re ects the average nancing premium faced by U.S. rms that are rated below investment grade. Finally, the VIX index is a common gauge for global investor risk appetite, uncertainty and demand for insuring equity market risk. It specically measures the option-implied expected forward 1-month volatility of the S&P 500 stock market index. 3 Table 1.1: Cross Asset Flight-to-Safety Behavior Z kd Underlying Asset Class FTS Behavior w k (avg) w k (PCA) Z 1d CBOE VIX Index Volatility + 1/6 0.17 Z 2d Wilshire 5000 Stock Index Equities - 1/6 0.18 Z 3d 10-year U.S. Treasury Yield Government Rates - 1/6 0.18 Z 4d 10-year German Bund Yield Government Rates - 1/6 0.19 Z 5d U.S. High Yield Spread Credit + 1/6 0.16 Z 6d FX Carry* Currencies - 1/6 0.12 *FX Carry is an equally-weighted index long New Zealand Dollar (NZD) and Australian Dollar (AUD) vis-a-vis the the Swiss Franc (CHF) and Japanese Yen (JPY). A measure of ight-to-safety will be estimated by relying on the cross-asset correlations typically observed during global ights-to-safety. The economics of FTS imply global port- folio rebalancing such that risky assets are sold and safe assets bid in the face of rising uncertainty. To capture this ight-to-safety signature, I dene a ight-to-safety or risk-o as a period based on the following sign restrictions over any given trading day: 3 Notice that four of the six benchmark assets are U.S. centric and therefore, I make the implicit assumption that global FTS are largely re ected in U.S. markets, and more generally across advanced economies. Similar interpretations are taken for the VIX when it is used as a gauge of global risk appetite. While this assumption may be reasonable, global economic centers shift over time. My approach is general enough such that one can easily add more nancial benchmarks to the set, (e.g. China) to account for other important or growing economic centers. 8 - Volatility (VIX) rises [ + ] - Equities fall [ { ] - Treasury and Bund yields fall [ { ] - High yield credit spreads rise [ + ] - Carry currencies (AUD, NZD) depreciate against safe currencies (JPY, CHF) [ { ] as depicted in Table 1.1. The precise inverse is dened as risk-on behavior, so the nal FTS index will capture both risk-on and risk-o movements. 1.2.1 Stage 1: Measuring asset market shocks FTS measure rst requires estimating individual asset price shocks before aggregating to the global ight-so-safety index. Denoter kd ;k2f1;:::Kg as the daily return of assetk over day d. The K = 6 assets considered are those mentioned: The VIX, the Wilshire 5000 index, 10-year Treasury yields, 10-year German Bund yields, FX Carry, and U.S. corporate high yield spreads. All returns are in log-dierences, except the two government yields, which are rst-dierences. The global FTS index is constructed as an aggregation of normalized daily innovations across these assets. I dene daily shocks in each asset by comparing the realized return on day d, r kd , to the square root of the conditional variance forecast for day d (i.e. the ex ante conditional volatility), made on day d 1: Z kd = r kd p E d1 [ 2 kd ] : (1.1) This procedure is similar to the approach of conditionally standardizing or devolatizing price returns (Engle [2002] and Pesaran and Pesaran [2010]). A key dierence is that I con- sider the forecasted, or ex ante volatility, while devolatizing traditionally considers realized volatility of the same period as the return, in our case day d. This step serves three impor- tant purposes. First, the volatility of returns vary substantially across assets and over time. 9 Standardizing asset returns by their conditional volatility produces a transformation which admits to comparing across assets classes and accounts for regime changes (i.e. volatility clustering). Second, under the assumption that Z kd follows an i.i.d. distribution (it is, af- ter all, a conditional z-score), the probability that return r kd was unexpected rises injZ kd j. From the econometricians perspective, large values of Z kd are increasingly likely to re ect exogenous price movements in the sense that they were unforeseeable. Third, large values of Z kd are naturally interpreted as tail realizations. While r kd is observed, E d1 [ 2 kd ] is not and must be estimated. To estimate E d1 [ 2 kd ], a model which allows for time-varying volatility must be specied. I assume that asset returns are mean zero with time-varying volatility following a GARCH process (Bollerslev [1986]): r kd = q E d1 [ 2 kd ]Z kd ; Z kd (0; 1); (1.2) where the return sequence is mean zero, and split into a stochastic i.i.d component (Z kd ) and a time-varying volatility component ( kd ). Notice that our estimates of asset-specic shocksZ kd corresponds to the the exogenous component of asset returns under the specied model. I parameterize Z kd as being drawn from a standard normal distribution, hence conditional returns are normally distributed but the unconditional distribution are allowed to be fat-tailed 4 . Specically the conditional variance at timed follows a GJR-GARCH(1,1) process: 5 E d1 [ 2 kd ] =! k + k E d2 [ 2 k;d1 ] + ( k + k I k;d1 )r 2 k;d1 ; where (1.3) 4 One can parameterizeZ kd as being drawn from a Student-T's distribution which allows for both fat tails in conditional and unconditional distributions, and the results are virtually unchanged. 5 See Glosten et al. [1993] for the extension of GARCH to GJR-GARCH. Alternatively one could use another model for time-varying volatility, for example stochastic (latent) volatility models. These typically rely on computationally intensive Bayesian approaches to estimate them and further assumptions on prior distributions and estimation design. Despite their dierences, GARCH, stochastic volatility, and realized volatility models, three workhorse models of time-varying volatility, perform quite similarly. 10 I k;d1 = 8 > > < > > : 0 if r k;d1 > 0 1 if r k;d1 < 0: (1.4) The conditional volatility model under a GJR-GARCH extends the classical GARCH framework by allowing for asymmetric volatility, a well-known stylized fact of nancial asset returns where the conditional variance of an asset is correlated with returns. The expected or ex ante volatility for day d conditional on day d 1 information is computed as: q E d1 [ 2 kd ] = q ! k + k E d2 [ 2 k;d1 ] + ( k + k I k;d1 )r 2 k;d1 : (1.5) Referring back to Equation 1.1, I recover shocks to asset k by dividing its observed real- ization on day d with the ex ante conditional volatility (Equation 1.5). In other words, we simply ask: to what degree was the realized move justied under the prevailing (ex ante) fore- cast distribution? Larger values imply tail realizations, and equivalently returns which are more likely to be unforeseeable and less likely to be generated from the ex ante distribution. With the Z kd for all 6 components estimated, the global daily FTS index (FTS d ) is constructed as the rotated cross-section average on each day d: FTS d = (w 1 Z 1d w 2 Z 2d w 3 Z w 4 Z 4d +w 5 Z 5d w 6 Z 6d )1 d ; X k2K w k = 1; (1.6) where the rotations ensure that positive values of FTS d coincide with ght-to-safety or risk-o, and negative values coincide with risk-on episodes. Hence, the shocks (Z kd ) corre- sponding to the VIX and high-yield credit spreads are added, while the rest are subtracted. I apply equal weights w a = 1=6 but more generally, one can assign arbitrary weights w k across assets. Similarly, an estimate of FTS d can be obtained by taking the rst principal component across asset shocks Z kd . The implicit weights assigned via PCA are reported in Table 1.1 under w k (PCA). In practice, there is very little dierence between estimates of FTS d obtained via PCA or equal weighting. Specically, the FTS d estimated as the cross- 11 section average shares a correlation of over 0.98 with the PCA approach. This is because the cross-section average and 1st principal component asymptotically converge to the same measure under true factor structure (likely in our case with nancial market returns, Wester- lund and Urbain [2015]). The added benet of taking cross-section averages is that it can be calculated each period without requiring information from the entire sample. By contrast, a key advantage of the PCA approach is that it can \self-learn" weights in high-dimensional settings when the set of variables in Z kd becomes large. 1.2.2 State 2: Imposing the ight-to-safety sign restrictions To then identify ight-to-safety shocks, FTS d is multiplied by an indicator 1 d which takes a value of 1 if that day's cross-asset co-movement was consistent with either ight-to- safety/risk-o or risk-on, and 0 otherwise (the ight-to-safety conditions shown in Table 1.1.): 1 d 8 > > > > > > < > > > > > > : 1 iffZ 1d ;Z 5d g>c\fZ 2d ;Z 3d ;Z 4d ;Z 6d g<c `Risk-O' 1 iffZ 1d ;Z 5d g<c\fZ 2d ;Z 3d ;Z 4d ;Z 6d g>c `Risk-On' 0 otherwise: (1.7) This way, I impose the sign restriction condition that all 6 asset returns move in the direction consistent with ight-to-safety, with the size of the move necessarily larger than some threshold c. If asset price movements do not satisfy this joint condition, there is no ight-to-safety, and FTS d = 0. If the set of sign restrictions is satised, the size of FTS d is continuous, and can be positive (`risk-o') or negative (`risk-on'). As a baseline, I set c = 0, meaning a ight-to-safety is identied simply based on sign, regardless of the size of the moves. One issue with this method is that some days may satisfy the FTS condition simply by random chance, and likely realize low values of FTS d , though this becomes increasingly unlikely as the number of sign restrictions increase. Taking a more conservative threshold for 12 c, accounts for both the direction and size of cross-asset moves. Considering this alternative, I also set a threshold of c = 1, meaning that all components must havejZ kd j> 1 on a given day (at least a 1-sigma) and also move in the direction consistent with ight-to-safety to trigger as an FTS. Note also that the thresholdc can be further generalized, setting dierent c for each asset price series Z kd . Moreover, given a particular target outcome variable (e.g. GDP growth), one could estimate a threshold c using maximum likelihood methods (MLE) as in Chudik et al. [2020]. For simplicity and in this particular case because all shock series are standardized I consider the same threshold c across Z kd . Finally, the daily FTS index FTS d can be aggregated to monthly frequency, FTS t : FTS t = D(t) X d=1 FTS d (t); (1.8) where D(t) is the number of days in month t, and FTS d (t) denote daily global ght-to- safety measures corresponding to montht. By summing the daily values ofFTS d , which can be positive (risk-o), negative (risk-on) or zero (non-event), each monthly value ofFTS t can be interpreted as the net of the daily positive and negative daily FTS values. A large positive monthly value of FTS t indicates that month had either/several large global ight-to-safety days (risk-o) relative to risk-on days and days which were neither risk-on or risk-o. Table 1.2: FTS Index: Sensitivity Analysis Correlation with Excluding: daily FTS d monthly FTS t CBOE VIX Index 0.95 0.96 Wilshire 5000 Stock Index 0.97 0.97 10-year U.S. Treasury Yield 0.97 0.96 10-year German Bund Yield 0.92 0.93 U.S. High Yield Spread 0.96 0.95 FX Carry 0.88 0.89 Leave-one-out analysis constructs the FTS d and FTSt indices but only aggregating four of the ve assets, excluding one at a time. Then the correlations are estimated against the full FTS index calculated with all ve assets, to test whether the index is sensitive to leaving any particular asset out of the calculation. Final row excludes two components: high yield spreads and the VIX index. Because only six assets are in the setK which constructs the FTS index, it's important to 13 assess how sensitive the index is to excluding any single asset. I provide results from a leave- one-out analysis as a robustness check in Table 1.2, showing that both the daily and monthly FTS series remains highly correlated with series constructed as an aggregate of ve of the six assets. Re-computing the index while excluding any of the assets maintains a correlation of 0.89 or greater with the monthly FTS index constructed from all six assets, and 0.88 or higher for the daily index. The inclusion of safe assets is also important for distinguishing ights-to-safety from indicators of the Global Financial Cycle, estimated as the common factor from a broad array of risky asset prices (Rey [2015] and Miranda-Agrippino and Rey [2020]) and do not consider safe asset prices. I omit gold from the FTS index because I wish to only consider nancial market assets. There are several additional reasons: First, the price of gold tends to be strongly determined by factors like its nite supply and the real interest rate. Second as a commodity, gold prices are disproportionately aected by global demand forces versus traditional nancial assets, and its market size is dwarfed by the size of other safe asset markets. As a result the allocation of major global investors and intermediaries to gold is disproportionately small in comparison to safe nancial assets. The U.S. Dollar, another safe asset, I also omit from the ight-to-safety index for similar reasons. The value for the Dollar is driven by several factors, being the benchmark trade invoicing currency and also the global reserve currency. I will show, however, that the U.S. dollar signicantly appreciates during a ight-to-safety, consistent with its safe-haven status, and I consider both gold and the USD as outcome variables when estimating the impact of ights-to-safety on world prices. 1.2.3 Global ight-to-safety: properties and stylized facts From January 2000 through August 2019, of the 5,130 days in the sample, 9.6% are consistent with a ight-to-safety or `risk-o', with 9.6% co-moving in a way consistent with `risk-on'. Note that these proportions do not say anything about the size of the moves (recall c = 0). Risk-o days are also particularly special in the sense that asset price moves are signicantly 14 Figure 1.2: Time-Series of Global Flights-to-Safety (FTS t ) 'Risk−Off' 'Risk−On' −2 0 2 Jan 2000 Jan 2005 Jan 2010 Jan 2015 Jan 2020 Global FTS First order auto-correlation = -0.01. Series is normalized to have unit standard deviation. larger { statistically and economically { than usual. For the Wilshire 5000 stock index, the average daily negative return is -0.7%. on a risk-o day, when negative equity returns are accompanied by rising volatility, falling bond yields, rising credit spreads and depreciating risky currencies the average daily Wilshire 5000 return doubles to -1.4%. Similar patterns apply across the other markets. When the VIX index rises, it rises on average 5.3%. On a risk-o day, it rises on average 8.4%. A time-series of monthly FTS shocks is shown in Figure 1.2. Unlike the standard VIX index or changes in the VIX, neither daily nor monthly measures of FTS shocks (FTS d or FTS t ) exhibit signicant serial correlation - an important feature which should be necessary, but not sucient, in a measure of global FTS shocks. The volatility of FTS shocks have also markedly increased since 2007 (Figure 1.3). Each month the realized volatility is computed by taking the standard deviation of daily FTS d shocks within that month. The volatility of FTS shocks after February 2007 is roughly 60% larger than before 2007. 15 Figure 1.3: Realized Monthly Volatility of Daily Global Flight-to-Safety Shocks 0.00 0.25 0.50 0.75 1.00 1.25 Jan 2000 Jan 2005 Jan 2010 Jan 2015 Jan 2020 Realized Monthly Volatility Each month's realized volatility of FTS is computed as the standard deviation of daily values of FTS d for each month. Structural break occurs in February 2007. 1.2.4 FTS and other measures of nancial stress The way FTS shocks are designed, they can be interpreted as a subset of more general global nancial uctuations: Those which are 1) abnormally large and 2) satisfy the ght-to- safety sign restrictions. Figure 1.4 shows that the FTS index (x-axis) are indeed correlated with other measures of nancial stress, but imperfectly so. These imperfect correlations suggest certain overlapping as the indices all respond to shocks that generate ights-to- safety. However, several types of adverse shocks that do not generate a ight-to-safety are still captured as uctuations in broad measures of nancial stress: Monetary policy shocks, stag ationary shocks, liquidity crunches, non-fundamental shocks. 6 Therefore, the FTS index more cleanly separates a specic type of shock which generates a ight-to-safety pattern. This is especially important if we believe that shocks generating distinct patterns across nancial markets bear dierent economic implications and signal. 6 For instance, contractionary monetary policy shocks and liquidity shocks may result in falling equity prices and rising bond yields. A similar co-movement approach to disentangle shocks is used in Jaroci nski and Karadi [2020] who disentangle monetary information shocks using co-movements with equity returns. 16 Global ights-to-safety can explain roughly 22% of the variation in log changes in the VIX (correlation of 0.47). The measureGVOL t is the change in logged global average equity realized volatility in the spirit of Cesa-Bianchi et al. [2020a]. 7 Monthly FTS can explain roughly 14% of the variation in GVOL t (correlation of 0.38). FTS t is also imperfectly correlated with monthly changes in the global nancial cycle indicator of Rey [2015] and Miranda-Agrippino and Rey [2020], GFCY t , though the correlation is stronger than that between FTS t and GVOL t or that between FTS t and changes in the logged VIX index. Roughly 37% of the variation in GFCY t is explained by global ights-to-safety (correlation of -0.61). Particularly interesting is that FTS, composed from just 6 components, exhibits the degree of correlation that it does with the Global Financial Cycle, which is estimated using over 1,000 asset prices series. Figure 1.4: Global Flight-to-Safety Shocks and other measures of Global Financial Stress ρ = 0.47 , p = 2.4e−14 −2 0 2 4 −2 0 2 Flight−to−Safety VIX Index ρ = 0.38 , p = 2.2e−09 −2 −1 0 1 2 3 −2 0 2 Flight−to−Safety Global Realized Volatility ρ = - 0.61 , p < 2.2e−16 −5.0 −2.5 0.0 2.5 −2 0 2 Flight−to−Safety Global Financial Cycle LHS: Monthly changes in logged VIX index on the y-axis. Center: Monthly changes in logged global realized volatility, GVOLt from Cesa-Bianchi et al. [2020a] on the y-axis. RHS: Monthly changes in the Global Financial Cycle, GFCYt from Miranda- Agrippino and Rey [2020] on the y-axis. These correlations weaken further when increasing the threshold to c = 1 which more conservatively identies ight-to-safety episodes as tail shocks. 7 The measure is calculated by rst computing monthly equity realized volatility from daily stock market index returns across 32 countries, and then taking the cross-section average to arrive at a global average realized volatility index. Finally for consistency, the measure is logged and then rst-dierenced. 17 1.2.5 Events underlying the largest ights-to-safety Comparing extreme values of the FTS index shows that it indeed captures global tail risk. Table A.5 provides a list of dates between 2000 and 2020 that, based on the daily measure FTS d , are identied as the largest ights-to-safety. The global nature of these shocks be- come apparent: `Brexit' (2016), `Chinese Correction' (2007), U.S. President Trump political controversies (2017), the Lehman bankruptcy (2008), and the Arab Spring (2011) round out the top ve daily global ights-to-safety. If we included early 2020 in the calculation, January 27 and February 24, 2020, the onset of the COVID-19 global pandemic, would have both scored within the top ten largest FTS d readings since 2000, specically the tenth and fourth largest, respectively. Using a dierent methodology, a similar list is reported in De Bock and de Carvalho Filho [2015b]. Several ight-to-safety episodes agged by FTS d are shared in their list, even with dierent approaches. None of the ten largest global FTS shocks correspond with the largest U.S. stock market crashes. Table A.6 lists the top 10 largest daily stock market percent declines between the same period Most of the largest stock market crashes occurred during the 2008 Global Financial Crisis, and another the popping 2000 Tech Bubble. Table A.7 shows the top 10 largest percent changes in the VIX index { three overlap with the top 10 daily largest FTS shocks. The largest VIX shock re ects the `Volmageddon' (2018), considered by many practitioners as a technical, non-fundamental event caused by overcrowded short volatility positions, highlighting the potential for non- fundamental movements in nancial stress indicators. 1.2.6 Global ghts-to-safety, the U.S. Dollar, and world prices Table 1.3 reports daily and monthly regressions of U.S. Dollar log returns on its own lagged returns, changes in the VIX index, and the FTS index. I consider the Dollar vis-a-vis an equally weighted basket of advanced (G10) and emerging market (EM) economies separately. The results are consistent with the Dollar's role as a global safe asset. When including the VIX and excluding the FTS index (columns 1, 3, 5, 7), USD appreciation is signicantly 18 Table 1.3: Global Flights-to-Safety and U.S. Dollar Appreciation Daily Returns Monthly Returns G10/USD EM/USD G10/USD EM/USD (1) (2) (3) (4) (5) (6) (7) (8) Intercept 0.001 0.001 0.008 0.008 0.021 0.021 0.115 0.119 (0.006) (0.006) (0.005) (0.005) (0.114) (0.113) (0.077) (0.076) Lagged USD 0.011 0.012 0.061 0.053 0.363 0.344 0.403 0.386 (0.016) (0.016) (0.025) (0.025) (0.059) (0.061) (0.067) (0.070) lnVIX 0.003 0.0005 0.015 0.007 0.023 0.015 0.040 0.034 (0.001) (0.001) (0.001) (0.001) (0.009) (0.009) (0.007) (0.008) FTS 0.034 0.091 0.278 0.243 (0.009) (0.007) (0.118) (0.092) Observations 5,129 5,129 5,129 5,129 234 234 234 234 R 2 0.003 0.007 0.081 0.127 0.171 0.187 0..366 0.386 Adjusted R 2 0.002 0.006 0.081 0.126 0.164 0.176 0.361 0.378 Robust standard errors with *,**,*** corresponding to 10, 5, and 1 percent signicance, respectively. USD returns are computed as log-changes from the previous period. G10 index is the USD return vis-a-vis an equal-weighted basket of currencies of: New Zealand, E.U., United Kingdom, Australia, Switzerland, Sweden, Norway, Denmark, Japan, Canada. EM index is the USD return vis-a-vis an equal weighted basket of currencies of: South Korea, Mexico, Brazil, India, Malaysia, South Africa, Taiwan, Thailand, Sri Lanka. FTS is normalized to unit variance, while other variables are in percentages. 19 associated with positive innovations in the VIX. However, the FTS index is signicantly and even more powerfully associated with Dollar appreciations. While both the VIX and FTS are highly signicant for the EM/USD exchange rate, the VIX loses it's explanatory power once the FTS is introduced in the G10 equations. This result implies that jumps in the VIX alone are not suciently capturing conditions that warrant a ight to the Dollar. Rather, G10 Dollar appreciations are only associated with the VIX when the VIX rises amid a ight-to-safety. To explore the responses to a global ight-to-safety across a broad spectrum of world prices, I estimate a second-order structural vector auto-regression (SVAR) of monthly log- dierences of U.S. short and medium term yields, USD exchange rates, commodities, and U.S. in ation expectations where FTS shocks, FTS t , are identied recursively: FTS shocks are ordered rst, such that they impact all other variables contemporaneously, consistent with the exogenous nature of unusual or unexpected events which trigger ights-to-safety. Because we model the response to FTS shocks, the ordering of the remaining variables does not matter. Figure 1.5 traces the impulse responses of a 1-SD FTS shock on a variety of commodity prices, gold, and the USD exchange rate vis-a-vis the G10. The solid line is the response to a 1-SD FTS shock, FTS t , with shaded areas indicating 90% bootstrapped condence bands. Figures A.2 and A.3 provide additional IRFs for U.S. interest rates, in ation expectations and additional commodity prices. Most responses exhibit signicant adjustment for several months following an FTS shock, U.S. yields and market-based in ation expectations fall along the entire maturity curve. Commodity prices fall and the U.S. Dollar appreciates in response to an FTS shock, both in time 0 and subsequent months. The response of commodities is sharp across metals and energy. The impact of FTS shocks are also apparent in some soft commodities (Figure A.3) like soybeans, one of the largest Chinese imports. These results further suggest that ights- to-safety are not pure shocks to risk aversion, rather there is an important change in global demand (i.e. physical risk) that occurs. The eect on gold is statistically indierent from 20 Figure 1.5: Response to a 1-Standard Deviation FTS Shock −0.2 0.0 0.2 0.4 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Gold −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Silver −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations USD/G10 −0.9 −0.6 −0.3 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Commodities −0.75 −0.50 −0.25 0.00 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Crude Oil −1.00 −0.75 −0.50 −0.25 0.00 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Copper Cumulative response (in standard deviations) to a 1-standard deviation structural ight-to-safety (FTS) shock,FTS t . 90% bootstrapped condence bands. Negative change in USD/G10 is U.S. Dollar appreciation. zero and the impact on silver is signicant but relatively small. This may be somewhat surprising given that some view precious metals as safe havens. However the dierential impact on the U.S. Dollar and gold highlights the nature of gold being a safe asset but also a commodity with industrial use. Because FTS also indicate dropping global demand, the demand eect osets the risk premia eect on gold prices, resulting in the null average response. By contrast, the U.S. Dollar appreciates when faced with both adverse global demand or risk aversion. I show this in Section A.3, where I attempt to separate the excess risk sentiment and global demand components embedded in global FTS shocks, subject to 21 a number of assumptions. 1.2.7 Robustness The responses of world prices to a global ight-to-safety are robust. Figure A.4 shows that uctuations inFTS t that are uncorrelated with changes in the VIX index still have signiant information content, highlighting the special nature of ights-to-safety beyond aggregate uctuations in the VIX index. Figure A.5 sets the FTS condition threshold to c = 1, so not only do all 6 assets need to move in specic directions, but they must all move in excess of 1 standard deviation, emphasizing tail events. Finally Figure A.6 orders the FTS shock in the SVAR last, allowing it to only impact world prices with a lag rather than contemporaneously. The benchmark results broadly hold under this setup as well. For additional robustness, I also propose a model-free estimator of global FTS shocks in Section A.2 of the Online Appendix. This simple approach identies FTS shocks as changes in the log VIX index on days which satisfy the ight-to-safety condition mentioned previously. These daily VIX changes amid risk-on/risk-o are then summed to a monthly aggregate FTS series, which turns out to be highly correlated with the baseline FTS shock series, FTS t . 1.3 The Impact of Global FTS on Emerging Markets Recent debate and research focuses the consequences of global nancial shocks on emerging markets (EMs), many of which are left particularly vulnerable from growing nancial inte- gration. I revisit this issue, specically to evaluating the dynamics of emerging markets in response to a global ight-to-safety shock. I collect monthly data on sovereign spreads and industrial production across 34 emerging markets from 2000 to 2019. 8 I build on several recent studies have investigated the global transmission of world nancial shocks on EM dy- 8 Data details are found in Section A.1. 22 namics (Uribe and Yue [2006], Akinci [2013], Caballero et al. [2019], Kalemli-Ozcan [2019a], Cesa-Bianchi et al. [2020a], Obstfeld et al. [2019b]). The traditional modeling approach used is a panel regression or VAR which estimates average eects and impulse response functions (IRF) to a global shock by pooling information across all countries. While pooling has the advantage of increasing statistical power, it ignores vital heterogeneity across countries, which surely exists among EMs. A key dierence in my modeling approach is that I allow for country-specic heterogeneity, following an approach similar to Cesa-Bianchi et al. [2020a]. I further show that this heterogeneity can be used to shed light on potential transmission mechanisms through which global shocks transmit to the real economy. In view of this consideration, I propose a heterogeneous multi-country VAR which com- bines elements from the large VAR literature (Global Vector Autoregressive (GVAR) Pesaran et al. [2004] and Factor-augmented Vector Autoregressive (FAVAR) Bernanke et al. [2005]). Like the benchmark panel VAR, it can be used to report average eects by pooling results across countries. However, like Fernandez et al. [2017] and Cesa-Bianchi et al. [2020a], my approach builds on previous analyses by also allowing for country-specic heterogeneity. Key modeling challenges of multi-country economic systems include accounting for 1) global com- mon factors 2) network eects or spillovers between countries 3) spillovers from advanced countries to emerging markets, and 4) heterogeneous transmission of shocks. Consider the baseline model which incorporates these features: 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 s i;t y i;t S i 0 ;t Y i 0 ;t Y US;t FTS t 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 = 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 s i y i S i Y i US i V i 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 + i (L) 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 s i;t1 y i;t1 S i 0 ;t1 Y i 0 ;t1 Y US;t1 FTS t1 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 + 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 u s i;t u y i;t u S i 0 ;t u Y i 0 ;t u US i;t v t 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 ; (1.9) 23 where s i;t is the change in the log sovereign spread { a proxy for domestic nancial conditions { of country i over month t. Country i's year-over-year change in industrial production (IP) in montht is given by y i;t . It's easy to see that a model with just these two variables represents a classic VAR(L) model. Country-specic lag polynomials are expressed as i (L) of nite order `. I set the number of lags equal to ` = 4 months. The specication is extended by modeling cross-country linkages through S i 0 ;t and Y i 0 ;t . These are cross- section averages of changes in the log sovereign spread and year-over-year IP growth over all countries excluding country i. Specically, S i 0 ;t = s i 0 ;t = X i 0 6=i w s i 0s i 0 ;t ; N1 X i 0 =1 w s i 0 = 1; Y i 0 ;t = y i 0 ;t = X i 0 6=i w y i 0 y i 0 ;t ; N1 X i 0 =1 w y i 0 = 1; where s i 0 ;t is a weighted average of the spread change for countries not including i, s i 0 ;t , weighted by w s i 0. I set equal weights (w s i 0 = 1=(N 1) for all i 0 ), therefore s i 0 t can be interpreted as the cross-section average of sovereign spread changes, exclusive of country i. The same is done for Y i;t , except I exclude Iraq from the calculations given large outlier values driven by the Iraq War in the early 2000's. Other approachs to obtaining weights would be to apply GDP weights, bilateral trade-weights, capital ow weights, or estimateing them via PCA forw s i 0. 9 However, in this particular setting, because cross-country correlations are high, these alternatives make no practical dierence. Including these global averages admit for cross-country interdependencies without run- ning into the `Curse of Dimensionality' issue most large VARs face (hence, also admitting to a GVAR interpretation). For example, S i 0 ;t and Y i 0 ;t can be thought of as the inclusion of lagged spreads and IP growth for all other countries in the equations for countryi. Without any coecient restriction, estimating a VAR(4) would entail the addition of 3342 = 264 9 I test both and the factor estimated via averages and that via PCA are highly correlated, close to a coecient of 1. 24 additional lagged variables, exceeding the number of observations. However, including cross- sectional averages imply a coecient restriction such that lag l spreads and IP growth from all other countries in countryi's equation have coecients equal to i (L) 1 N1 . I also include Y US;t changes in U.S. economic activity, measured using the Chicago Fed National Activity Index (CFNAI) to account for spillovers between advanced economies and emerging markets. Finally, FTS shocks FTS t enter the system as a common variable across all countries to which countries respond dierentially (as re ected in the country-specic coecients V i ), and the shock is identied recursively. That is, FTS t can be viewed as a common factor that unlikeS i 0 ;t and Y i 0 ;t is completely external to the system. Recall that FTS t is mea- sured from nancial variables either based out of the U.S. or advanced economies, while the endogenous variables in Equation 1.9 belong to emerging markets except for Y US;t . 1.3.1 Estimating the multi-country SVAR and impulse responses The shock FTS t is structural, in that it is identied under the recursive assumption that FTS t contemporaneously aects fast-moving nancial variables s i;t and S i 0 ;t , while slower- moving macroeconomic variables y i;t ,Y i 0 ;t andY US;t respond to FTS shocks with a 1-month lag. It is, after all highly plausible that a global nancial shock passes through to countryi's nancial conditions while an idiosyncratic shock to countryi does not trigger a global ight- to-safety { so long as countryi is not a dominant country in the economy. 10 Within a sample of emerging markets this assumption is reasonably satised. The recursive assumption re- lated to y i;t ,Y i 0 ;t andY US;t requires the FTS shock variableFTS t to be contemporaneously orthogonalized against the three slow-moving economic activity variables. The results of the impulse response analysis are robust to alternative ordering restrictions, specically one such that FTS t contemporaneously aects all other variables but no other variable contempora- 10 An excellent example corroborating this assumption is the case of Chile in 2019, suering from increasing political unrest and protest. While these events disrupted Chile's domestic nancial and economic condi- tions, it did not trigger a reaction across global nancial markets. By contrast, a few months later, panic over COVID-19 induced a global nancial market shock which severely impacted Chile among many other countries in an indiscriminate fashion. 25 neously aects FTS t . The large T dimension of the data allows the multi-country SVAR to be estimated country-by-country, estimating country-specic SVARs for 34 emerging markets. This es- timation procedure is akin to estimating a Global VAR (Pesaran et al. [2004], Chudik and Pesaran [2016]) with similar approaches also being applied in Fernandez et al. [2017] and Cesa-Bianchi et al. [2020a]. The heterogeneous modeling approach still allows estimation of average or pooled eects as done in traditional panel models. Estimating the average IRF over the panel is simple using the Mean Group (MG) estimator of Pesaran and Smith [1995] and Chudik and Pesaran [2019]. 11 Following Cesa-Bianchi et al. [2020a], the horizonh mean group, or average, impulse response function for the endogenous variable, denoted X it , to a 1-SD FTS shock is computed as: MGIRF (h) = 1 N N X i=1 E[X i;t+h jv t = 1;! t1 ] 1 N N X i=1 E[X i;t+h jv t = 0;! t1 ] = 1 N N X i=1 E[X i;t+h jv t = 1;! t1 ]; (1.10) where E[X i;t+h jv t = 1;! t1 ] is the horizon h impulse response of country i, denoted as the conditional expectation of X i;t+h given a 1-SD structural FTS shock (v t = 1), and ! t1 denotes the full information set available as of time t 1. Intuitively, the impulse response function of Equation 3.10 examines howX i;t+h responds to a 1-standard deviation FTS shock at time t given the information available at time t 1, comparing it to a counterfactual scenario of no FTS shock (v t = 0) at time t with the same information available at time t 1. The associated non-parametric cross-sectional standard errors computed as: SE(h) = v u u t 1 N 1 N 1 N X i=1 E[X i;t+h jv t = 1;! t1 ]MGIRF (h) 2 : (1.11) 11 Alternatively, the Common Correlated Eects Estimator (CCE) of Pesaran [2006] and Chudik and Pesaran [2015] can also be applied. 26 It can be easily seen that the MG IRF is simply the cross-section average of alli country- specic IRFs, each being denoted E[X i;t+h jv t = 1;! t1 ], at each horizon h. 95% dispersion intervals for each horizon h which I report in the results are equal to MGIRF (h) 1:96SE(h): (1.12) These methods have been applied successfully to large, heterogeneous macroeconomic panel data of similar size to address a variety of research questions. 12 1.3.2 The average response to a global ight-to-safety shock I rst estimate the model with the global FTS shock, FTS t and examine the average dy- namics of economic activity and sovereign risk across EMs. Figure 1.6 traces the average, or MG estimate impulse response of both logged sovereign spreads and IP growth to a 1-standard deviation FTS t shock. Sovereign spreads react strongly and the response is front-loaded, displaying over-shooting behavior in the rst few months following the shock. Economic activity signicantly contracts over about 18 months. All units are measured in standard deviations to correct for heteroscedasticity across coun- tries. For the sake of interpretation, the 18-month cumulative response in IP growth is ap- proximately equivalent to a 4% contraction. For comparison I also show that U.S. economic activity (thin solid line) signicantly contracts with a lag following an FTS shock, re ecting their global nature. The total U.S. contraction and recovery occurs faster and more sharply. The dashed line indicates the response of economic activity to a 1-standard deviation id- iosyncratic country spread shock. As a proxy for country-specic nancial shocks, the results indicate that global shocks are much more potent than their local counterparts. Both the response in sovereign spreads and the subsequent contraction in IP growth remain signicant after orthogonalizing the FTS shock against changes in the VIX index, 12 See for example Dees et al. [2007], Chudik et al. [2017], Hernandez-Vega [2019], Cesa-Bianchi et al. [2020a]. 27 Figure 1.6: Emerging Markets: Average Response to a 1-Standard Deviation FTS Shock (Solid), Response to a Country-Specic Sovereign Spread Shock (Dashed) 0.0 0.2 0.4 0.6 0 10 20 30 Months Standard Deviations Log Sovereign Spread Country−Specific Financial Shock U.S. −0.75 −0.50 −0.25 0.00 0 10 20 30 Months Standard Deviations Industrial Production Cumulative MG Response (Equation 3.10) to a 1-standard deviation structural ight-to-safety shock, FTSt (solid), and after controlling for contemporaneous VIX innovations (dashed). 95% non-parametric dispersion bands as computed in Equation 3.12. Log sovereign spread in monthly changes. Industrial production as year-over-year log change. Thin line in RHS gure is the IRF of U.S. national economic activity. suggesting a distinct role for FTS shocks in shaping macroeconomic dynamics (Figure A.7. Figure A.8 shows that these results are robust to an FTS index identied under a more conservative ight-to-safety condition of c = 1, where both direction of asset price moves and also size are taken into account. 1.3.3 Comparing ights-to-safety and U.S. monetary policy shocks To quantify the relative importance of global FTS shocks to other global nancial shocks, I compare emerging market dynamics following a contractionary U.S. monetary policy shock. Both ights-to-safety and U.S. monetary policy lead to tighter global nancial conditions, but the two shocks fundamentally dier and are uncorrelated. Monetary policy shocks are unanticipated changes in the U.S. policy rate while FTS shocks re ect unpredictable 28 economic tail events or news. I use high-frequency (30-minute) changes in the 3-month Treasury futures contract around FOMC announcements to capture U.S. monetary shocks following Kuttner [2001] and Gertler and Karadi [2015]. These are then aggregated to the monthly frequency. The correlation between FTS shocks and U.S. monetary policy shocks is statistically indistinguishable from zero (correlation of 0.10), and as expected, there is no lead-lag relationship between the two shock series. Figure 1.7 reports MG IRFs, after replacing the FTS shock series with U.S. monetary policy shocks in Equation 1.9. Figure 1.7: Emerging Markets: Average Response to a 1-Standard Deviation Contractionary U.S. Monetary Policy Shock U.S. Monetary Policy Shock 0.0 0.2 0.4 0.6 0 10 20 30 Months Standard Deviations Log Sovereign Spread U.S. Monetary Policy Shock −0.75 −0.50 −0.25 0.00 0 10 20 30 Months Standard Deviations Industrial Production Cumulative MG Response (Equation 3.10) to a 1-standard deviation contractionary U.S. monetary policy shock. Monetary shocks are recovered from prices changes in the 3-month treasury futures contract within a 30-minute window of FOMC announcements. 95% non-parametric dispersion bands as computed in Equation 3.12. Log sovereign spread in monthly changes. Industrial production as year-over-year log change. Negative values imply exchange rate percent depreciation. International reserves in monthly log changes. In response to a 1-standard deviation contractionary U.S. monetary shock, sovereign spreads widen and economic activity contracts { much like a ight-to-safety shock. A 29 1-standard deviation ight to safety elicits a response in sovereign spreads and economic activity roughly three and two times that of 1-standard deviation monetary policy shock, respectively. The standard deviation of the monetary policy shock series is equal to 3.54 basis points, while the standard deviation of changes in overall 3-month treasury yields over a similar period is 17.92 basis points. If we were to scale the monetary policy shocks up to have a similar standard deviation as the nominal 3-month yield, then the impact of a 1-standard deviation FTS shock on emerging markets would be roughly equal to the eects from a contraction of U.S. monetary policy between +36 and +48 basis points. This could alternatively be interpreted as the extent of monetary accommodation required to oset the impact of the average global ight-to-safety on emerging markets. 1.3.4 Incorporating exchange market pressure Exchange market pressure (EMP), introduced early on in Girton and Roper [1977] along with its many variants (Hossfeld and Pramor [2018]), is a useful gauge of international pressure on the exchange rate either resisted through foreign exchange intervention or relieved through currency depreciation. EMP severity tends to capture periods of large, volatile capital in ows or out ows - often straining exchange rates and nancial liquidity. Many recent studies highlight the role of global shocks in driving pressure on international markets via exchange or capital ow pressures across EMs. 13 To consider the implications of EMP in the presence of global ights-to-safety, I augment the multi-country VAR with two additional country-specic endogenous variables: logged changes in USD exchange rates and international reserves. Global ight-to-safety shocks likely bear implications for EMP and its interaction with economic activity. Currency mis- match, for example, is a mechanism through which EMP may impact the real economy, as exchange rate depreciation increases the cost of foreign-denominated liabilities (Eichengreen and Hausmann [1999], Hofmann et al. [2019], Carstens and Shin [2019].). Known as the - 13 Fratzscher [2012], Aizenman and Binici [2016], Goldberg and Krogstrup [2018]. 30 Figure 1.8: Emerging Markets: Average Response to a 1-Standard Deviation FTS Shock (Solid) and After Controlling for Changes in Logged VIX (Dashed) 0.0 0.2 0.4 0.6 0 10 20 30 Months Standard Deviations Sovereign Spread −0.75 −0.50 −0.25 0.00 0 10 20 30 Months Standard Deviations Industrial Production −1.5 −1.0 −0.5 0.0 0 10 20 30 Months % Change Exchange Rate −2 −1 0 0 10 20 30 Months % Change International Reserves Cumulative MG Response (Equation 3.10) to a 1-standard deviation structural ight-to-safety shock, FTSt before (solid) and after controlling for contemporaneous VIX innovations (dashed). 95% non-parametric dispersion bands as computed in Equation 3.12. Log sovereign spread in monthly changes. Industrial production as year-over-year log change. Negative values imply exchange rate percent depreciation. International reserves in monthly log changes. nancial channel of exchange rates, the pecuniary externality caused by currency depreciation in the presence of currency mismatch osets the classical trade channel where depreciation is considered stimulative. For this reason I focus on USD exchange rates given the recent evidence on the overwhelming role of the U.S. Dollar in the international monetary and price system. 14 Figure 1.8 traces the Mean Group IRF from a 1-SD FTS shock from the model including exchange rates and international reserves. In addition to sovereign spreads widening and economic activity contracting, there is signicant exchange market pressure across emerging markets. EMP manifests as both currencies rapidly depreciating against the USD and sig- nicant running down of international reserves. Within the rst few months, exchange rates depreciate on average of 1.1%. After 10 months, reserves growth drops almost 1.5%. Both of these eects remain signicant after orthogonalizing FTS shocks against the VIX index (Figure A.7) and when setting the FTS condition threshold to c = 1. In section A.2 of the Online Supplement I estimate the baseline results shown in Figure 14 The majority of trade is invoiced in USD, most countries peg to the USD, most international reserves are held in USD, most international nancing is denominated in USD. 31 1.8 using the Local Projection method of Jord a [2005] rather than the SVAR approach as a robustness check. The results remain consistent and signicant regardless of modeling procedure. I also explore asymmetries, where positive (risk-o) and negative (risk-on) FTS shocks are allowed to impact emerging market dynamics dierently, nding that after allow- ing for asymmetries, risk-o (positive FTS) shocks have substantially larger absolute eects compared to risk-on (negative FTS) shocks. This implies that the IRFs from the baseline symmetric VAR understate the macroeconomic and nancial impact of adverse ights-to- safety. 1.4 Cross-Country Heterogeneity Emerging Markets, on average, are subject to signicant adjustments in response to a global ight-to-safety yet these eects may vary widely at the individual country level. An issue worth exploring then is whether these cross-country heterogeneities are large, and system- atically linked. Specically, nancial market responses (sovereign spreads, exchange rates, international reserves) are relatively immediate compared to the adjustment of macroeco- nomic activity. This section uses cross-country heterogeneity to explore particular trans- mission channels moderating the impact of global FTS shocks on emerging market business cycles. 1.4.1 Shedding light on the nancial transmission channels Explicit identication of transmission channels at the international macro level remains a challenge. Generally speaking, there are two main approaches. The rst is to develop a structural model while the second is a reduced form approach. An example of the reduced form approach is taken in Akinci [2013] when attempting to quantify whether or not global nancial shocks transmit to the real economy through their eect on domestic nancial con- ditions. A basic counterfactual exercise is done by comparing the variance decomposition 32 of a nancial shock to real economic activity under the baseline VAR, to the same variance decomposition after shutting down eect of global nancial shocks on domestic sovereign spreads (i.e. setting the coecients in the sovereign spread equations associated with global nancial shocks equal to zero). The results suggest that indeed, global shocks are amplied through their eect on sovereign spreads. However, the author also notes that this coun- terfactual exercise is subject to the Lucas Critique, as it is questionable whether all other coecients characterizing the system would in fact stay constant when setting one particular coecient to zero. Figure 1.9: Heterogeneous Impact of Global FTS Shocks: 6-month Change in Sovereign Spreads (LHS), USD Exchange Rates (Center), International Reserves (RHS) vs. 18-Month Change in Economic Activity ρ = - 0.52 , p = 0.0017 Argentina Belarus Brazil Chile China Colombia Cote d Ivoire Croatia Ecuador Egypt El Salvador Hungary Iraq Jordan Kazakhstan Lithuania Malaysia Pakistan Peru Russia Senegal South Africa Tunisia Turkey Venezuela −2 −1 0 1 0.0 0.2 0.4 0.6 0.8 t+6 Change in Log Spread in SDs t+18 Change in Log IP in SDs ρ = 0.28 , p = 0.1 Argentina Belarus Brazil Chile China Colombia Cote d Ivoire Croatia Ecuador Egypt El Salvador Gabon Indonesia Iraq Kazakhstan Malaysia Peru Poland Russia Senegal Sri Lanka Tunisia Turkey Ukraine Uruguay Venezuela Vietnam −2 −1 0 1 −3 −2 −1 0 t+6 % Change in USD Exchange Rate t+18 Change in Log IP in SDs ρ = - 0.4 , p = 0.02 Argentina Brazil Chile Colombia Cote d Ivoire Croatia Ecuador Egypt El Salvador Indonesia Iraq Kazakhstan Lithuania Malaysia Mexico Pakistan Peru Russia Senegal Tunisia Ukraine Uruguay Vietnam −2 −1 0 1 −0.04 −0.02 0.00 0.02 t+6 % Change in Log Reserves t+18 Change in Log IP in SDs Cumulative Responses (in standard deviations) to a 1-standard deviation structural ight-to-safety (FTS) shock, FTSt (Equa- tion 1.9). Given the heterogeneity provided by my modeling approach, I extend upon the approach of Akinci [2013] by exploiting cross-country dierences to infer potential transmission chan- nels. By comparing countries with dierential responses to FTS shocks, we can shed light on transmission mechanisms without imposing such controversial restrictions on the counterfac- tual estimation. For example, I investigate whether the impact of FTS shocks on economic activity is signicantly stronger for the subset of countries where FTS shocks also impact severely sovereign spreads. Figure 1.9 LHS shows across the 34 countries in the panel, the 6-month cumulative change 33 in the log sovereign spread against the 18-month cumulative change in industrial production induced by a 1-SD FTS shock. The LHS correlation coecient equals -0.52 and is statistically signicant. Countries which realize wider short-run sovereign spread adjustment in response to an FTS shock are subject to deeper long-run economic contractions. While Akinci [2013] nds that transmission of global nancial shocks through country spreads account for two- thirds of the impact on macroeconomic activity, I nd that the role of tightened country spreads explain closer to 27% of the variation in macroeconomic adjustment. Similarly, the center gure shows that countries which experience greater currency depre- ciation vis-a-vis the USD amid an FTS shock also realize larger subsequent IP contractions. By contrast,the RHS gure shows that countries which more aggressively expend reserves also realize shallower subsequent contractions in industrial production (correlation equal to -0.40). Taken together, these associations suggest that the impact of FTS shocks on the real economy are partly determined by the sensitivity of domestic nancial factors along with the intensity of policy responses. 1.4.2 U.S. Integration: Do ETFs amplify the impact of Flight-to- Safety shocks? The extent to which FTS shocks eventually impact economic activity in emerging markets suggestively depend on the sensitivity to domestic nancial factors { response of domestic nancial conditions and the policy response with international reserves { as shown. Addi- tionally, nancial openness or integration with advanced economies may be a critical factor which also shapes the business cycle response to foreign nancial shocks. In this context, the advent of exchange-traded funds (ETFs) in advanced economies has been of growing inter- est, giving global investors considerable access EM investments with the promise of superior liquidity. With them comes the potential for much greater capital ow volatility. In recent work, Converse et al. [2020] document that equity and bond ETF ows are signicantly more sensitive to global nancial conditions than mutual fund ows, amplyng the global 34 nancial cycle in emerging markets. Figure 1.10: Distribution of the number of ETFs traded on U.S. exchanges each EM has presence within 0 5 10 15 0 200 400 600 Number of ETFs Frequency ρ = - 0.53 , p = 0.0013 Argentina Belarus Brazil Chile China Colombia Cote d Ivoire Croatia Ecuador Egypt El Salvador Indonesia Iraq Kazakhstan Malaysia Mexico Pakistan Peru Philippines Russia Senegal Tunisia Ukraine Uruguay Vietnam −2 −1 0 1 0 2 4 6 Ln(number of U.S. ETFs +1) t+18 Change in Log IP in SDs LHS: Frequency distribution of the number of U.S. ETFs a country has presence within (as of October 2020). Source: etfdb.com. RHS: x-axis plots the ln(number of U.S. ETFs +1) against the 18-month cumulative IP growth response to a 1-SD FTS shock. To capture the role of nancial integration with the U.S. through ETFs, I investigate whether the impact of FTS shocks dier systematically in countries which have either equity or bond ETFs available for trade on U.S. exchanges compared to those which do not (or have very few). The former countries, by virtue of selection, are likely to have more advanced nancial markets and more open capital accounts { not just with the United States. Greater nancial development implies that these countries enjoy lower nancing costs on average. At the same time, these countries may be particularly sensitive to ight-to-safety shocks and associated sudden capital out ows as global investors withdraw capital from emerging markets, deemed relatively risky investments. Meanwhile, Converse et al. [2020] argue that ETFs may attract dierent investors than mutual funds, specically those which put greater value on liquidity, and do not put as much value on local fundamentals when allocating capital. This ETF-specic channel can amplify the impact of external shocks even conditional on nancial openness. Table A.4 provides the number of U.S. traded ETFs granting exposure to each country in the sample as of October 2020. Brazil, China, Mexico and South Africa 35 each have more than 200 U.S. traded ETFs which at least some nancial assets based in those countries. By contrast, several countries have little or no investment through U.S. ETF holdings: Belarus, Cote d'Ivoire, Croatia, Ecuador, Vietnam, among others. A clear demarcation is observed between Ukraine, which a U.S. investor can gain exposure through 7 ETFs and the next country Pakistan, for which the number of ETFs jump to 47. Figure 1.10 shows the frequency distribution (LHS) of countries by number of U.S. based ETFs. Roughly half of the countries have little or no ETF presence in the United States. On the RHS, the relationship between the logged number of ETFs per country on the x-axis and the response of IP growth to a FTS shock is plotted. It's quite clear from a cursory look that economic contractions induced bye global FTS shocks are deeper in countries with greater presence among U.S. ETFs. Figure 1.11: Average Response to a 1-Standard Deviation FTS Shock for Countries with U.S. ETF presence (dashed) and those without (solid) 0.0 0.2 0.4 0.6 0 10 20 30 Months Standard Deviations Sovereign Spread −1.0 −0.5 0.0 0 10 20 30 Months Standard Deviations Industrial Production −1.5 −1.0 −0.5 0.0 0 10 20 30 Months % Change Exchange Rate −4 −3 −2 −1 0 0 10 20 30 Months % Change International Reserves Cumulative MG response (Equation 3.10) to a 1-standard deviation structural ight-to-safety shock, FTSt. 95% dispersion bands as computed in Equation 3.12. Log sovereign spread in monthly changes. Industrial production as year-over-year log change. Negative exchange rate response equals percent depreciation against the USD. International Reserves as log monthly change. Countries with U.S. ETF presence: Argentina, Brazil, Chile, China, Colombia, Egypt, Hungary, Indonesia, Malaysia, Mexico, Pakistan, Philippines, Poland, Russia South Africa, Turkey. Figure 1.11 traces the IRFs to a 1-SD FTS shock for two dierent groups of EMs. The dashed line refers to countries with a substantial presence in the U.S. ETF space (Argentina, Brazil, Chile, China, Colombia, Egypt, Hungary, Indonesia, Malaysia, Mexico, Pakistan, Philippines, Poland, Russia, South Africa, and Turkey). The solid line is the MG IRF 36 for countries with little to no ETF presence (Belarus, Cote d'Ivoire, Croatia, Ecuador, El Salvador, Gabon, Iraq, Jordan, Kazakhstan, Lithuania, Peru, Senegal, Sri Lanka, Tunisia, Ukraine, Uruguay, Venezuela, and Vietnam). The minimum number of ETFs available among the countries with substantial presence is 47 (Pakistan) and the max is China (571). The minimum for the group with low ETF presence is zero (Belarus, Cote d'Ivoire, Ecuador, Gabon, Lithuania, Senegal, Tunisia, Uruguay, Venezuela) and the maximum is Ukraine with 7 ETFs. Despite similar responses in sovereign spreads to a global FTS shock, The group of coun- tries with heavy ETF presence are subject to signicantly deeper { roughly four times deeper { economic contractions than the group without U.S. ETF presence. While both groups of countries experience heavy exchange market pressure following a global FTS shock, the groups dier by whether the pressure is relieved through currency depreciation or expending reserves. Countries with heavy ETF presence realize relatively sharper currency depreciation while expending relatively less international reserves, with the reverse holding for the group without an ETF presence. Controlling for Financial Openness Countries with signicant ETF presence on developed market exchanges may simply be more nancial integrated and developed with other countries other than the U.S., and thereby more sensitive to global nancial shocks. To test whether the ETF dierential is soley proxying for broad nancial openness, I take the 16 EMs with high ETF presence and re- sort these countries into those with high versus low capital in ow controls. Data on capital in ow controls are taken from the Fern andez et al. [2016] data set, and I arrive at country- specic values by averaging values of aggregate capital in ow control index from 2000-2019 for each country. 15 Splitting the ETF EMs into two equal-sized groups, the countries with ETF presence but high or above-median capital controls are: China, Colombia, Indonesia, 15 Of the 34 EMs in the sample, 27 have capital control data available. 37 Figure 1.12: Average Response to a 1-Standard Deviation FTS Shock for Countries with U.S. ETF presence, sorted into High (Dashed) and Low (Solid) Capital In ow Controls 0.0 0.2 0.4 0.6 0 10 20 30 Months Standard Deviations Sovereign Spread −1.5 −1.0 −0.5 0.0 0 10 20 30 Months Standard Deviations Industrial Production −2.0 −1.5 −1.0 −0.5 0.0 0 10 20 30 Months % Change Exchange Rate −4 −3 −2 −1 0 1 0 10 20 30 Months % Change International Reserves Cumulative MG response (Equation 3.10) to a 1-standard deviation structural ight-to-safety shock, FTSt. 95% dispersion bands as computed in Equation 3.12. Log sovereign spread in monthly changes. Industrial production as year-over-year log change. Negative exchange rate response equals percent depreciation against the USD. International Reserves as log monthly change. Countries with U.S. ETF presence and low capital controls: Argentina, Brazil, Chile, Egypt, Hungary, Poland, South Africa, Turkey. Countries with U.S. ETF presence and high capital controls: China, Colombia, Indonesia, Malaysia, Mexico, Pakistan, Philippines, Russia. Malaysia, Mexico, Pakistan, Philippines, Russia. The countries with ETF presence but low or below-median capital in ow controls are: Argentina, Brazil, Chile, Egypt, Hungary, Poland, South Africa, Turkey. The idea here is that if general nancial openness matters, then there should be a dierential impact of FTS shocks observed based on nancial openness even after conditioning on high ETF intensity. Figure 1.12 traces MG impulse response functions for these two groups of ETF-intensive EMs. It becomes immediately clear that nancial openness is generally not driving the strong impact of global FTS shocks on EM economic activity with high ETF presence. The response of sovereign spreads and industrial production between these two sub groups are statistically indistinguishable. Although note that this doesn't mean that capital controls are ineective: Those countries with stricter capital ow controls do experience less exchange market pressure (both as less exchange depreciation and reserves depletion) amid global ights-to-safety. This evidence corroborates Converse et al. [2020] in that the growth of ETFs poses a potential amplication mechanism for the transmission of global shocks, and the channel may not simply be due to nancial openness, but the unique integration with 38 the U.S. or liquidity-preferring behavior of ETF investors. 1.4.3 Explaining cross-country macroeconomic adjustment to ights to safety Taking stock of the systematic heterogeneity between domestic nancial factors and the transmission of global shocks, The evidence presented thus far suggests that a global FTS shock can induce deeper subsequent contractions in industrial production when the early response in sovereign spreads are sharper, when the exchange rate depreciates more, or when there is greater U.S. ETF presence. Meanwhile, actively running down international reserves in response to an FTS shock is associated with a buering eect on economic activity. These domestic nancial factors may interact with each other, or possibly explain the same underlying nancial exposure. To analyze the joint in uence of these nancial factors on the long-run impact of FTS shocks on economic activity, I propose the following cross-sectional regression: E i [y i;t;t+18 jFTS t ] = + 1 E i [s i;t;t+6 jFTS t ] + 2 E i [fx i;t;t+6 jFTS t ]+ 3 E i [res i;t;t+6 jFTS t ] + 4 ln(ETF i + 1) + 5 X i +e i ; (1.13) where the dependent variableE i [y i;t;t+18 jFTS t ] is countryi's cumulative response in IP growth to a 1-SD FTS shock after 18 months. E i [s i;t;t+6 jFTS t ], E i [fx i;t;t+6 jFTS t ], and E i [res i;t;t+6 jFTS t ] are the 6-month cumulative response of country i's sovereign spread, USD exchange rate, and international reserves to a 1-SD FTS shock, respectively. Finally ETF i is the number of U.S. traded ETF's countryi maintains a presence within andX i refers to additional controls. Specically, I include an indicator denoting whether the country is a commodity exporter to capture economic composition. 16 Standard errors are robust to het- 16 Commodity exporter is dened as in Aslam et al. [2016], a country with greater than 35% of exports in 39 eroscedasticity. Both the dependent variable and the independent variables are estimates, thus subject to measurement error. In the case of uncorrelated measurement error, atten- uation will bias the coecients estimated by least squares towards zero. Therefore, a most plausible scenario is one where the standard errors are biased upwards and the point esti- mates are biased downwards, so estimated associations from Equation 1.13 would understate the true association strength. Table A.8 reports the regression results from estimating Equation 1.13. Deeper sub- sequent IP growth contractions are associated with countries which initially realize wider sovereign spreads or currency depreciation in response to a FTS shock. Countries which expend more reserves as a buer against an FTS shock realize economic contractions which are comparatively smaller. Moreover, having a larger presence in the U.S. ETF investable space is associated with deeper economic contractions following ights-to-safety, and this relationship is signicant and robust. To consider the interaction of international reserves and exchange rate movements which together characterize total exchange market pressure, I include the interaction term, E i [fx i;t;t+6 jFTS t ]E i [res i;t;t+6 jFTS t ]; which is abbreviated in the table for succinctness. The interaction term is highly signi- cant and negative, while the marginal eect of exchange rate depreciation is insignicant,and the marginal eect of expending reserves remains highly signicant. Therefore a possible in- terpretation of the three estimates is that expending reserves (i.e. leaning against the wind) buers against the real economic impact following a global FTS shock, and this eect weak- ens with greater coincident exchange rate depreciation. In other words, following a global FTS shock, the buering eects of expending reserves on subsequent economic growth ap- pears most eective when the exchange rate is successfully stabilized. Column 6 shows that commodities and greater than 5% of all trade in commodities, on average over 1960-2014. These countries are: Argentina, Brazil, Chile, Colombia, Cote d' Ivoire, Ecuador, Gabon, Indonesia, Iraq, Kazakhstan, Malaysia, Peru, Russia, Uruguay, Venezuela. 40 results persist after controlling for commodity intensity to capture dierences in economic structure across countries. Taken together, these domestic nancial factors explain up to 60% of the cross-country variance in the macroeconomic sensitivity to a global FTS shock. Moreover, the results sug- gest ample evidence that these nancial heterogeneities are not simply confounded with one another, rather they explain distinct cross-country variation in the macroeconomic adjust- ment to FTS shocks. Tables A.9 and A.10 replicate the regression results after replacing FTS shocks with the VIX or Global Financial Cycle in the multi-country VAR (Equation 1.9) for robustness. The signicance of most nancial factors disappear, and the explanatory power of the regression drops substantially (adjustedR 2 falls from 60% to between 40%-50%). One possible reason for this is that FTS shocks are more cleanly identifying a specic nancial shock which transmits to real economic activity via the reported margins of heterogeneity. Fluctuations in the global nancial cycle and the VIX index of course account for shocks that generate ights-to-safety, but also many other types of adverse shocks. If these other shocks transmit via other margins of heterogeneity, we'd see a similar sort of attenuation in the results as we moved from FTS shocks to the VIX or global nancial cycle variable. All three measures of global nancial stress highlight a signicant role of ETF intensity and exchange market pressure in explaining cross-country macroeconomic adjustment to a global FTS shock. 1.5 Concluding Remarks This paper presents a new measure of global nancial shocks specically re ecting ight- to-safety to test their impact on domestic nancial and economic conditions in emerging markets. The largest daily FTS shocks do not correspond with the largest stock market crashes nor a majority of the largest jumps in the VIX index. Flight-to-safety shocks do map to economically disruptive historical events, informing current and future changes in 41 interest rates, exchange rates, commodities, in ation expectations, the U.S. Dollar, and contain both components re ecting shifting risk sentiment and global demand. In Section A.3 of the Online Supplement, I further investigate the separation of FTS shocks into excess risk sentiment and global demand components. I investigate how global FTS shocks shape macroeconomic dynamics in the U.S. and a panel of 34 emerging markets within a multi-country VAR framework. In response to a global FTS shock, sovereign spreads widen dramatically, exchange market pressure increases and economic activity subsequently contracts in both emerging markets and the U.S. over a period of 18 months. These eects persist when using variation in FTS shocks that is uncorrelated with the VIX index. I further show that there is signicant country-specic heterogeneity in the impact of FTS shocks across EMs. Countries realizing sharper adjustment in their sovereign spreads and greater currency depreciation are subject to deeper subsequent economic contractions. Meanwhile, countries which aggressively expend international reserves, leaning against the wind in response to an FTS shock, are subject to smaller subsequent economic contractions, especially when the exchange rate is successfully stabilized. Moreover, the impact of FTS shocks on economic growth is signicantly amplied among countries with substantial pres- ence within U.S. traded ETFs. These features are supportive of a broad range of risk-centric macroeconomic models where shocks to risk premia propagate through the real economy, and policy intervention can mitigate these eects. The role of domestic nancial factors moderating the pass-through of global shocks to local economic conditions coincides with the ndings of Akinci [2013], Aizenman et al. [2016] and Kalemli-Ozcan [2019a] and recent risk-centric theoretical frameworks of Caballero and Simsek [2020b], Caballero and Simsek [2020a], Jeanne and Sandri [2020] and Davis et al. [2020]. Along the international dimension, the buering eects of running down interna- tional reserves suggest an important role for monetary policies to serve as macroprudential policy-puts, buering against external tail shocks in a nancially integrated world. The am- 42 plication mechanism of global shocks through highly volatile investment ows, particularly through ETFs and nancial integration with the U.S., also warrants further research given the rapidly expanding footprint of the industry. 43 Chapter 2 Monetary Policy Spillovers under Intermediate Exchange Rate Regimes 2.1 Introduction The international policy Trilemma Mundell [1963] states that no country can meet all three objectives: Independent monetary policy, free capital ows, and exchange rate stability. The importance of these implications has grown sharply amid the onset of rapid nancial globalization, remaining an enduring topic of discussion among academics and policymakers alike. However, research on the policy Trilemma almost exclusively focuses on the eects of corner policy choices (e.g., exchange rates are either considered xed or oating, capital accounts are either open or closed) because of the challenges associated with constructing continuous measures of Trilemma policy variables. Despite the substantial presence of inter- mediate exchange rate regimes around the world, we know relatively little of the implications of middle-ground policy choices on monetary autonomy. This study aims to address this gap in the literature. A growing body of evidence suggests that the Trilemma generally holds in the short and long-run: conditional on open capital ows, international transmission of monetary policy 44 from base countries tend to be stronger under xed exchange rates than under oating (Frankel et al. [2004], Shambaugh [2004], Obstfeld et al. [2005], Miniane and Rogers [2007], Klein and Shambaugh [2015], Herwartz and Roestel [2017], Eichengreen [2018], Han and Wei [2018]). 1 Typical estimates of monetary pass-through suggest that transmission is incomplete (i.e. less than 1-for-1), and less complete in emerging markets, with the unanticipated component of base country monetary policy changes exhibiting greater pass-through rates (Bluedorn and Bowdler [2010]). While the literature on international monetary spillovers under the policy Trilemma is highly active and growing, most empirical studies resort to categorizing exchange rate regimes in a binary fashion (xed or oating) due to various challenges, including data limitations and the practical diculties associated with classifying exchange rate regimes. Frankel et al. [2004] and Klein and Shambaugh [2015] break this trend by studying monetary autonomy while considering intermediate exchange rate regimes as a class of their own. Both studies nd that intermediate regimes buy some monetary autonomy relative to xed exchange rates. While oering several important contributions, these studies are limited in terms of allowing for heterogeneity within intermediate exchange rate regimes. 2 Given the wide spectrum of intermediate peg intensities, this may be an overly restrictive classication. Specically, whether monetary policy spillovers are linearly, or non-linearly related to exchange rate policy remains an open question requiring greater detail on peg exibility within the class of intermediate exchange rate regimes. In this paper, I depart from the literature by introducing an exchange rate regime measure which is fully continuous. My particular approach brings with it three distinct advantages. First, it relaxes the constraint that all intermediate exchange rate regimes are identical. 1 In contrast, a number of studies debate that the Trilemma has broken down to a `Dilemma', rendering exchange rate policy irrelevant for monetary independence due to several reasons related to nancial glob- alization (Calvo and Reinhart [2002], Frankel et al. [2004], Rey [2015], Miranda-Agrippino and Rey [2020], Georgiadis and Zhu [2019]). However, Klein and Shambaugh [2015] and Han and Wei [2018] specically con- sider these factors and still nd that monetary policy pass-through to foreign interest rates is signicantly stronger (weaker) under xed ( oating) exchange rate regimes. 2 Though importantly, Frankel et al. [2004] do dierentiate between bands and managed oats, two regimes falling under the intermediate classication. 45 Second, It allows one to investigate the open question of whether monetary policy transmis- sion under the Trilemma is linear in exchange rate exibility, as typically assumed. If it is not, what are the policy implications? What mechanisms may be generating an empirical non-linearity? These are important issues that I attempt to address. Third, this approach allows for testing monetary spillovers under basket pegs, which itself remains unexplored in the empirical Trilemma literature. Continuous exchange rate regime measures themselves are not new. A separate yet related line of research aims to study the Trilemma conguration using continuous policy measures. Aizenman et al. [2010], Aizenman et al. [2013], and Ito and Kawai [2014] investigate the Trilemma middle-ground under a continuous policy setting, but rather than focusing on monetary policy spillovers, they focus on macroeconomic outcomes and determinants of such middle-ground policy congurations (Aizenman and Ito [2014], Jord a et al. [2015], Frankel et al. [2019] and Obstfeld et al. [2019a]). Studies combining the two approaches { testing monetary policy spillovers under con- tinuous measures of exchange rate exibility { are few and far apart. One closely related paper, Herwartz and Roestel [2017], studies monetary pass-through in such a fashion among a sample of advanced economies, documenting a nearly linear trade-o between exchange rate stability and monetary autonomy. I build on this issue, diering from the previous study in several ways. First, I consider a larger panel of countries across both advanced economies and emerging markets. Second, I introduce a dierent continuous, de facto mea- sure of exchange rate regime by drawing on the literature related to estimating currency zones. 3 I estimate non-overlapping, quarterly de facto peg intensities vis-a-vis three candi- date base currencies using daily exchange rate returns. The method is exible enough to allow for multiple exchange rate targets, allowing for spillover tests under basket peg poli- cies. By contrast, Herwartz and Roestel [2017] rely on the exchange stability index proposed in Aizenman et al. [2008], which is a transformation of the annual standard deviation of 3 Haldane and Hall [1991] and Frankel and Wei [1992]. 46 monthly exchange rate changes. By using higher-frequency, daily exchange rate data my approach provides more consistent estimates of quarterly de facto exchange rate variability. I then go a step further in attempting to identify the underlying mechanisms which may lead to a non-linear relationship between exchange rate exibility and monetary autonomy, namely exchange market intervention via international reserves, and international limits to arbitrage. The main contributions of this paper are three-fold. First, under a new continuous exchange rate regime measure, I conrm prevailing evidence of existing monetary policy spillovers within the context of the international Trilemma. Second, I document new evi- dence suggesting that monetary policy spillovers can be diversied under basket pegs. Third, I test the linearity of the Trilemma through leveraging both standard econometric methods and more recent machine learning models such as Generalized Additive Models (GAMs). In both sets of tests, I identify the eects of foreign monetary policy shocks on domestic monetary policy using the instrumental variables (IV) approach of Jord a et al. [2015] and Jord a et al. [2020]. Both the standard econometric and GAM specications point to a signif- icant non-linear relationship between exchange rate exibility and monetary independence along intermediate exchange rate regimes: greater exchange rate stabilization translates to disproportionately smaller or larger losses in monetary autonomy along certain parts of the peg intensity spectrum. This contrasts Herwartz and Roestel [2017], who nd a near linear relationship between exchange rate stability and monetary autonomy. Moreover, net `gains' in monetary autonomy are allocated dierently across advanced economies and emerging markets. Advanced economies tend to put greater emphasis on output stabilization while emerging markets focus on in ation. Among emerging markets, active reserves management appears to be a plausible mechanism generating these empirical non-linearities. These nd- ings are robust to a variety of sensitivity tests, including: testing for short-run and long-run monetary spillovers; accounting for the zero lower bound; alternative exchange rate regime classications; using exogenous U.S. monetary policy shocks around FOMC events; omitting 47 the 2008 Global Financial Crisis period; changes in the SDR basket components. These results also bear implications for the Two-Corners Hypothesis which gained pop- ularity after the late 90's early 2000's chain of nancial crises experienced across the world. The argument is that middle ground exchange rate regimes are unstable and crisis prone, therefore exchange rate policy should converge to either xed or oating (Frankel et al. [2000]). However, empirically this hypothesis has been continuously rejected, as middle- ground exchange rate policies are alive and well (Fischer [2001], Masson [2001], Williamson [2002], Frankel [2019], Frankel et al. [2019]). Most of the world follows an intermediate exchange rate regime. As of 2018, 46.6% of the 189 IMF member countries report adminis- tering intermediate pegs - up from 40% in 2010. 4 In addition, extensive empirical evidence suggests that many of the world's oating exchange rates are actually managed oats - i.e., intermediate pegs of varying exibility. Calvo and Reinhart [2002] and Ilzetzki et al. [2019] both highlight the systematic `Fear of Floating' exhibited by exchange rates of countries which presumably claim to oat, despite pervasive contradicting evidence. My ndings sup- port this view such that across countries and over time, a substantial proportion of countries in the sample appear to partially target the exchange rate. The rest of the paper is structured as follows: Section 2.2 brie y goes over the data. Section 2.3 discusses measurement and estimation of continuous de facto exchange rate regimes. Section 2.4 goes on to discuss notable trends and statistics in de facto exchange rate regimes across countries over the last two decades. Section 2.5 covers the baseline empirical strategy for analyzing monetary policy transmission under the policy Trilemma. Section 2.6 then goes over baseline results. Section 2.7 pays particular focus on testing for potential non-linear monetary policy spillovers under intermediate exchange rate regimes and Section 2.8 then explores potential underlying mechanisms which may generate these non-linearities. Section 2.9 covers a battery of robustness checks and Section 2.10 concludes. 4 Source: IMF Annual Report of Exchange Arrangements and Exchange Restrictions (AREAER) for the year 2018 48 2.2 Data I consider a panel composed of 46 countries which does not include the U.S. and E.U. over the period Q1 2000 to Q4 2018 (quarterly frequency). 5 12 are Advanced Economies and 34 are Emerging Markets. The list of countries are reported in Table B.1 in the Appendix. 6 The data was collected from multiple sources. Quarterly central bank policy interest rates are taken from the BIS and IMF IFS databases. Additional data on interest rates were collected from individual central bank websites and Global Financial Data. When ocial central bank policy rates could not be used, short-term treasury bills, repos, or discount rates are used. The use of short-term rates ensures that proper testing of the Trilemma, based on UIP, can be conducted such that maturities broadly match across countries. In ation and CPI data is primarily drawn from the BIS, IMF IFS, and the World Bank. For country-quarter observations where data was not available, annual in ation rates (divided by four) were used for imputation. In ation is year-over-year. Nominal GDP data is from the IMF IFS database. Growth rates are computed as year-over-year. Missing observations were imputed using annual frequency growth rates from the World Bank. Daily exchange rate data is taken from the BIS and are used to estimate de-facto exchange rate peg intensity. Moreover, daily log returns are aggregated to the quarterly frequency, and combined with in ation data to recover quarterly real exchange rate returns. A positive change in the real exchange rate corresponds to local depreciation. Daily commodity price data for gold, copper, crude oil, coee and sugar are taken from Bloomberg. Specically, I rely on front month futures contract prices. Data on daily and quarterly CBOE VIX index values, a common gauge for global risk appetite, are from FRED. Annual capital controls measures are taken from the Chinn-Ito index (Chinn and Ito [2006]) derived from the IMF AREAER, and repeated over each quarter within the year. 5 The country choice is subject to data coverage. The data is taken from all publicly available sources. After cleaning and merging data from various sources, 46 countries in total have sucient sized samples to conduct the analysis. 6 Select Tables and Figures are moved to the Appendix for brevity. Table an gure numbers labeled with `A' refer to those in the Appendix. 49 For Serbia, capital control measurements are taken from the Wang-Jahan index, which is also derived from the IMF AREAR index. Remaining missing values for Serbia are extrapolated (2000-2004, and 2014-2018). Since the index is updated through 2017, I extrapolate 2017 values to 2018. Developed and Emerging/Developing Economy classications are taken from IMF WEO (2019). Data on foreign exchange reserves are taken from the IMF International Reserves and Foreign Currency Liquidity database. International reserves are measured as the sum of total foreign currency reserves, IMF reserve positions and SDRs. Gold holdings are excluded from calculation. For robustness, additional tests are run using alternative denitions of exchange rate regime. Specically, I use the Ilzetzki et al. [2019] data set on de-facto exchange rate regimes and anchor currencies, which has 14 classes of exibility which I consolidate into a smaller set. IRR exchange rate regime only thorough Q 4 2016. I take quarterly averages of monthly exchange rate regimes. Fed Fund Futures data are taken from Bloomberg. First contract month yield changes are computed over the day of a scheduled FOMC meeting. Daily monetary policy shocks are then aggregated to the quarterly level (simple sum). In the process of cleaning the data, I remove country-quarter observations which are deemed outliers based on: Interest rate changes greater than 5 percentage points in absolute value, interest rate levels greater than 50%, and in ation greater than 40%. 7 2.3 De-Facto Peg Intensities A key limitation across studies on the policy Trilemma is the coarse classication of exchange rate regimes. Most studies resort to a binary (or at best, discrete) splitting of observations into either ` oating' or `xed' exchange rate regimes. While this is an important consideration when focusing on the corner congurations of the policy Trilemma, little can be said about the monetary autonomy trade-o under more complex exchange rate targeting policies, such as an intermediate peg or basket peg. Moreover, intermediate exchange rate regimes are 7 Comparable to Ilzetzki et al. [2019]. 50 not all equal: policymakers choose the degree of exibility which potentially gives way to a spectrum of exchange rate regimes (peg intensities) which vary both across countries and over time. As a parsimonious solution for estimating a continuous measure of the de-facto exchange rate regime, I follow and extend the methodology introduced in Haldane and Hall [1991], Frankel and Wei [1992], and later on in Benassy-Quere et al. [2006]. This regression-based technique estimates continuous `peg intensities' that are directly associated with a base currency. 8 The rst-step here is to estimate non-overlapping de-facto peg intensities at the quarterly frequency. These estimates, which characterize country's exchange rate regime, can then be applied in the main analysis testing for monetary policy transmission. I extend the methodology along two dimensions. First, I rely on higher frequency (daily) data to estimate non-overlapping, lower frequency (quarterly) peg intensities. This contrasts the traditional approach of estimating peg intensities on an overlapping or rolling basis. Second, I control for global common factors and shocks which may impact exchange rate uctuations both in the country of interest and the base country { specically world commodity prices and global investor risk aversion. Like Haldane and Hall [1991] I use daily exchange rate data which yields a sucient number of observations for consistent quarterly peg intensity estimates. However at the daily frequency the issue of asynchronous trading hours across international exchange rate markets might pollute the regression analysis. One solution would be to use weekly exchange rates (Frankel and Wei [1992] and McCauley and Chan [2014]), but the number of observations to estimate quarterly peg intensities would drastically drop. To overcome the issue of potential non-overlapping trading hours while preserving the number of observations, I compute 2-day rolling average exchange rate returns following Forbes and Rigobon [2002] and Wang et al. [2017]. Then over each quarter, I estimate the following regression with daily data: 8 Variants of this methodology have been recently implemented in McCauley and Chan [2014], Ito and Kawai [2016] and Ito and McCauley [2019] to study cross-country patterns in trade invoicing currencies, global imbalances and the composition of central bank foreign reserves. Frankel et al. [2019] consider con- tinuous de facto exchange rate regimes to study their eects on economic growth. 51 e i d (t) = i (t) +W e it e e d (t) +W U it e U d (t) +W $ it e $ d (t) + i d (t); (2.1) where e i d (t) is the day d (of month t) change in the log exchange rate of country i vis-a-vis the IMF's Special Drawing Rights currency basket (SDR) and base currencies on the RHS denoted e b d , b2fe;U;$g, are the Euro, Japanese Yen, and U.S. Dollar vis- a-vis the SDR, respectively. I choose these three currencies as the possible set of base currencies because of their disproportionately large role in international trade and nance. The U.S. Dollar and the Euro together make up the large majority of: base currency pegs, international reserves holdings, external debt currency denomination, and trade invoicing currency globally. 9 Furthermore, following the literature, the specication implicitly assumes that these three base currencies are de facto pure oaters, making up the potential candidate target currencies for all other countries. Note that the question of which numeraire to use is discussed extensively in the literature as it aects the interpretation of the error term when the currency does not follow a perfect hard peg. 10 To circumvent this issue, I follow Frankel [1993] and Ma and McCauley [2011] by considering SDRs as the numeraire. Meanwhile, other solutions have been proposed: Frankel et al. [2001] use a basket of currencies { not unlike the SDR { and Frankel [1993] use consumer price indices as the numeraire. 11 Another proposed solution which does not consider a basket-type numeraire but still attempts to deal with the collinearity of exchange rates induced by triangular arbitrage is to simply use the USD as a numeraire, but have the regressions explicitly omit the USD exchange rate from the RHS. For example, Ito and McCauley [2016] and Ito and McCauley [2019] denominate exchange rate returns in USD, 9 See Gopinath [2015], Maggiori et al. [2019], Goldberg and Lerman [2019] and the recent ECB note (ECB [2019]). 10 Additionally, if the numeraire moves closely in line with one of the candidate base currencies, then that base currency will have very small variance and may be confused with the constant term (Benassy-Quere et al. [2006]). 11 I do not consider using price indices as the numeraire because price index data is not available at the daily frequency. One could alternatively consider trade-weighted eective exchange rates as a solution to the numeraire problem (though results are likely to remain similar as the SDR and trade-weighted exchange rate returns are highly correlated). 52 but on the right hand side include base country currencies but not the USD. Then, to estimate the weight on the USD, the authors take the dierence between 1 and the sum of the estimated weights on the other base currencies. The advantage of this approach is that it simplies the problem of choosing an appropriate numeraire. Meanwhile, a potential drawback is that the weight on the USD base is restricted such that the weights across all base currencies necessarily sum to 1. Ma and McCauley [2011] further demonstrate that the results from Frankel and Wei [1993] are robust to using either the SDR or the U.S. Dollar as the numeraire. Equation 2.1 implies that the movements of each currencyi are decomposed to a weighted average of the base currencies plus an idiosyncratic error term. These weights translate to peg intensities against base currencies. For example, with a currency that pegs perfectly to the U.S. Dollar (e.g. Ecuador, which has been Dollarized since 2000),W $ it would equal 1 and the other weights would equal zero. In contrast, a purely oating exchange rate would have weights statistically indierent from zero across all three base currencies, and an exchange rate which targets a basket (e.g. Singapore) would have non-zero weights on multiple base currencies. Therefore, the strength of the peg is given by a value between 0 and 1, where 0 is no weight ( oat), and a 1 is interpreted as a hard peg to the base currency. This way we arrive at a continuous measure of peg intensity for each country, for each quarter, through exploiting currency movements at the daily frequency. An important note to emphasize is that a peg intensity estimate equal to 1 does not necessarily imply pegging, especially if the estimated regression results in a poor model t, which would most likely coincide with statistical insignicance. To correct for such scenar- ios, I follow the algorithm of Ito and McCauley [2019] to clean peg intensity estimates. 12 . 12 To clean and remove spurious results when estimating Equation 2.1: before estimating Equation 2.1, I rst omit observations of daily log exchange rate changes exceeding 5% in absolute value to prevent crisis- related outliers from in uencing peg intensity estimates (as similarly done in Ilzetzki et al. [2019] who remove in ation observations exceeding 40% in their analysis). Then, after estimating Equation 2.1, any statistically signicant negative coecient estimates of the peg intensities (W b it ) is set to be a missing value (large negative weights are theoretically inconsistent). Statistically insignicant negative values are set to zero (because a weight of zero is not rejected in this case). Values statistically signicantly greater than one are taken to be missing values (positive values exceeding one are theoretically inconsistent), and values insignicantly 53 Additionally Figure B.1 and Table 2.1 report the distributional characteristics of R 2 across all country-quarter observations where a strong peg is estimated (i.e. there is a ^ W b it = 1). Table 2.1: Summary statistics of R 2 from all country-quarter regressions where ^ W b it = 1 Statistic N Mean St. Dev. Min Pctl(25) Pctl(50) Pctl(75) Max R 2 1,634 0.740 0.309 0.04 0.48 0.92 1 1 Immediately notice the very high median R 2 of 0.92 and that the majority of values lie between 0.48 and 1, validating that most of the identied country-quarters under strong pegs in fact bear appropriately high model ts, thereby further conrming the reliability of the rst-stage results. 2.3.1 Controlling for common shocks A potential issue with the standard estimation of Equation 2.1 is that it doesn't recognize the role of global factors or common shocks which may in uence jointly country i's and base country b's exchange rate, thus generating what may appear as large or sudden shifts in exchange rate policy if not controlled for. For example, common factors may include uctuations in global commodity prices. Through driving variation in the terms-of-trade, commodities are known to in uence exchange rates of resource-dependent economies. Ex- change rates exhibiting such behavior are often dubbed `commodity currencies' (chin Chen and Rogo [2003], Ahmed [2020], Beckmann et al. [2020], among several others). In addition to commodities, global investor risk appetite appears to play an increasingly potent role in driving broad currency risk (Avdjiev et al. [2019]). Periods of high risk aver- sion tend to coincide with episodes of Dollar and Yen appreciation as they are viewed as global safe assets. At the same time, risk aversion drives risky asset prices lower, which may greater than 1 are set to 1 (becuase a weight of 1 cannot be rejected in this case). 54 include Emerging Market or carry trade currencies. Thereby, risk aversion shocks can induce correlations in foreign exchange markets which are not necessarily be driven by the exchange rate targeting mechanism. I control for these common drivers by augmenting Equation 2.1 with global factors: e i d (t) = i (t) + f W e it e e d (t) + f W U it e U d (t) + f W $ it e $ d (t)+ K X k=1 B tk c kd (t) +C t vix d (t) + i d (t): (2.2) In Equation 2.2, c kd (t) refers to daily log returns from commodityk over quartert, and vix d (t) refers to daily log changes in the VIX index - a proxy for global risk appetite. 13 For commodities, I consider K = 5 heavily traded world commodities: Gold, copper, crude oil, coee and sugar. The two estimation procedures result in two sets of de facto peg intensities: the conventional measures ^ W b it and the estimates upon controlling for global factors f W b it which I'll refer to as the augmented measures. For robustness, I'll typically consider both when testing for monetary spillovers. 2.4 Trends in Exchange Rate Policy I estimate peg intensities for a sample of 52 currencies against the U.S. Dollar, Euro, and Japanese Yen (Table B.2 and continued on Table B.3). 14 Because of the broadly low peg levels against the Yen, I focus on the cross-country dynamics of USD and EUR peg intensities. Figure 2.1 shows percentages of countries falling into each exchange rate classication over the 2000-2018 period. Floats, intermediates and pegs are dened as peg intensity estimates 13 The CBOE VIX index is a model-free measure of 30-day expected volatility of the S&P 500 stock index derived from options prices. 14 Exchange rate data is available for 52 countries, but due to varying data coverage, after merging all data sets together the main analysis is conducted on a panel of 46 countries as discussed in Section 2.2. 55 ^ W b it 2 f[0;:1]; (:1;:9]; (:9; 1]g, respectively. 4-quarter averages are plotted for clarity. A striking consistency is how persistent the proportion of intermediate exchange rate regimes have been over the past two decades across both base currencies, particularly the USD. Roughly a third of the sample follows an intermediate peg at any given period. Moreover, the proportion of countries oating against the USD nearly doubled from 20% in 2000 to 40% by 2018. This trend was driven by countries transitioning away from a hard USD peg, rather than intermediate pegs becoming more exible. Figure 2.1: Exchange Rate Regimes Across Countries, vis-a-vis USD (left), EUR (right) Floats, intermediates and pegs are dened as peg intensity estimates ^ W b it 2f[0;:1]; (:1;:9]; (:9; 1]g, respec- tively. Rolling 4-quarter averages. A striking statistic in the data is the number and proportion of actual pure oats across the sample (Figure 2.2). In 2000, the only currency which had estimated peg intensities of less than or equal to 0.20 against all three base currencies was the British Pound. Including the three base currencies, that amounts to just four pure oats at the turn of the century. Proportionately, it is clear from the gure that pure oating currencies are historically scarce and continue to be so. In 2018, the number rose to ten if we include the base currencies USD, EUR and JPY under the assumption that they are oats. Additional identied countries 56 are Brunei and Singapore, the Chinese Yuan, Korean Won, Thai Baht, Canadian Dollar and British Pound. The Emerging Market cases are of particular interest. The currency of Brunei is ocially pegged to Singapore's, therefore its exibility vis-a-vis the USD, EUR, or JPY rises as Singapore's exibility rises despite not being a true oating currency. Throughout 2018, the Thai Baht / Singapore Dollar exchange rate was exceptionally stable, suggesting that Thailand was likely de facto targeting vis-a-vis the SGD. Singapore itself has realized steady gains in exchange rate exibility over the past two decades. The Chinese Yuan saw its peg intensity to the USD weaken dramatically since 2016 amidst rising trade tensions between China and the United States. South Korea has been under an in ation targeting monetary regime since the early 2000's. If Brunei and Thailand are dropped from the list of true oats due to their de facto targeting of the SGD, and the case of China is considered transient, that leaves just 7 currencies under a truly pure oat in 2018, with Singapore and South Korea being potentially new and notable independent oaters. Figure 2.2: Sample Proportion of `Pure' Floaters, 2000-2018 I dene a currency as a pure oater in any particular quarter if all three weights, ^ W b it where, b2fUSD, EUR, JPYg, are estimated to be less than 0.20. Rolling 4-quarter average of ^ W b it is used. Total sample contains 55 countries; number is inclusive of USD, EUR, and JPY as these assumed to oat freely given their role as potential exchange rate targets by other countries. Figure 2.3 sorts peg intensities from lowest to highest across countries, for the year 2000 57 and 2018. 15 The number of hard U.S. pegs (intensity greater than 0.90) have fallen drastically over the past two decades, while the number of oaters rose. In contrast, peg intensities against the EUR have risen over the past 20 years. 16 Moreover, the number of countries under intermediate pegs remains substantial in 2018 (roughly 60% of the sample considering both USD and EUR), and the `intensity curves' are relatively smooth - highlighting the importance of considering intermediate pegs across a broad spectrum. Figure B.2 in the Appendix shows 2000-2018 changes in peg intensity by currency. Against the USD, many countries which were hard pegs in 2000 have relaxed their policy by 2018, most of them following de facto intermediate policies. At the same time, most countries did increase the pegging weight attributed to the EUR. Focusing on USD pegs, Romania, South Korea, China, Brazil, Mexico, and Thailand round out the countries exhibiting the largest changes. Over this time period, Romania transitioned from a hard peg to the USD to targeting the EUR, explaining the near-maximal drop in USD peg intensity coinciding with a large rise in EUR peg intensity. In 2015, China begun transitioning from a hard de facto USD peg amidst the country's push to globalize it's currency, while the other countries are notable emerging markets that have adopted in ation targeting monetary policy over the period, thereby allowing market forces to increasingly drive their currency movements. An important possibility to consider is whether countries which moved away from the USD are switching to EUR as a base currency to peg against. The estimated correlation between 2000-2018 changes in USD peg intensities and 2000-2018 changes in EUR peg in- tensities is equal to -0.23 (t=-1.64) but not highly signicant in the statistical sense. The weak negative correlation implies that changes in USD peg intensity can explain roughly 5% of the variation in changes in EUR peg intensity. The evidence, therefore suggests that base currency substitution was not a major factor driving transitions in exchange rate policy. Taking a look at exchange rate intensities over time, I plot 4-quarter rolling average 15 The plotted intensities are 4-quarter averages. 16 Ito and McCauley [2019] attribute this partly to commodity currencies moving away from the pure U.S. Dollar zone to a more intermediate position between the Dollar and Euro. 58 Figure 2.3: Peg intensities in 2000 vs 2018, vis-a-vis USD (left), EUR (right) Annual 2000 and 2018 estimates of ^ W b it are 4-quarter averages. USD and EUR intensities for selected countries in Figure B.3 and aggregate, cross-country averages in Figure B.4. Romania's early-2000's transition from a USD peg to a EUR peg becomes clear. Singapore has steadily reduced it's peg against the USD to nearly zero, through for a large part of the 2000's the country seems to have targeted a basket with partial pegs against both the EUR and USD. Switzerland had a strong yet imperfect peg against the EUR over most of the sample period, though the EUR peg intensity dropped considerably during the 2011 European Debt Crisis, then returning to high levels until Switzerland surprised the world with their sudden re-valuation in January 2015 when the Franc appreciated roughly 30% against the Euro. Since then, the peg intensity has continued to steadily weaken. China's hard peg to the USD is very apparent in the early 2000's (despite the government claiming to target a basket). The country continued to administer a strong (though not perfect) USD peg up until Q4 2015, and since then the USD peg intensity has dropped sharply to less than 0.10 amidst the country's push towards introducing the Yuan as a global currency. This drop is not substituted with increased EUR intensity. 17 17 It is also possible that this sharp drop in China's targeting the USD was driven by the U.S.-China trade 59 Overall trends in USD and EUR peg intensities across all countries in the sample are shown in Figure B.4. What is clear is that the average USD peg intensity has crept lower steadily over the past 20 years (from over 0.60 to below 0.45), with the exception of 2011 dur- ing the European Debt Crisis where a sharp rise in USD peg intensity appears to have been driven by countries substituting away from targeting the EUR, which realized a coinciding sharp drop in intensity. Moreover the persistent rise of intermediate pegs accompanying a persistent scarcity of pure oats are not supportive of the Two Corners hypothesis, high- lighting the important need to more carefully study middle-ground exchange rate policies. The question of what might determine a country's choice of exchange rate policy is a natural (extensively-studied) follow-up. Many potential factors might drive this choice. For example, Edwards [1996] nds that political economy factors play a major role, as the choice between xed and oating is related to the country's historical degree of political instability, the probability of abandoning a pegged rate, and the policy objectives of the do- mestic monetary authorities. Devereux and Engel [1998] argue that what matters is whether prices are set in the currency of the consumer or producer. Recent studies also consider the choice of operating an intermediate exchange rate regime. Ito and Kawai [2014] suggest that countries opt for more exible exchange rate regimes when the country has: greater international reserves, more trading partners, a lower proportion of commodity exports, and greater domestic savings, while McCauley and Chan [2014] report that the composition of foreign exchange reserves strongly explains cross-country variation in (continuous measures of) exchange rate peg intensities. Armed with continuous peg intensities against the USD and EUR, the two globally dominant base currencies, one can eectively measure monetary policy spillovers with ner granularity. That is, we can shift our attention from the corners of exchange rate policy to interior choices, i.e. intermediate regimes. The following analysis leverages these estimated peg intensities to study whether and to what degree monetary policy spillovers are consistent war in an eort to insulate against the eects of taris. 60 with the Trilemma, particularly under intermediate pegs. 2.5 Testing the Trilemma: Empirical Strategy There are a number of steps that must be taken before arriving that the nal econometric specication to test monetary policy spillovers. For illustrative purposes, consider a modied Uncovered Interest Rate Parity (UIP) condition which allows for both open and closed capital ow regimes: R it = (1 it )(R bit +E t [e ib;t+1 ] + it ) + it R it ; it 2f0; 1g; (2.3) where whether countryi administers closed (open) capital ow is given by it : a value of 0 for open and 1 if closed. Under free capital ow ( it = 0), the interest rate of countryi,R it should equal the interest rate of the base country,R bit plus the expected percent appreciation of base country b's currency vis-a-vis country i's currency denoted E t [e ib;t+1 ], plus a risk premium it . Under a perfectly credible hard peg, E t [e ib;t+1 ] equals zero. So under a hard peg and assuming a zero risk premium and = 0, its easy to see that R it = R bit . That is, countryi does not have any monetary autonomy as the base country interest rate fully passes through. In contrast, under a exible exchange rate and/or time-varying risk premia, R it can indeed deviate from the base country interest rate. The Trilemma implies that limiting capital ows by introducing capital controls can reduce this policy pass-through and grant greater monetary autonomy. This is shown in Equation 2.3 under it = 1. Under a closed capital account, UIP no longer applies and country i's interest rate is fully independent, R it =R it . A major simplifying assumption of the illustration just presented is that exchange rates can be either xed or oating, and capital controls can either be open or closed. Despite this unrealistic assumption, most studies on the policy Trilemma are restricted to such cases. By leveraging continuous measures of peg intensity, I aim to relax this assumption. Second, 61 interest rate levels tend to be very persistent, thus raising the issue of potential unit roots and spurious regression results. Therefore, following the literature, we test for the monetary pass-through using interest rate changes. Third, as in Han and Wei [2018], it is important to condition interest rates on domestic variables which the central bank may target as we wish to capture interest rate changes exclusively driven by the Trilemma and remove bias driven by policy responses to domestic economic conditions. Additionally, it is crucial to condition base country interest rates on domestic variables (Jord a et al. [2020]) to identify base country monetary policy movements that are unrelated to domestic economic conditions. 2.5.1 Identication of Base Country Monetary Shocks The base interest rates under consideration are the U.S. and E.U. (ECB) policy interest rates, b2fUS;EUg. 18 A key identifying assumption here and in the broad majority of related studies is that all other countries take changes in U.S. and E.U. monetary policy as exogenous. That is, country i's economic condition does not factor into monetary policy decisions for the U.S. and E.U., where only domestic conditions strictly determine the interest rate. Though plausible, this assumption may or may not be reasonably satised at all times. For example, a country's business cycle may be correlated with that of the base country. Therefore, as a robustness check I also consider a measure of unanticipated U.S. monetary policy shocks later in Section 2.9.4. To remove potential endogeneity arising from policy changes driven by domestic economic conditions, instead of using interest rate changes directly, I rst run the following regression resembling a Taylor-type rule where the monetary policy responds to output and in ation: R bt = 1 + 2 y bt + 3 bt +D b;ZLB [ 1 + 2 y bt + 2 bt ] +Z bt ; (2.4) where R bt is the quarterly change in interest rate for base country b, in this case either the U.S. or E.U. y bt and bt are year-over-year GDP growth and in ation, respectively. 18 These two countries make up the lions share of globally held international reserves, and currency pegs. 62 Because of the drastic change in monetary policy after hitting the Zero Lower Bound (ZLB), I allow for a structural break in the regression coecients conditional on base country interest rates hitting their eective lower bound. This is captured by an indicator variable, D b;ZLB which takes a value of 1 if base country b's policy rate is at the eective lower bound, and 0 otherwise. For the U.S., the interest rate is at the eective lower bound when the policy rate is 0.125% or lower, and for the E.U. when the policy rate equals 0%. For both countries, the lower bound period is persistent, occuring mostly after the 2008 Financial Crisis. The estimated residual policy rate change ^ Z bt 2f ^ Z US;t ; ^ Z EU;t g { cleaned of domestic confounders { is then a measure of base country monetary policy changes that are uncorrelated with domestic economic conditions. Naturally, most identication approaches come with drawbacks. For example, while this method allows for a structural break at the ZLB, during period of zero rates, there is nearly zero variation in the policy rate, and unconventional policies dominated the central bank toolkit. Moreover, the `residual' approach may not always be sucient for identifying the exogenous component of monetary policy. To validate the robustness of the results, I apply two additional approaches for estimating ^ Z bt to capture changes in the monetary policy stance despite at the ZLB. First, I replace R bt for the U.S. and E.U. with their respective shadow rates (Wu and Xia [2016]). Second, for the U.S. specically, I construct a series of identied monetary policy shocks from Fed Fund futures data which yields an entirely dierent series of policy innovations. Results under these alternative schemes are reported in Section 2.9.3 and Section 2.9.4, respectively. The second step required for identication is motivated by the IV strategy of Jord a et al. [2015] and Jord a et al. [2020], and more generally consistent with the broader literature on the policy Trilemma. That is, the eect of base countryb's monetary policy shock on country i's interest rate depends on: country i's peg intensity with respect to the base currency of countryb given by ^ W b it , and countryi's capital account openness,K it . Both of these variables lie within [0; 1], where 0 indicates fully oating exchange rate/closed capital accounts, and 1 63 indicates fully pegged exchange rate and full capital openness. Taken together, the variable of interest in the baseline regression specication will be the interaction term ^ Z bt ^ W b it K it . The key dierence between this measure and prevailing studies is that here, the variable measuring exchange rate regime, ^ W b it is continuous and lies within [0; 1]. 19 Importantly, the identication assumption that must be satised is monotonicity: @E[R it jx] @[ ^ Z bt ^ W b it K it ] 0: (2.5) What the assumption requires is that the change in countryi's interest rate (conditional on controls, x), is increasing in the denominator. Think of peg intensity and capital openness as measures of how exposed countryi's interest rate is to the base country's, and we ideally, wish to compare two identical countries in terms of fundamentals and capital controls, but varying in exchange rate exibility. For zero exposure, either ^ W b it or K it must equal zero. That is, the country must administer either a pure oat, or a closed capital account for complete monetary autonomy { precisely what the Trilemma implies. Conversely, exposure to the base country's monetary policy is conditionally maximized (i.e. minimal monetary autonomy) when ^ W b it and K it equal 1; when country i administers a hard peg under free capital ows. The interaction term imposes the structural assumption that the Trilemma trade-os are linear in that monetary autonomy linearly decreases as exchange rate exibility or capital account openness rises. 2.5.2 Econometric Specication The baseline regression to be tested is: 19 Jord a et al. [2015] denes exogenous monetary policy shocks in the same way { as the interaction of the base country's monetary policy change, the exchange rate regime and degree of capital openness - but using binary measures of exchange rate regimes. 64 R it = i + 1 R i;t1 + 2 y it + 3 it + 4 RER it + 5 VIX t + 6 R t + US [ ^ Z US;t ^ W $ it K it ] + EU [ ^ Z EU;t ^ W e it K it ] + it : (2.6) The baseline regression assumes that country i's interest rate responds according to an open economy Taylor-type rule (Aizenman et al. [2011], Engel [2011], Han and Wei [2014], Han and Wei [2018]) and conditions on key domestic variables which the policy rate may react to. Changes in country i's policy rate are regressed on lagged policy rates, 20 R i;t1 , nominal GDP growth y it , changes in in ation it , and changes in the log real exchange rate RER it vis-a-vis the USD. Positive changes in the real exchange rate indicate country i depreciation. Including the real exchange rate also will capture any possible evidence of Fear of Floating, one phenomena which challenges the sustainability of the Trilemma (Calvo and Reinhart [2002]). The choice of real exchange rates vis-a-vis the USD is intentional: it is the most relevant exchange rate, as the USD dominates among invoicing currencies in international trade, and is also the currency of choice in international nance (Gopinath [2015], Maggiori et al. [2019], ECB [2019]). 21 Additionally the validity of the Trilemma has been actively debated in light of new evi- dence of a global nancial cycle (Rey [2015], Miranda-Agrippino and Rey [2020]), hence the specication also controls for global factors: log changes in the VIX index given by VIX t , and R t which denotes changes in the global average interest rate. 22 The merged panel 20 The specication taking the form of a dynamic panel model is well known to suer from Nickell [1981] bias when the time dimension is small. However, our quarterly sample provides T ranging from mid-40 to mid-70 depending on the sub-sample and country. Judson and Owen [1999] show through Monte-Carlo studies that the LSDV estimator performs well in comparison with GMM and other estimators when T=30. 21 Moreover, real eective exchange rate changes are highly correlated with USD exchange rate changes such that using either do not result in meaningful changes to estimates of monetary spillovers. 22 The global average interest rate is computed each period t as the cross-section average of R it across all countries i, excluding base countries. It proxies for the common factor in interest rate uctuations and absorbs common trends across countries (Pesaran [2006]). 65 data are unbalanced as data sources vary in their coverage (Table B.1 includes a description of countries along with the number of interest rate observations per country). Standard errors are clustered at the country level. Its worth brie y pointing out that the monetary shocks ^ Z b;t and peg intensities ^ W b it are both estimated, and therefore subject to the classical case of measurement error (errors-in-variables problem). Because the measurement error is embedded in independent variables, under the standard assumption that the measure- ment error is random and uncorrelated with the independent and dependent variables in the regression, this biases the coecients towards zero, and biases the associated t-statistic downwards. Measurement error therefore induces attenuation bias such that the resulting monetary spillover estimates are likely to be relatively conservative in the sense that they would otherwise be larger in the absence of measurement error. 23 The nal two terms preceding the residual it of Equation 2.6 are the focus of this study. Coecients US and EU capture the degree of spillover from base interest rates (U.S. mon- etary policy and ECB monetary policy, respectively) to country i's interest rate. Given a foreign monetary policy shock to the base country, ^ Z bt , the total spillover to country i is an increasing function of peg intensity and capital account openness, b [ ^ W b it K it ]. 24 A potential drawback of the regression specication is the imposed homogeneity of co- ecients across countries. For example, weights on Taylor Rule coecients might dier across countries which aim to prioritize dierent policy objectives: emerging markets may prioritize targeting the real exchange rate, while this may not be an objective at all among some advanced economies (Aizenman et al. [2011] and Ahmed et al. [2019]). Despite this limitation, much of the literature stands by the pooled panel regression specication as it 23 The way to adjust standard errors when a regressor is estimated typically varies on a case-by-case basis. Often however, bootstrapping the entire estimation procedure (rst stage plus second stage, etc.) is done. However, when there are many stages or many estimations in a single bootstrap round, this approach can become exceedingly intensive in terms of computation time. The approach applied in this paper is one of those scenarios: in the rst stage, I estimate for each of 46 countries, and for each quarter, peg intensities, which then enter into a second stage panel regression (the rst stage yielding roughly 3,450 estimates). This would then have to be bootstrapped hundreds of times. 24 Ito and Kawai [2012] and Ito and Kawai [2014] apply a similar method to estimate a country's monetary independence, but they do not pre-condition base country interest rates on domestic variables or account for nancial openness. 66 buys considerable statistical power when dealing with cross-country panels 25 . In support of the homogeneous coecients restriction, Han and Wei [2018] nd that after estimating country-specic Taylor-type regressions, weights assigned to in ation for in ation targeting countries and non-in ation targeting countries are not statistically dierent. However to account for potential heterogeneity in regression coecients, I estimate the regression on ad- vanced and emerging market sub-samples of countries along with the full sample. Moreover in Section 2.7 I allow the coecients to be estimated separately across countries binned by exchange rate peg intensity, re ecting the possibility that countries with greater monetary autonomy under a exible exchange rate can put more weight on domestic policy objectives compared to countries administering stronger pegs (Klein and Shambaugh [2015]). 2.5.3 Tests and Hypotheses The policy Trilemma assumes that b = 1 from Equation 2.6. That is, under a perfect peg and open capital ows ( ^ W b it = 1, K it = 1), interest rate pass-through should be one-for-one, while under a pure oat ( ^ W b it = 0) or closed capital ows (K it = 0), there is no interest rate pass-through (i.e. complete monetary autonomy). However, in practice it is dicult to expect this assumption to hold. First, the policy Trilemma relies on UIP being satised, but there is extensive empirical evidence of UIP being violated in the data. Second, as Klein and Shambaugh [2015] show, one cannot expect Trilemma-consistent pass-through if country i's interest rate changes are correlated with other factors that in uence their policy rate such as expected exchange rate changes, risk premia or global shocks. Nonetheless, there are a number of valuable tests that can be conducted. If b is sta- tistically signicant and positive, that itself is evidence in favor of the Trilemma despite imperfect pass-through. A positive coecient implies a statistically signicant relationship between base country policy rates and country i's policy rates which strengthens as the ex- change rate policy becomes increasingly rigid, or as capital accounts become more open. A 25 Obstfeld et al. [2005], Klein and Shambaugh [2015], Han and Wei [2018], Obstfeld et al. [2019a] all employ the pooled specication in their baseline analysis. 67 continuous measure of exchange rate regime will let us infer whether intermediate exchange rate regimes oer intermediate degrees of monetary policy autonomy. Given the linear form of the interaction term, it is simple to calculate spillovers under any combination of exchange rate exibility and capital account openness. To focus on the trade-o between monetary autonomy and exchange rate exibility, the discussion focuses on the case whereK it = 1, or conditional on an open capital account for ease of interpretation. This way, we can make comparisons on the monetary autonomy between two hypothetical countries, both with open capital accounts, but dierent exchange rate policies. A similar design, though with discrete exchange rate regimes, is taken in Han and Wei [2018]. In fact, this is not a binding constraint { we can x the capital account openness to any value ofK it and still infer the monetary autonomy - exchange exibility trade-o between countries given the same capital account openness. This point is particularly important to note because the scenario ofK it = 1 may not be borne out in the data particularly among emerging markets. Fortunately, under the assumptions of the Trilemma (i.e. monetary spillovers are linearly increasing in K it ), the case of K it = 1 is easily inferred from the model even for emerging markets. Dierent coecient estimates of US and EU suggest that monetary policy spillovers are heterogenous, and may be dierent depending on the base currency. Finally, a signicant coecient on both US and EU in a regression including both suggest (but do not conclude) that basket pegs, where the same total weight W b it is allocated across base currencies, can oer diversication benets compared to a hard peg (where the equivalent total weight is allocated to a single currency) against a single base currency so long as the base country monetary policies are not perfectly correlated with one another. For example, a country targeting a basket of two exchange rates with weights of 50% on each, would be imperfectly exposed to both monetary policies, versus committing 100% weight towards single currency. Despite equal total foreign exposure (weights sum to 1 in both cases), the country targeting a basket is subject to less monetary pass-through, on average, from base countries in each 68 period so long as the two base countries do not conduct synchronized monetary policy. If the two base country monetary policies are imperfectly correlated, pass-through is is reduced under a basket peg for any given quarter. If the two base country monetary policies were perfectly correlated, then there would be no dierence between the two weighting schemes (and in fact, one of the RHS regressors, either the US or EU monetary policy shock, would drop out of the regression). The latter two tests would bring novel insights to the literature. A key assumption of the regression specied in Equation 2.6 is the implicit linearity imposed on monetary pass-through. The eect of monetary pass-through implied by b [ ^ W b it K it ] is linear in peg intensity and capital account openness, and it follows that under open capital accounts, the trade-o between monetary autonomy and exchange rate stability is also linear. The Trilemma trade-os however are not necessarily required to be linear, though have been assumed to be so in some studies (Ito and Kawai [2014]). There is no consensus on the linearity of Trilemma trade-os. Aizenman et al. [2010] and Herwartz and Roestel [2017] test the linearity assumption and nd supportive evidence. In contrast, Obstfeld et al. [2019a] nd non-linear eects of (non-monetary) spillovers under varying degrees of exchange rate exibility. Because of the important policy implications of (non) linearity, I explore this issue in more detail in Section 2.7 by exploiting the continuous nature of peg intensity measures. 2.6 Baseline Results The results for the full sample of countries are reported in Table 2.2. The rst three columns represent dierent variants of the peg intensity estimate ^ W b it . The second and third columns use a 2-quarter and 4-quarter rolling average of ^ W b it , respectively denoted with (RA, 2) and (RA, 4), to replace the unsmoothed measure (column 1). Smoothing out the peg intensity estimate with past observations helps makes a more conservative choice to ensure that pegs, which tend to be persistent, are well-established (Jord a et al. [2015], Jord a et al. [2020]). 69 Table 2.2: Baseline Regression Results: All Countries (1) (2) (3) (4) (5) (6) ^ W b it ^ W b it (RA, 2) ^ W b it (RA, 4) ^ W b it (RA,2) ^ W b it (RA, 2) f W b it 2 (0; 1) ^ US 0.351*** 0.370*** 0.402*** 0.486*** 0.412** 0.390*** (0.108) (0.124) (0.136) (0.177) (0.147) (0.098) ^ EU 0.511*** 0.486*** 0.581*** 0.328* 0.703*** 0.392*** (0.124) (0.133) (0.178) (0.178) (0.116) (0.120) Adj. R 2 0.15 0.14 0.14 0.06 0.15 0.16 F-Statistic 69.51 58.77 47.31 47.80 44.91 75.25 NT 2,882 2,532 1,937 2,532 1,727 2,909 Country FE Y Y Y Y Y Y Time FE N N N Y N N ***,**,* refer to signicance at the 1%, 5% and 10% level, respectively. Robust standard errors clustered at the Country level. Regression specication of Equation 2.6. Estimation period: Q2 2000 - Q4 2018. Column 5 estimates on the sub-sample of intermediate pegs (peg intensities between 0 and 1, for both U.S. and E.U.). Column 6 uses f W b it , the estimated peg intensities (Equation 2.2) after controlling for common shocks. Within R-squared reported. Moreover, smoothing even over 2 quarters helps ensure that results are not driven by outliers and helps eliminate episodes of opportunistic pegging and sudden short-lived devaluations. Regardless, estimates are consistent and signicance is broadly robust across columns. Col- umn 4 reports results after substituting a time xed eect for global controls. Column 5 reports results the sub-sample of country-quarter observations under intermediate pegs, and Column 6 reports results under the augmented peg intensity measure, f W b it for additional robustness. 2.6.1 All Countries Signicant non-zero estimates on both ^ US and ^ EU indicate Trilemma-consistent monetary spillovers from both base countries to others (Table 2.2). Under free capital ows (K it = 1), as peg intensity rises (falls), the pass-through of base country interest rates strengthens (weakens). Note that the eects are statistically dierent from both 0 and 1, implying imperfect Trilemma pass-through. That is, under a perfect peg and free capital ows, a 1 70 percentage point change in the base country (US, EU) interest rate is associated with interest rates roughly (+0.37, +0.49) percentage points higher (Column 2). Column 4 introduces time xed eects as a robustness check - the eects of monetary pass-through broadly hold under this specication as well, and the results are robust to using the augmented measure f W b it . 2.6.2 Advanced economies Table 2.3 reports estimates for the sub-sample of advanced economies. Both base country Trilemma coecients are highly signicant across the varying specications of peg intensity and remain robust to both country and time xed eects. Both U.S. and E.U. base country pass-through is roughly 0.70 for advanced economies, much higher than it is for the full sample. In fact, in many instances the condence interval includes 1 { indicative of near- perfect monetary policy pass-through when targeting either base currency. Moreover, a hypothetical advanced economy with free capital ow targeting a 50-50 USD-EUR basket would import about half of each country's monetary policy change. So long as these policy rate changes in the U.S. and E.U. do not occur simultaneously, targeting a basket would appear to oer potential diversication benets. 2.6.3 Emerging markets Table 2.4 reports pass-through estimates for the sub-sample of emerging markets. Across all four specications (columns 1 to 4), coecient estimates suggest positive yet imperfect pass- through, but there is little evidence of signicant monetary policy spillovers from the E.U., despite a number of emerging market economies pegging, at some point, to the Euro. 26 In contrast, the eect of U.S. monetary policy is statistically signicant in most specications, ranging from 0.26 to 0.44, indicating that under a perfect peg and free capital ows, monetary spillovers from the U.S. are imperfect, with emerging market interest rates rising on average 26 These countries include but are not limited to: Albania, Bulgaria, Croatia, Czech Republic, Hungary. 71 Table 2.3: Baseline Regression Results: Advanced Economies (1) (2) (3) (4) (5) (6) ^ W b it ^ W b it (RA, 2) ^ W b it (RA, 4) ^ W b it (RA, 2) ^ W b it (RA, 2) f W b it 2 (0; 1) ^ US 0.656*** 0.742*** 0.797*** 0.701*** 0.737*** 0.529*** (0.213) (0.209) (0.220) (0.198) (0.178) (0.159) ^ EU 0.799*** 0.759*** 0.700*** 0.422*** 0.663*** 0.701*** (0.071) (0.117) (0.131) (0.121) (0.088) (0.076) Adj. R 2 0.42 0.43 0.42 0.186 0.40 0.41 F-Statistic 70.40 62.91 46.60 39.59 40.04 68.5 NT 746 644 486 644 444 777 Country FE Y Y Y Y Y Y Time FE N N N Y N N ***,**,* refer to signicance at the 1%, 5% and 10% level, respectively. Robust standard errors clustered at the Country level. Regression specication of Equation 2.6. Estimation period Q2 2000 - Q4 2018. Advanced Economies sub-sample only. Column 5 estimates on the sub-sample of intermediate pegs (peg intensities between 0 and 1, for both U.S. and E.U.). Column 6 uses f W b it , the estimated peg intensities (Equation 2.2) after controlling for common shocks. Within R-squared reported. +0.35 percentage points for every +1 percentage point rise in U.S. interest rates. 2.6.4 Intermediate pegs Column 5 of Tables 2.2, 2.3 and 2.4 consider the sub-sample of country-quarter observations which do not include pure oats or hard pegs (i.e. excluding values of 0 or 1 for ^ W b it ). This is done to verify whether corner policies are driving the results of the regression tests, or whether the range of intermediate pegs actually oer a spectrum of monetary autonomy. Across the full sample, the eects of both U.S. and E.U. peg intensity remain highly signicant upon omitting corner policy observations, suggesting that the intensive margin of peg intensity also matters for monetary policy. The advanced economy sub-group signals the same message: the eects of monetary policy pass-through hold for both the intensive and extensive margin of exchange rate regimes. For the emerging market sub-group, the signicance of the coecient estimate on ^ US disappears (though remains positive) when removing observations containing corner policies 72 Table 2.4: Baseline Regression Results: Emerging Markets (1) (2) (3) (4) (5) (6) ^ W b it ^ W b it (RA, 2) ^ W b it (RA, 4) ^ W b it (RA, 2) ^ W b it (RA, 2) f W b it 2 (0; 1) ^ US 0.266** 0.265** 0.263* 0.444** 0.165 0.356*** (0.108) (0.121) ( 0.135) (0.198) (0.143) (0.116) ^ EU 0.199 0.181 0.458 0.066 0.868*** 0.064 ( 0.167) (0.179) (0.322) (0.218) (0.261) (0.177) Adj. R 2 0.13 0.13 0.12 0.04 0.13 0.14 F-Statistic 46.09 39.18 31.00 32.57 29.94 49.98 NT 2,135 1,887 1,451 1,887 1,282 2,131 Country FE Y Y Y Y Y Y Time FE N N N Y N N ***,**,* refer to signicance at the 1%, 5% and 10% level, respectively. Robust standard errors clustered at the Country level. Regression specication of Equation 2.6. Estimation period Q2 2000 - Q4 2018. Emerging Markets sub-sample only. Column 5 estimates on the sub-sample of intermediate pegs (peg intensities between 0 and 1, for both U.S. and E.U.). Column 6 uses f W b it , the estimated peg intensities (Equation 2.2) after controlling for common shocks. Within R-squared reported. (Column 5, Table 2.4). This may have several interpretations. One is that across emerging markets, intermediate pegs may not oer intermediate monetary autonomy, but rather dis- proportionately greater monetary autonomy than a hard peg, indicating a non-linear relation- ship between exchange rate exibility and monetary autonomy: a country which introduces a little bit of exchange rate exibility can potentially buy a lot of monetary independence. There are other possible interpretations as well: for these countries, increasing exibility of the exchange rate might disproportionately increase the sensitivity of monetary policy to non-Trilemma factors (domestic objectives, Fear of Floating, nancial cycles or commodity cycles, risk premia, etc.). So, while the base country's monetary policy spillovers are less in uential, the costly rising importance across other external factors may oset any bene- ts from monetary autonomy. In the next section, we will investigate these non-linearities further, and allow regression coecients to vary across peg intensities to possibly re ect changing weights on policy objectives as countries move from pegs to oats. Finally, in an interesting twist when considering only intermediate peg observations, 73 monetary spillovers under the Trilemma with regards to E.U. monetary policy becomes sta- tistically signicant (^ EU ), implying that under intermediate peg intensities, E.U. monetary policy passes through to countries which partially target the Euro and the pass-through increases as the country approaches a peg. However surprisingly, hard pegs to the Euro do not exhibit Trilemma-consistent monetary spillovers in emerging markets. 2.6.5 Discussion To summarize, signicant evidence of monetary policy spillovers is present in both advanced Economies and emerging Markets, but estimated monetary policy pass-through is consider- ably stronger among advanced economies. For the full sample and advanced economies in particular, there is robust evidence consistent with Klein and Shambaugh [2015] that the Trilemma holds under interior policy choices (i.e. peg intensities between 0 and 1), poten- tially allowing for partial monetary autonomy under a managed oat. These results validate the prevailing literature testing the Trilemma. Both monetary policy spillovers and overall regression t (R 2 ) are lower for the emerging markets sub-sample compared to advanced economies. This could be due to the presence of important factors which are correlated with countryi's interest rate. For example, monetary pass-through estimates may be low in emerging markets because risk premia tend to be highly volatile (Kalemli-Ozcan [2019b]). Fear of Floating and Global Financial Cycles, operating through the real exchange rate and nancial conditions respectively, may also impact country i's policy choices (Calvo and Reinhart [2002] and Rey [2015]). Some emerging markets are heavily reliant on commodity trade, hence exposing themselves to commodity cycles which in turn can in uence policy objectives (Aizenman et al. [2011]). Finally, recent evidence suggests that the burgeoning debt positions of emerging markets (and advanced economies) brought in by unprecedented monetary easing after the 2008 Financial Crisis may be interacting with monetary policy objectives (Ahmed et al. [2019]). A new insight is the signicance of joint pass-through from both U.S. and E.U. monetary 74 policy { bearing a key policy implication: basket pegs can potentially mitigate monetary policy spillovers from a single country occurring under a unitary peg by taking on monetary spillovers from an additional country, eectively diversifying spillover risk. Interestingly, Emerging Markets do not seem to exhibit Trilemma-consistent monetary policy spillovers under intermediate pegs. However, this may imply that among these countries, moving from a hard peg to an intermediate peg buys a disproportionate amount of monetary independence { either unconditionally or relatively by assigning greater weight on other policy objectives. Potential non-linearities in the exchange rate regime { monetary spillover function are ex- plored in the next section. 2.7 Non-linear Trilemma Trade-os Thus far, I've provided evidence conrming that monetary spillovers subject to the Trilemma are present in both advanced economies and emerging markets. However, as I mentioned, the regression design implicitly imposes that the spillover country i faces is linear in exchange rate exibility: b ^ W b it , given free capital ows (K it = 1) and a unitary monetary shock ( ^ Z bt = 1, though the size of the shock can be arbitrary). We saw, however, that the Trilemma seems to hold for intermediate pegs among advanced economies and corner policies appear to drive the signicant results among emerging markets. This brings the implication of linear monetary spillovers into question { a research area which has received limited attention. In this section we further explore monetary policy pass-through under intermediate peg intensities, asking specically whether the relationship with exchange rate exibility is non- linear. Testing for non-linearities in U.S. and E.U. spillovers jointly is not feasible under the baseline regression design due to the size of the sample. 27 Therefore, I focus rst on non- linearities in U.S. monetary policy. Then, I modify the regression analysis to a setting which can jointly analyze the linearity of monetary spillovers under intermediate exchange rate 27 It would require interacting all covariates twice, and sub-samples already are limited in the number of observations they include. 75 regimes for both U.S. and E.U. Finally to further test whether the observed non-linearities are statistically signicant, I extend the baseline regression to a Generalized Additive Model (GAM) specication adopted from the machine learning literature. 2.7.1 Peg intensity bins I start simple with a baseline analysis which allows the researcher to investigate how well the imposed linearity assumption of the original specication is satised without adding complexity. To do this, I relax the linear-implied specication of the baseline regression (Equation 2.6) and estimate separate sub-samples, sorting by peg intensity. Again, using the 2-quarter rolling average peg intensities, ^ W b it (RA, 2). Country-quarter observations are sorted into the following 6 bins: Pure Float Hard Peg 1 2 3 4 5 6 W b it [0,0.1] (0.1,.30] (0.30,.50] (0.50,0.70] (0.70,0.90] (0.90,1] The regression specication must be modied due to the more limited number of ob- servations per sub-sample after dividing the data into 6 separate groups. Moreover, I only consider peg intensities to one base country at a time, starting with the U.S (Results for E.U. shocks can be found in Table B.4). Constructing bins which condition both on U.S. and E.U. peg intensity would lead to too few observations per group. 28 The regression takes the following form: 28 One could take Equation 2.6 and interact ^ Z b with binned peg intensities, which would potentially allow for both U.S. and E.U. to be jointly tested for non-linear pass-through. However, this comes at the cost of constraining all other regression coecients to be pooled together across the entire sample. Because policy weights can vary across countries which peg or don't peg, It's crucial to allow for coecient exibility, something that can be achieved by estimating on sub-samples. Results from this approach are reported in Table B.4 and are broadly consistent with other specications. 76 R it = i + 1 y it + 2 it + 3 RER it + 4 VIX t + 5 R t + US [ ^ Z US;t K it ] + it : (2.7) there are two key dierences between Equation 2.7 and the previous specication, Equa- tion 2.6. The rst is that the lagged dependent variable is removed from the RHS. This is due to data limitations { by constructing sub-groups using more rened exchange rate regime categories, each group will not have sucient data along the time dimension to reduce the bias that a xed eects dynamic panel specication generates. Moreover, each observation is now increasingly valuable for statistical power, and therefore lost observations from in- cluding a lagged dependent variable becomes costly for inference. On a positive note, since the regression specication is in interest rate changes the data is not persistent, thereby excluding a lagged dependent variable will not in uence the results in a meaningful way. 29 The second change is related to peg intensity. First, I only consider U.S. monetary policy spillovers, so the variable capturing shocks from the E.U. is removed. Second, peg intensity, ^ W US it is removed from the trio of interactions. This is simply because now we condition the entire sample on ^ W US it by estimating separate regressions per intensity bin. An advantage of this specication aside from its simplicity is that, by running separate bin-specic regressions, we allow all of the coecients to be heterogeneous across peg intensity bins, lending to more realistic and exible inference, and addressing some of the limitations mentioned previously over the original pooled specication. 29 If the regression was estimated in levels, removing the lagged dependent variable would very likely have a major impact on coecient estimates. To demonstrate the robustness of omitting the lagged dependent variable in Equation 2.3, the coecients on ^ US and ^ EU from Table 2.2 column 1 would change from 0.351 to 0.364 and 0.511 to 0.500, respectively. The results remain statistically signicant at the 1% level. 77 2.7.2 All countries Table 2.5 reports spillover estimates from U.S. monetary policy across bins (^ US ), but also re- ports coecients on the other covariates. This way we can infer whether monetary spillovers are non-linear in peg intensity, but also if greater monetary autonomy indeed translates to greater weights on domestic variables, namely in ation or output. The sixth row reports the spillover coecients given by ^ US , and as the Trilemma implies, the coecients roughly increase with peg intensity, with hard pegs having the largest spillover coecients (0.48). However, there is evidence of potential non-linearity in spillovers based on peg intensities. Under weak to moderate peg intensities ranging of 0 to 0.50 (bins 1 to 3), evidence of mon- etary spillovers is statistically indierent from zero { the same as if under a fully oating policy. Evidence of monetary spillovers begin to manifest under more rigid exchange rate policy (bins 4 to 6, peg intensities from 0.5 to 1). Moreover, moderately strong pegs (bin 4 and 5) exhibit weaker monetary pass-through from the U.S.compared to hard pegs (bin 6), 0.27 and 0.20 versus 0.48, respectively. This evidence has policy implications, as it sug- gests that a little bit of exchange rate exibility can potentially buy a considerable degree of monetary autonomy, and that some exchange rate stability can be bought without sacri- cing monetary autonomy. Hence, the policy Trilemma trade-o appears to be non-linear in the data, which diers from ndings of Aizenman et al. [2010], Ito and Kawai [2014], and Herwartz and Roestel [2017]. Moreover, coecients on in ation tend to remain highly signicant even under weak to moderate peg intensity (bins 2 and 3) and are approximately 7 times larger than under a hard peg (bin 6), suggesting that the gains from monetary autonomy are associated with greater emphasis on targeting domestic policy objectives, particularly in ation, The evidence suggests that pure oating is not necessary to achieve these gains. There is also some evidence that under a both oating and xed exchange rates, and a particular intermediate pegs (bins 1, 2 and 5, 6), monetary policy is increasingly in uenced by global nancial conditions proxied by changes in the VIX index. Under a exible (xed) exchange rate, interest rate changes 78 tend to respond positively (negatively) to changes in the VIX. Because U.S. monetary policy tends to ease in the presence of heightened risk, pegged monetary policy also falling when the VIX rises is consistent with the Trilemma. Under oating exchange rates, interest rates tend to rise { this is shown to be driven by the emerging markets sub-sample, who tend to tighten monetary polciy, instead of ease, during periods of heightened risk aversion in hopes of mitigating sudden capital out ows. 2.7.3 Advanced economies Table 2.6 reports results across advanced economies. Again, monetary policy pass-through estimates are nearly monotonically increasing in peg intensity. Hard pegs (bin 6) suggest full pass-through with a coecient of approximately 1. A weakly non-linear trade-o between exchange rate regime and monetary autonomy is present among the advanced economy sub- sample. A moderate to strong peg (bins 4 and 5) have spillover estimates of 0.43 and 0.62, respectively, suggesting that giving up a little exchange rate stability can cut monetary spillovers by 50%. Weaker pegs (bins 2 and 3) suggest even greater gains in monetary autonomy which are not statistically dierent than monetary autonomy under a oating exchange rate. The evidence suggests that a country which oats it's exchange rate can administer stabilization with little cost in monetary independence, while a country running a hard peg can give up a little stability to buy a considerable degree of monetary autonomy. Across advanced economies, there is consistent evidence that intermediate exchange rate regimes oer countries greater weight allocation to domestic objectives, particularly output growth, but not in ation. Under oating and most intermediate exchange rate regimes, out- put growth has a signicant coecient (bins 1, 2, 4 and 5) which is not present under a hard peg. Evidence that global nancial conditions have strong in uence over advanced economy interest rates is weak (mostly insignicant coecient estimates on VIX it ). Taking this point together with the results on domestic policy objectives, it appears that for advanced economies, exibility allows countries to focus on domestic objectives without surrendering 79 autonomy to global nancial forces. 2.7.4 Emerging markets Table 2.7 reports results for emerging markets. Across the emerging market sample under hard pegs there is signicant evidence of U.S. monetary pass-through, though imperfect (coecient of 0.367). Consistent with hard pegs to the U.S. Dollar, changes in the VIX index are associated with interest rate cuts among hard pegging emerging markets. In addition, these countries exhibit the strongest evidence of responding to real exchange rate depreciation by hiking interest rates (Fear of Floating, Calvo and Reinhart [2002]). Like their advanced economy counterparts, across bins monetary policy pass-through appears non-linear in exchange rate peg intensity. Moving from a hard peg (bin 6) to a moderately strong peg (bin 5) can reduce on average, interest rate pass-through by two- thirds (from 0.37 to 0.13). Even more striking, is that bins 2 through 4 show no evidence of signicant monetary pass-through. That is, light pegs (bin 2) and even moderate pegs (bins 3 and 4), on average, aord as much monetary autonomy as a free oating exchange rate (bin 1). Emerging market monetary pass-through, in comparison to advanced economies, appears much more non-linear in exchange rate exibility. Moderate pegs (bins 2 and 3) appear to put as much weight on targeting in ation as free oating emerging markets (bin 1) and about 7 times as much weight compared under a hard peg (bin 6). However, contrasting with advanced economies, there is evidence of global nancial conditions signicantly impacting the monetary policy of emerging markets under free oats or moderate oats (bins 1 and 2). Therefore, exible exchange rates in emerging markets may be double-edged: while it buys monetary autonomy and greater allocation to domestic objectives, policy choices will also be in uenced by global factors (Miranda-Agrippino and Rey [2020]). The sweet spot seemingly lies in the intermediate range { U.S. peg intensities between 0.30 and 0.5 { where policy rates are able to adjust to domestic in ation, while buying a signicant degree of monetary policy autonomy and 80 insulation from global nancial shocks. Figure 2.4: U.S. spillover estimates ^ US by Peg Intensity Bins Peg intensity of 1 corresponds to pure oat. Peg intensity of 6 corresponds to hard peg vis-a-vis the USD. Estimates of ^ US from Equation 2.7. Dark-shaded bars are statistically signicant at the 10% level. 2.7.5 A Generalized Additive Model Approach The baseline non-linear regression analysis sheds light on new evidence of a varying trade-o between exchange rate exibility and monetary independence, especially under intermediate exchange rate regimes. However, without a formal test, we cannot conclude whether the evidence points to an actual non-linear trade o, or whether the results are caused by mea- surement noise. For example, it is possible that for emerging markets, the relationship is indeed linear, but just so weak that under more exible exchange rates it is too dicult to dierentiate from a null eect. To test more rigorously for non-linearities, I adopt a exible non-paratmetric regression framework by estimating a generalized additive model (GAM), an approach rst introduced in the machine learning and statistical learning literature by Hastie and Tibshirani [1990]. The concept is quite simple. Unlike linear regression which assumes that the dependent 81 variable and the independent variable are linearly related, under a GAM, the relationship is allowed to be linear or a non-linear smooth function. Typically, this is denoted as: Y it =X 1it + s(X 2it ) +e it ; (2.8) where X 1it takes on a traditional linear relationship with Y it , but X 2it does not have to. The function s() is an unspecied smooth (non-parametric) function, often constructed from a number of basis functions (e.g. splines). While the method was introduced decades ago, GAMs have only recently gained popularity in application due to advances in computing power, as estimation can become computationally intensive under high dimensional settings. I recast the baseline regression model (Equation 2.6) in a GAM setting specically tailored to address the question at hand: R it = i + 1 R i;t1 + 2 y it + 3 it + 4 RER it + 5 VIX t + 6 R t + US [ ^ Z US;t s( ^ W $ it )K it ] + EU [ ^ Z EU;t s( ^ W e it )K it ] + it : (2.9) Notice in Equation 2.9, I leave everything as is, but now allow the the functional re- lationship with peg intensity, ^ W b it to be non-linear. Moreover, this specication allows us to jointly investigate spillovers from the U.S. and E.U. because the model is able to in- corporate information from the full panel, hence no sub-sampling is required. The smooth function s( ^ W b it ) is estimated via penalized cubic splines. 30 Two main estimation approaches are typically used for tting GAMs, cross validation or generalized cross validation (GCV) or (restricted) maximum likelihood (REML). GCV is shown to be unbiased asymptotically, but in application with small samples, typically suers from under-smoothing. For these 30 Penalized cubic splines are cubic splines, but changes at knots are penalized, shrunk towards zero. This helps prevent over tting even in the presence of many knots. 82 reasons, I estimate the GAM via REML, which is typically robust to under-smoothing but more computationally intensive (Wood [2017]). There are alternative modeling approaches to GAMs which also allow for smooth non- linear relationships in regression analysis. For instance, smooth transition models have been used prominently for modeling exchange rate dynamics (Franses et al. [2000], Taylor et al. [2001]). GAMs, however, are substantially dierent from smooth transition models. First, GAMs are not restricted to discrete regimes, while smooth transition models a priori assume discrete, usually two, regimes, while the transition between the regimes is smooth. 31 More- over, the smooth transition between regimes typically has a pre-specied functional form (e.g. logistic or exponential), which itself imposes symmetry in the transition probabilities. Other potential issues with smooth transition models are that identifying the transition function may be dicult in cases where the underlying data does not provide sucient information, and that ndings can also depend on the starting values. GAMs are not restricted by any of these parametric assumptions. Finally, and crucially, GAMs are suciently exible to allow for a single non-linearity in the model within an interaction term. By only allowing spillovers to vary non-linearly with exchange rate regime while keeping everything else similar to the standard econometric specication (Equation 2.3), we can call out the marginal eects of introducing non-linearity along the single, focal dimension tailored to our specic research question. I estimate the model for all countries, and the two sub-samples (advanced economies and emerging markets). For each model, the estimation procedure selected 10 knots. Figure 2.5 shows U.S. spillover estimates under the GAM specication, with 95% credible intervals. Red dashed lines are the spillover estimates implied by the linear baseline specication, Equation 2.6. It's clear that for some regions of peg intensity, the non-linearity is statistically signicant at the 5% level or lower across both the full sample and sub-samples. Across the full sample, the Trilemma eects don't appear to kick in until peg intensity 31 More than two regimes quickly increases the number of parameters that need to be estimated. 83 Figure 2.5: GAM Estimates: U.S. spillover estimates by Peg Intensity Spillover estimate is under free capital controls (K it = 1). Estimates are from Equation 2.9. Shaded areas are 95% credible intervals. Number of knots selected: 10 via REML. Red dashed line is the implied linear spillover under Equation 2.6. reaches north of 0.50, suggesting that reasonably managed exchange rates can potentially enjoy a high degree of monetary independence. However, the Trilemma conditions appear to take eect sharply beyond a peg intensity of 0.75, accelerating rapidly. The monetary transmission function is estimated to be highly non-linear for emerging markets, making a wave-like pattern, only turning statistically signicant for pegs and near-pegs. For peg intensities ranging from 0 to 0.75, monetary policy spillover estimates are statistically in- dierent from zero for emerging markets. The advanced economy sub-sample also indicates non-linear monetary spillovers, with statistically insignicant estimates from a peg intensity of 0 to 0.5, but then spillover estimates accelerate sharply as peg intensity rises further. Binned analysis results for E.U. spillovers are reported in Table B.4, with GAM estimates for E.U. spillovers are reported in Figure B.5. Unlike U.S. spillovers, E.U. spillovers do not increase monotonically across bins (but do roughly increase in peg intensity), exhibiting some non-linearity. However, under the GAM specication, these non-linearities related to E.U. spillovers are statistically insignicant. Finally, for robustness, I also present a set of results from the GAM estimation under a more conservative selection of 5 knots rather than the 84 10 knots selected by the estimation algorithm (Figure B.6), which increases the smoothness of the spillover function. The results and non-linear shapes presented here broadly hold, suggesting that the estimates are robust under varying tuning parameters. 2.7.6 Discussion The evidence from this section points to a non-linear trade-o between exchange rate exi- bility and monetary autonomy across the full sample and advanced economy and emerging market sub samples, bringing into question the traditional assumption of a linear Trilemma. Initially, under the simple binned analysis, evidence pointed to non-linear Trilemma trade- os between monetary autonomy and exchange rate stability in both advanced and emerging countries. Weak and moderate pegs come with more stability than oating exchange rates while providing just as much monetary independence. Even moving from a hard peg to one that is strongly managed appears to reduce disproportionately the degree of monetary policy pass-through a country is exposed to. These non-linear patterns are further conrmed under the more sophisticated GAM model, and the non-linearities test as statistically signicant among both advanced economies and emerging markets. It's also apparent that under varying degrees of peg intensity, countries allocate to domes- tic targets dierently, and this may be enabled by gains from a non-linear trade-o, or weak adherence to the Trilemma. Among advanced economies, greater monetary autonomy bought with exchange rate exibility is associated with stronger weights on domestic policy objec- tives (output growth), with no evidence of a global nancial cycle eect on monetary policy. For emerging markets, exchange rate exibility and greater monetary autonomy translates to heavier emphasis on in ation as a domestic policy target. Global nancial cycle eects on monetary policy are present under both oating/near- oating and near-hard/hard peg regimes in emerging markets, therefore mid-intensity pegs appear to oer the best trade-o for this group of economies in terms of monetary independence and exchange rate stability. 85 2.8 What Induces Non-Linear Monetary Spillovers? 2.8.1 Active reserves management Is this empirical non-linearity between exchange rate exibility and monetary autonomy a free lunch, or generated through some economic friction? To address this, I explore two possible mechanisms which could result in a non-linear trade o between exchange rate stability and monetary independence. The rst of these is the role of reserves accumulation as an additional policy tool. The potential for foreign exchange interventions to allow a country to violate the Trilemma constraint has been discussed in the literature. Obstfeld et al. [2010] argue that the demand for reserves is crucially motivated by the objective of nancial stability amid increased nancial integration. Empirically, they nd that countries under soft pegs tend to hold signicantly greater levels of reserves. 32 These countries may wish to actively intervene in exchange markets to prevent external nancial shocks from causing large exchange rate devaluations. Aizenman et al. [2010] document the trend of several emerging markets choosing to target intermediate levels of exchange rate stability and nancial openness while maintaining high levels of monetary autonomy, thereby violating the Trilemma. These countries also tend to hold sizable levels of international reserves. Steiner [2017] and Angrick [2018] also report evidence suggesting that the policy Trilemma constraint can be relaxed with active reserves management. Using international reserves to relax the policy Trilemma constraints applies whether UIP holds or is violated. If UIP holds, a country may choose to intervene in foreign exchange markets as an alternative way to stabilize the exchange rate rather than altering the interest rate directly. Specically, sterilized interventions would, in theory, achieve exchange rate stability without changing the money supply. On the other hand, unsterilized interventions would alter the money supply, but with a lag, and therefore unsterilized interventions can also grant exchange rate stability with monetary independence { in the short-run. If UIP 32 The eect of a hard peg was found not to be statistically signicant, but economically signicant and quantitatively similar to that under a soft peg. 86 fails to hold (as it seems to empirically) then that itself causes the Trilemma constraints to break down. In this situation, matching the monetary policy of the base country may simply not be sucient to maintain the desired level of exchange rate stability, with direct intervention being more eective. To investigate the role of active reserves management, I test whether the accumulating and expending of country i's foreign exchange reserves are associated with base country monetary policy changes. To do this, I simply replace the dependent variable of the baseline equation (Equation 2.6) with a measure of changes in international reserves: IR it = i + 1 IR i;t1 + 2 y it + 3 it + 4 RER it + 5 VIX t + 6 R t + US [ ^ Z US;t ^ W $ it K it ] + EU [ ^ Z EU;t ^ W e it K it ] + it ; (2.10) where IR it is the quarterly change in logged international reserves (excluding gold) of country i in quarter t. Reserves are measured in terms of USD. Considering the growth of reserves accounts for dierences in levels of international reserves across countries, and the RHS of the equation controls for dierent GDP growth rates across countries { hence this specication nests the case where reserves are measured per GDP, logged IR/GDP. Data on international reserves is taken from the IMF International Reserves and Foreign Currency Liquidity database. A negative coecient on ^ b suggests that a reduction in reserves is associated with a positive shock to the foreign interest rate, and this reduction strengthens in the degree of exchange rate rigidity. Under a rigid exchange rate regime, a higher foreign interest rate, without a reciprocated change in the local country interest rate, would cause capital out ows and currency depreciation. However, this could be mitigated without an interest rate change (i.e. preserving monetary independence) if the central bank steps in by selling reserves to 87 maintain exchange rate stability. Table 2.8 reports coecient estimates. The rst two rows consider the full spectrum of exchange rate exibility including pure oat and pure pegs (1,926 full sample observa- tions) while the second two rows are considering only intermediate exchange rate regimes ( ^ W b it 2 (0; 1)) (1,330 full sample observations) to assure that the results aren't driven by corner policies. Reserves seem to be more sensitive to U.S. shocks than E.U. shocks, with the latter not statistically signicant across sub-samples. This is consistent, and may be associated with the role of the U.S. Dollar making up a majority of reserve assets and ex- change rate pegs. The signicant negative coecients on U.S. monetary shocks suggest that countries tend to reduce international reserves in response to a U.S. tightening, possibly to stabilize the exchange rate and prevent excessive depreciation. This eect strengthens in peg intensity, and is particularly signicant among emerging markets, consistent with previous studies. The eects become more pronounced when considering the sub-sample of interme- diate exchange rate regimes, with E.U. shocks turning negative and economically signicant (but not statistically signicant) for advanced economies. For emerging markets under inter- mediate pegs to the U.S. Dollar, a coecient of -4.12 implies that under a strongly managed peg (peg intensity of 0.80), a 1 percentage point U.S. interest rate shock is associated with a reduction of international reserves equal to [+1% x -4.12 x 0.80] = -3.2%. The signicant response of international reserves to monetary shocks in emerging mar- kets, which is particularly strong under intermediate exchange rate regimes, provides some evidence supporting their role in relaxing the policy Trilemma, thereby enabling a non-linear trade o between exchange rate stability and monetary autonomy. 2.8.2 Limits to international arbitrage A second mechanism that may produce a non-linear trade o between monetary autonomy and exchange rate stability is if there exists costly frictions which inhibit the free ow of capital (e.g. transaction costs, intermediation fees, illiquidity), thereby violating the UIP 88 condition (Fama [1984], Engel [1996], Bansal and Dahlquist [2000]). In the presence of such frictions, interest rate dierentials between two pegged countries can persist, only to be arbitraged when the dierential widens enough to compensate the investor for the associated costs. This suggests that monetary policy spillovers should not just be an increasing function in a) nancial openness and b) exchange rate rigidity, but also the c) interest rate dierential between the base country and countryi. In other words, when the interest rate dierential is small, countryi has more monetary autonomy, therefore the pass-through of a U.S. monetary policy shock should be smaller, than when the interest rate dierential is large (all else xed). I test for evidence consistent with this hypothesis with a simple extension to the baseline regression (Equation 2.6): r it = i + 1 r i;t1 + 2 y it + 3 it + 4 RER it + 5 VIX t + 6 R t + US [ ^ Z US;t ^ W $ it K it jr i;t1 r US;t1 j] + EU [ ^ Z EU;t ^ W e it K it jr i;t1 r EU;t1 j] + it ; (2.11) where the monetary policy shock instrument [ ^ Z US;t ^ W b it K it ] is further interacted with the absolute lagged interest rate dierential,jr i;t1 r b;t1 j. Under this specication, a positive estimate on ^ b implies that for a given degree of exchange rate exibility and nancial openness, monetary policy spillovers will be larger when interest rate dierentials are wider. Table 2.9 reports estimates of ^ b , testing whether the interest rate dierential in uences monetary policy transmission. While broadly, coecient estimates are positive (consistent with limits to arbitrage), statistical signicance varies. The strongest evidence supporting limits to arbitrage is present in advanced economies targeting the Euro (estimate of 0.317), and this eect is robust for the sub-sample of intermediate pegs (estimate of 0.291). A sig- 89 nicant eect of interest rate dierentials on monetary pass through is also seen in emerging markets targeting the USD, however, this eect is driven by corner policies (namely emerging markets under xed or oating exchange rate regimes). For advanced economies targeting the Euro, the limits to arbitrage mechanism shows the strongest evidence of driving a non-linear exchange rate regime-monetary autonomy trade-o. Overall however, evidence of a limits to arbitrage friction is weaker than the evidence supporting the role of active reserves management. The use of international re- serves, specically among emerging markets, may be an important factor allowing countries to `lean against' the Trilemma constraint, corroborating Aizenman et al. [2010]. Thus, active use of international reserves results in what appears to be exchange rate stability, without necessarily losing monetary autonomy. 2.9 Robustness 2.9.1 Long-Run Monetary Policy Adjustment The main analysis focuses on short-run associations between countryi and the base country's monetary policy, while Shambaugh [2004] highlight the possibility of long-run adjustment in the policy rate which might also depend on the Trilemma conguration. That is, even if policy rates across countries respond immediately to one another, it's also possible for countryi's interest rate to be increasingly cointegrated with the base country's interest rate as peg intensity rises, so interest rate adjustment occurs over both the short-run and over a longer period of time. 33 To test for this, I extend Equation 2.6 to include two error-correcting terms: a cointegrating vector between country i's interest rate and the base country (U.S. and E.U. interest rates, respectively), interacted with peg intensity and capital openness: (r i;t1 C b r b;t1 ) ^ W b i;t1 K i;t1 : (2.12) 33 This could be due to various nancial market imperfections or practical limits to arbitrage. 90 Typically one estimates C b in a rst-stage, but I pre-set C b = 1, eectively dening the cointegrating vector as the interest rate dierential between country i and base country b. 34 A negative coecient on this term implies that when country i's interest rate exceeds the base country's, it will induce adjustment in the policy rate to catch down to the base country's. The interaction with peg intensity allows the rate of reversion to strengthen with peg intensity as expected under the policy Trilemma. The interaction with capital openness allows for comparison across countries with identical openness yet diering peg intensities. Table 2.10 reports long-run spillover eects. 35 Short run estimates are included to verify that they are not sensitive to the inclusion of error-correction terms. Across the sample, there is evidence of longer-run adjustment in country i's interest rate to both base coun- tries E.U. and U.S. which increases in country i's peg intensity to either base country. The negative coecient sign is theoretically consistent: when the interest rate dierential is pos- itive (negative), country i's policy rate adjusts in the direction of the base country interest rate. When stratifying the sample into advanced and emerging market sub-samples, it's the emerging markets which exhibit evidence of statistically signicant error-correction in their policy rates under both U.S. and E.U. pegs, while advanced economies generally only exhibit evidence of strong short-run monetary spillovers. If the sample is limited to only intermedi- ate pegs ( ^ W b it 2 (0; 1)), the long-run eect against E.U. peg intensity turns signicant at the 1% level while the long-run eect vis-a-vis the U.S. turns insignicant, 36 precisely matching patterns in short-run eects for emerging markets under intermediate pegs, thereby support- ive of potentially non-linear policy trade-os between exchange rate stability and monetary autonomy. Given the high rate of short-run pass-through among advanced economies, it is plausible that base country monetary policy spillovers occur rather quickly and to their full extent 34 Constraining the cointegrating vector to the interest rate dierential by setting C b = 1 is theoretically consistent with UIP. 35 Robust standard errors clustered at the Country level. Regression specication of Equation 2.6 plus error correction terms (Equation 2.12). Estimation period: Q2 2000 - Q4 2018. Peg intensity used: ^ W b it (RA, 2). 36 This result is not reported in Table 2.10. 91 among these countries. The signicant long-run adjustment among emerging markets at least in part, may explain their relatively weak and imperfect short run pass-through, suggesting that across emerging markets the monetary spillover from base countries may take longer. These results are consistent with the fact that emerging markets are considerably less na- nially developed and host to generally weaker institutions { both factors potentially inducing greater nancial market frictions compared to their advanced economy counterparts. 2.9.2 Alternative Measures of Exchange Rate Flexibility As a robustness check, I also consider the ne exchange rate regime classications of Ilzetzki et al. [2019] (IRR). For this exercise, I only consider U.S. shocks rather than both U.S. and E.U. shocks since the construction of the IRR data doesn't consider de facto basket anchors. The IRR exchange rate regime data, which are monthly, are aggregated to quarterly averages. There are ve levels: Floating, Weak Managed Float, Moderate Managed Float, Strong Managed Float, and Fixed (U.S. is the anchor currency). Denote them: 1, 2, 3 and 4 and 5, respectively. The original IRR ne classication contains 15 dierent regimes. I consolidate levels 2 through 13 into the respective bins described in Table 2.11. Using this alternative exchange rate regime classication, I test for evidence of non-linear monetary policy transmission (with respect to the exchange rate regime). The regression specication used is the same as Equation 2.6, but with only U.S. shocks, and now the discrete IRR exchange rate regimes: R it = i + 1 R i;t1 + 2 y it + 3 it + 4 RER it + 5 VIX t + 6 R t + US (IRR)[ ^ Z US;t D(IRR) $ it K it ] + it : (2.13) The coecient ^ US (IRR) represents the spillover coecients across the ve dierent IRR 92 exchange rate regime classications. If the estimates are not signicantly and/or monotoni- cally increasing in exchange rate rigidity, the story of non-linear monetary spillovers remains consistent with the primary analysis. Table B.5 reports estimates of ^ US (IRR) across all countries, advanced economies, and emerging markets. The general pattern persists: under more rigid exchange rates (3, 4 and 5), there is disproportionately less monetary independence. The hard peg (bin 5) estimates, interestingly, are statistically insignicant for emerging markets, but highly rigid oats (bin 4) are indeed signicant and subject to high monetary pass-through (estimate of 0.862). Across all three groups of countries, pass-through under free oating regimes less than 0.20 (but statistically signicant among advanced economies), suggesting considerable monetary independence from the U.S. under a oating exchange rate. To summarize, under a dierent measure of exchange rate regime, the results of monetary pass-through tend to be consistent with the baseline analysis. In addition, the robustness check conrms suggestive non-linearities in most cases, where dierent degrees of exibility within intermediate exchange rate regimes indicate disproportionate gains/losses in monetary autonomy. 2.9.3 Accounting for the Zero Lower Bound As mentioned in Section 2.5, both the U.S. and the E.U. saw prolonged episodes where the policy rate was pinned to the eective lower bound. The baseline specication treats these episodes as having little to no variation in monetary policy. Despite this, the use of unconventional policy tools were widespread in both countries, and therefore, it's important to allow for variation in monetary conditions which may not be directly observable through the policy rate. To do this, I take U.S. and E.U. shadow policy rates (Wu and Xia [2016]), which replace actual policy rates pinned to the ZLB with model-implied shadow rates. After the global nancial crisis, while observed policy rates were at near-zero, unprecedented levels of monetary easing drove shadow rates into negative territory. 93 The approach is simple, I replace actual policy ratesR bt with the shadow rate value, ifR bt is at the eective lower bound. 37 Then, I recompute the residual monetary shock ^ Z bt from the series of R bt spliced with shadow rates. The results after augmenting policy shocks with shadow rates are reported in Table B.6, and are largely consistent with the baseline analysis. In the full sample, signicant evidence of the spillovers under the Trilemma continues to be present, and monetary policy pass-through strengthens among the advanced economy sub- sample. For emerging markets, there is no evidence of spillovers under a EUR target, but there is signicant, albeit weaker evidence of spillovers under a USD target. These results mirror those found under the baseline analysis. 2.9.4 Unanticipated U.S. Monetary Policy Shocks It's very possible that residual changes in interest rates ^ Z US;t and ^ Z EU;t , used as interest rate `shocks' are still containing endogenous movements related to omitted or unobserved expectations and macroeconomic forces. As an additional robustness check, I replace ^ Z US;t with identied U.S. monetary policy shocks, exploiting the movement in Fed Fund futures contracts around FOMC announcements (Kuttner [2001]). 38 The slight alteration to the baseline regression then yields the following specication: R it = i + 1 R i;t1 + 2 y it + 3 it + 4 RER it + 5 VIX t + 6 R t + US [FFS US;t ^ W $ it K it ] + EU [ ^ Z EU;t ^ W e it K it ] + it : (2.14) Notice that the only alteration is that U.S. interest rate residuals ^ Z US;t are replaced with 37 This is precisely how the shadow rate is dened. 38 Bluedorn and Bowdler [2010] replace changes to U.S. interest rates with these `Fed Funds shocks' to test the Trilemma, reporting highly signicant results and near complete monetary pass-through to pegged countries. 94 Fed Fund shocks FFS US;t . These shocks are computed by taking the change in the front- month Fed Funds futures contract over the day of a scheduled FOMC meeting. Then, these daily changes are aggregated to the quarterly frequency. 39 Table B.7 reports the baseline spillover estimates, but now with Fed Funds shocks replac- ing the U.S. interest rate residual. Consistent with Bluedorn and Bowdler [2010], estimates across the full sample, advanced economies, and emerging markets all suggest ^ US = 1 within 95% condence bands, suggesting approximate 1-for-1 U.S. interest rate pass-through un- der open capital ows and a xed exchange rate. The full country sample and advanced economy sub-sample estimates are statistically signicant at the 1% level (estimates of 0.944 and 1.049, respectively), while the emerging market estimate of ^ US using FFS US;t is sta- tistically signicant at the 11% level (estimate of 0.867). Overall estimates of monetary pass-through under continuous exchange rate regime measures are robust to using either actual or unanticipated changes in U.S. monetary policy. 2.9.5 Omitting 2008-2010 Global Financial Crisis I omit Q1 2008 - Q4 2010 and re-estimate the baseline regression (Equation 2.6) to infer to what degree the 2008 Global Financial Crisis may be driving estimates of monetary pass-through. It's the conventional view that over this period, global factors were driving synchronized uctuations in real activity and nancial volatility across countries. Therefore it may be possible that correlations between monetary policy of dierent countries were actually responding to domestic conditions which happened to be synchronized. Table B.8 reports the results of the baseline tests (Equation 2.6) after omitting the crisis period, Q1 2008 - Q4 2010. Across all countries, advanced economies, and emerging markets, the pass-through eects remain robust to omitting the crisis period. In fact, the pass-through eects on both U.S. and E.U. coecients rise in the `all country' sample after omitting the crisis period (to 0.522 and 0.398, respectively). Across advanced countries, 39 There is no severe serial correlation generated through aggregation. Unit root tests on the quarterly FF shock series reject the null of a unit root. 95 spillover estimates remain stable and highly signicant. The pass-through coecient for emerging markets rises considerably (to 0.474) after omitting the crisis period. The evidence of intermediate exchange rate regimes aecting the pass-through of monetary policy remains a highly robust feature of the data, insensitive to the Global Financial Crisis. 2.9.6 Time-varying SDR Basket Weights Eective October 2016, the IMF added the Chinese Yuan as an additional currency in the SDR basket. As of that date, the currencies and corresponding weights were U.S. dollar 41.73%, euro 30.93%, renminbi (Chinese yuan) 10.92%, Japanese yen 8.33%, British pound 8.09%. Due to the time-varying nature of SDR component weights, it's possible that our peg intensity measures, and spillover estimates are sensitive to abrupt changes in SDR composi- tion. As a simple check to assess whether the overall results are sensitive to SDR rebalancing, I estimate the baseline regressions over the pre-2016 sample period, before the Yuan was in- troduced as an SDR component. Results are reported in Table B.9. Overall, the results from the pre-Yuan estimation very closely match the baseline results estimated over the entire sample period. 2.9.7 Including lower-dimension interaction terms Our baseline equations for testing monetary policy spillovers include the interaction ^ Z b;t ^ W b it K it but no lower-dimension interactions of these covariates nor do they enter indi- vidually. It may be of interest to see if any additional insight may be provided under the specication which includes all lower-dimension terms: 96 R it = i + 1 R i;t1 + 2 y it + 3 it + 4 RER it + 5 VIX t + 6 R t + X b2US;EU ( 1b ^ Z b;t + 2b ^ W b it ) + 3 K it + X b2US;EU ( 4b [ ^ Z b;t ^ W b it ] + 5b [ ^ Z b;t K it ] + 6b [ ^ W b it K it ]) + X b2US;EU 7b [ ^ Z b;t ^ W b it K it ] + it : (2.15) Under this expanded specication, the impact of a 100 basis point country b monetary policy shock ( ^ Z b;t ) on country i's policy rate would be equal to 1b + 4b ^ W b it + 5b K it + 7b [ ^ W b it K it ]: (2.16) For instance, a country with a xed exchange rate and open capital account would have a spillover coecient of 1b + 4b + 5b + 7b . While there is a structural interpretation of the instrument ^ Z b;t ^ W b it K it , it does not necessarily follow that including the lower-dimension terms is theoretically appropriate. Therefore this exercise is mainly exploratory. Table B.10 reports the estimates for all coecients linked to U.S. and E.U. monetary policy shocks (^ 1b , ^ 4b , ^ 5b , ^ 7b ). Interestingly, the at marginal eect of a U.S. monetary policy shock given by ^ 1;US is negative for all three sub-samples (all countries, advanced economies, emerging markets). Consistent with the Trilemma conditions, U.S. monetary spillovers are increasing in peg intensity (^ 4;US ) and capital account openness (^ 5;US ) while the estimate on the three-way interaction, ^ Z b;t ^ W b it K it , given by ^ 7;US is statistically insignicant across all three sub-samples, with a positive estimate for the advanced economy sub-sample and a negative estimate on the emerging market sub-sample. As with the baseline results, the eects of E.U. monetary spillovers are not statistically signicant in this extended 97 specication. 2.10 Concluding Remarks In this study, I investigate monetary policy spillovers under the Trilemma with a particu- lar focus on intermediate exchange rates. Specically, I test empirically the shape of the Trilemma, which often assumes a linear trade-o between exchange rate stability and mon- etary autonomy. To address this issue, I propose a continuous de facto measure of exchange rate regime which considers the entire spectrum of exchange rate exibility. I test and nd signicant evidence of a non-linear Trilemma, such that gains in exchange rate stability may not come with a proportionate loss in monetary autonomy along some parts of the peg in- tensity spectrum. Moreover, I show some evidence suggesting that for emerging markets, active reserves management may be generating these empirical non-linearities. Gains in monetary autonomy from this non-linear trade o are allocated dierently across advanced economies and emerging markets. Advanced economies tend to put greater emphasis on output stabilization while emerging markets focus on in ation. However, emerging market monetary policy also becomes increasingly vulnerable global nancial shocks as they move towards more exible exchange rates. I also draw implications for monetary policy spillovers under basket pegs, showing that targeting multiple exchange rates may help diversify against foreign interest rate shocks. The fact that the Two-Corners hypothesis has been continuously rejected, combined with the scarcity of pure oats, suggests that the de facto dominance of intermediate exchange rate regimes is here to stay. This paper's ndings, specically those suggesting a non-linear Trilemma trade-o concerning monetary independence, may provide one possible explana- tion as to why the majority of countries consistently choose middle-ground exchange rate policies. To bolster this argument, future research includes developing a simple model which investigates under what conditions some exchange rate stabilization may be optimal in mini- 98 mizing a central bank's loss function based on domestic targets. The solution will depend on both the sensitivity of domestic economic activity to real exchange rate uctuations and to domestic policy rate changes, both of which depend on the pass-through of foreign monetary policy. 99 Table 2.5: Spillover Eects across Peg Intensity Bins: All Countries Bin 1 2 3 4 5 6 ^ W US it (RA, 2) [0,0.1] (0.1,.30] (0.30,.50] (0.50,0.70] (0.70,0.90] (0.90,1] it 0.094*** 0.115*** 0.093*** 0.056 0.170*** 0.014* (0.022) (0.043) (0.015) (0.044) (0.025) (0.007) RER it 0.015 -0.009 0.003 0.017** -0.009 0.037** (0.011) (0.007) (0.007) (0.008) (0.007) (0.018) y it 0.029** 0.014 0.018 0.023 0.004 0.005 (0.012) (0.012) (0.015) (0.017) (0.006) (0.006) VIX t 0.165** 0.302** 0.027 0.143 0.173* -0.138** (0.070) (0.118) (0.074) (0.130) (0.088) (0.060) R t -0.048 0.186 0.045 0.157* 0.023 0.020 (0.082) (0.121) (0.072) (0.088) (0.034) (0.045) ^ US 0.001 -0.010 0.142 0.276** 0.207*** 0.482*** (0.008) (0.092) (0.150) (0.139) (0.067) (0.132) Adj. R 2 0.01 -0.01 0.05 0.03 0.24 0.03 F-Statistic 7.16 6.15 10.39 8.91 26.68 9.84 NT 385 356 409 356 389 684 ***,**,* refer to signicance at the 1%, 5% and 10% level, respectively. Robust standard errors clustered at the Country level. Regression specication of Equation 2.7. Estimation period Q2 2000 - Q4 2018. Country Fixed Eects included. Within adjusted R-squared reported. . 100 Table 2.6: Spillover Eects across Peg Intensity Bins: Advanced Economies Bin 1 2 3 4 5 6 ^ W US it (RA, 2) [0,0.1] (0.1,.30] (0.30,.50] (0.50,0.70] (0.70,0.90] (0.90,1] it 0.009 0.115** 0.101** 0.003 -0.099 0.092*** (0.024) (0.047) (0.049) (0.031) (0.068) (0.012) y it 0.050*** 0.049*** 0.009 0.016*** 0.039*** 0.010 (0.017) (0.017) (0.013) (0.006) (0.009) (0.022) RER it -0.001 -0.013 0.002 -0.015*** 0.024 0.011 (0.003) (0.009) (0.007) (0.005) (0.015) (0.025) VIX t -0.093 0.062 -0.042 0.132* 0.043 -0.121 (0.069) (0.091) (0.054) (0.066) (0.119) (0.107) R t 0.056** 0.129*** 0.043 0.003 -0.037** -0.031 (0.025) (0.042) (0.045) (0.033) (0.017) (0.039) ^ US 0.060* 0.012 0.137** 0.433*** 0.616*** 1.021*** (0.035) (0.065) (0.057) (0.030) (0.106) (0.115) Adj. R 2 0.17 0.28 0.01 0.09 0.30 0.56 F-Statistic 8.71 10.73 2.51 3.05 4.54 19.03 NT 167 130 100 50 37 84 ***,**,* refer to signicance at the 1%, 5% and 10% level, respectively. Robust standard errors clustered at the Country level. Regression specication of Equation 2.7. Estimation period Q2 2000 - Q4 2018. Country Fixed Eects included. Advanced Economy sub-sample only. Within adjusted R-squared reported. . 101 Table 2.7: Spillover Eects across Peg Intensity Bins: Emerging Markets Bin 1 2 3 4 5 6 ^ W US it (RA, 2) [0,0.1] (0.1,.30] (0.30,.50] (0.50,0.70] (0.70,0.90] (0.90,1] it 0.101*** 0.109** 0.093*** 0.058 0.172*** 0.014** (0.022) (0.051) (0.015) (0.045) (0.026) (0.007) y it 0.026* 0.005 0.020 0.023 0.003 0.005 (0.015) (0.015) (0.018) (0.018) (0.007) (0.006) RER it 0.026 -0.011 0.004 0.020** -0.010 0.040** (0.018) (0.008) (0.010) (0.008) (0.007) (0.018) VIX t 0.296*** 0.438*** 0.060 0.151 0.175* -0.133** (0.082) (0.155) (0.095) (0.154) (0.098) (0.067) R t -0.122 0.223 0.047 0.232** 0.039 0.025 (0.146) (0.234) (0.105) (0.110) (0.041) (0.050) ^ US -0.140 0.051 0.1490 0.210 0.134** 0.367*** (0.214) (0.439) (0.333) (0.187) (0.060) (0.113) Adj. R 2 0.00 -0.07 0.04 0.06 0.26 0.01 F-Statistic 4.99 3.04 7.91 8.67 25.65 6.42 NT 218 226 309 306 352 600 ***,**,* refer to signicance at the 1%, 5% and 10% level, respectively. Robust standard errors clustered at the Country level. Regression specication of Equation 2.7. Estimation period Q2 2000 - Q4 2018. Country Fixed Eects included. Emerging Market sub-sample only. Within adjusted R-squared reported. 102 Table 2.8: International Reserves and Monetary Spillovers Dep. Variable All Advanced Emerging IR it Countries Economies Markets ^ US -2.00** -0.787 -3.296*** (0.815) (1.445) (1.033) ^ EU 1.400 0.456 1.923 (3.630) (6.723) ( 2.102) Excluding Corner Exchange Rate Policies ^ US , ^ W US it 2 (0; 1) -2.079** -1.434 -4.117*** ( 0.958) (1.344) (1.303) ^ EU , ^ W EU it 2 (0; 1) -1.344 -3.001 0.949 (5.065) (7.268) (4.162) ***,**,* refer to signicance at the 1%, 5% and 10% level, respectively. Robust standard errors clustered at the Country level. Estimation period: Q2 2000 - Q4 2018. Peg intensity used: ^ W b it (RA, 2). Table 2.9: International Arbitrage and Monetary Spillovers Dep. Variable All Advanced Emerging r it Countries Economies Markets ^ US 0.042*** 0.182 0.0419*** (0.006) (0.161) (0.007) ^ EU 0.004 0.317** -0.047 (0.054) (0.131) ( 0.060) Excluding Corner Exchange Rate Policies ^ US , ^ W US it 2 (0; 1) 0.013 0.208 -0.007 ( 0.029) (0.164) (0.026) ^ EU , ^ W EU it 2 (0; 1) 0.150** 0.291** 0.108 (0.070) (0.130) (0.084) ***,**,* refer to signicance at the 1%, 5% and 10% level, respectively. Robust standard errors clustered at the Country level. Estimation period: Q2 2000 - Q4 2018. Peg intensity used: ^ W b it (RA, 2). 103 Table 2.10: Short vs. Long-run Monetary Spillovers Dep. Variable All Advanced Emerging r it Countries Economies Markets ^ US 0.398*** 0.783*** 0.283** (0.116) (0.199) (0.112) ^ EU 0.419*** 0.684*** 0.206 (0.142) (0.120) ( 0.239) (r i;t1 r US;t1 ) ^ W US i;t1 K i;t1 -0.022** -0.035 -0.023** ( 0.011) (0.025) (0.011) (r i;t1 r EU;t1 ) ^ W EU i;t1 K i;t1 -0.075*** 0.001 -0.084*** (0.022) (0.031) (0.024) ***,**,* refer to signicance at the 1%, 5% and 10% level, respectively. Robust standard errors clustered at the Country level. Estimation period: Q2 2000 - Q4 2018. Peg intensity used: ^ W b it (RA, 2). Table 2.11: Consolidating Ilzetzki et al. [2019] (IRR) Fine Classications IRR (2019) To 13 1 (Float) 11, 12 2 (Weak Managed) 9, 10 3 (Moderate Managed) 6, 7, 8 4 (Strong Managed) 2, 3, 4, 5 5 (Peg) IRR level 1, 14 and 15 are omitted. They correspond to, respectively: 1: no legal tender, 14: collapsing currency, 15: dual market with missing data. 104 Chapter 3 Global Demand Spillovers and Financial Stability Near the Zero Lower Bound 3.1 Introduction The period following the Global Financial Crisis of 2008 was characterized by unprecedented low interest rates across the developed world, with policy rates of several countries hitting the zero lower bound (ZLB) as shown in Figure 3.1. This posed a problem, as central banks would no longer be able to aggressively counter macroeconomic shocks with expansionary monetary policy. With the onset of the global COVID-19 pandemic in 2020, the number of countries are at or near the ZLB continues to rise, making these challenges ever more relevant. Specically, standard models predict that the macroeconomic impact of shocks become more severe when interest rates are at or near the ZLB, due to the limited capacity of monetary policy to oset these shocks. Not only is monetary easing constrained by the lower bound, shocks which are de ationary would lead to higher rather than lower real interest rates in this environment, adding further to the economic contraction underway. Contrary to the 105 theory, empirical evidence supporting the implications of these models remains inconclusive, thus giving to the rise of the `ZLB irrelevance hypothesis' [Debortoli et al., 2020]. The objective of this paper is to revisit the ZLB irrelevance hypothesis by taking a novel multi- country perspective to address several of the challenges faced in the empirical evaluation of this hypothesis. Figure 3.1: The Median Advanced Economy Policy Interest Rate 0 5 10 15 1980−1 1990−1 2000−1 2010−1 2020−1 Policy Interest Rate (%) Median policy interest rates across 17 advanced economies, quarterly. Shaded region is the inter-quartile range (25th and 75th percentiles). There are several empirical challenges which make it dicult to test the ZLB irrelevance hypothesis in a clean way. First, the hypothesis implies a threshold eect such that the macroeconomic adjustment to shocks becomes severe when interest rates are near the ZLB. However, assuming that this threshold is exactly at zero is overly restrictive. Policy rates can still be positive, but low enough that the suciently large rate cut becomes infeasible. In other words, the interest rate threshold is not known ex ante. Second and most critically, there are substantial endogeneity issues with attempting to identify shocks in the data. The prevailing literature relies on recovering innovations in domestic nancial or macroeconomic aggregates to identify shocks. But the size of these uctuations are endogenously determined 106 by the prevailing monetary policy space. For instance, these approaches fail to account properly for the endogeneous major policy changes occuring at the ZLB associated with major central bank Quantitative Easing policies initiated in the aftermath of the Global Financial Crisis. In a separate context, one can identify shocks more reliably using high- frequency data around macroeconomic news releases, but these studies are limited in testing the short-run eects of shocks mainly on nancial market variables, and cannot adequately test changes in broader and longer-run macroeconomic adjustment near the ZLB. Third, we simply do not have many historical episodes where interest rates are near the ZLB for any given country. For the United States, one has to use data going back to the pre-war era to collect more observations, which introduces additional problems { estimates using longer time-series become prone to structural breaks. Fourth, adding to the latter point, we also do not have many observations of contractionary or recessionary episodes within a low interest rate environment within any single country. This paper addresses these issues by taking a multi-country perspective, considering quarterly data on a panel of 17 advanced economies from 1979 to 2019. The time period and breadth of the data provide valuable variation in both time-series and cross-sectional dimensions, allowing us to increase the number of `near ZLB' episodes and the macroeco- nomic variation occurring near the ZLB. I construct a measure world shocks, proxying for global demand pressure by estimating the common factor of world real commodity prices. For these shocks to be considered exogenous, I rely on a hierarchical assumption { no single country can idiosyncratically in uence global demand, while global demand spillovers aect any single country. Therefore, uctuations in global demand pressure do not endogenously take into account an individual country's monetary space or economic structure, and can be viewed as a largely external shock. Then, I take a pooled maximum-likelihood approach to identify a policy interest rate threshold at which global demand spillovers to domestic equity returns change signicantly. The threshold is not zero, instead estimated to lie between 2.5% and 3.5%, with the likelihood maximized at a threshold of 3.2%. When policy interest rates 107 fall below 3.2%, global demand spillovers to domestic equity returns and equity volatility rise signicantly. This threshold is robust along both time-series and panel dimensions, and persists after accounting for secular time trends, large country eects (i.e. U.S.), economic boom-bust cycles, and is not driven by the post-2008 period when low interest rates were ac- companied by unconventional monetary policy. This threshold also aligns well with the fact that the United States has on average, cut rates 400 to 500 basis points during a recession, which would imply that with rates sitting between 2.5-3.5 percent, the risk of hitting the ZLB given a recessionary shock is high, and the prevailing monetary space is insucient. Finally the analysis is extended to test the ZLB irrelevance hypothesis more completely, by considering the adjustment of aggregate economic activity to global demand spillovers near the ZLB. To do this, I build a threshold-augmented multi-country VAR including both nancial and macroeconomic variables. Global demand shocks are identied recursively under the restriction that global demand spills over contemporaneously to nancial market returns and volatility, but spillovers to real output, in ation, house prices, and central bank policy rates occur with a one quarter lag. Non-linearities enter the VAR system in two ways: First, the lagged response of variables to global demand pressures is allowed to change depending on the policy rate threshold; second, the contemporaneous response of equity returns and volatility to global demand pressures is dependent on the policy rate threshold. Resulting impulse response functions to a 1-standard deviation negative global demand shock show that not just nancial returns and volatility are signicantly impacted, but also real GDP contractions are signicantly deeper when policy rates near the ZLB. When policy rates are above the threshold of 3.2%, the average 5-quarter contraction in real GDP amounts to roughly -0.15%, but the same negative demand shock induces a contraction twice as large, closer to -0.32% ,when interest rates are below the 3.2% threshold. The impact on in ation is also stronger and more de ationary when interest rates are near the ZLB, but unsurprisingly, the response in the policy interest rate is severely truncated. By contrast, policy rates are lowered roughly 40 basis points when rates are not near the ZLB. This results in the larger 108 GDP contraction near the ZLB to be accompanied by a positive change in real interest rates { consistent with a de ationary spiral. These changes in macroeconomic adjustment to shocks near the ZLB are consistent with the predictions of macroeconomic models and reject the hypothesis that the ZLB is irrelevant for the macroeconomy. Christiano et al. [2011] and Eggertsson [2011] argued that shocks are accompanied by a de ationary spiral when interest rates bind at the ZLB, causing the impact on output to become amplied. Basu and Bundick [2017] argue that the ZLB matters, having both rst-order eects and implications for aggregate uncertainty while Gust et al. [2017] use a structural model to show that the U.S. Great Recession of 2008-2009 and the subsequent recovery was signicantly impacted by the lower bound constraint on interest rates. The eectiveness of accommodative monetary policy may itself weaken in a low interest rate environment through expectations, debt overhang, or inherent non-linearities [Borio and Hofmann, 2017], leading to deeper recessions. In the international context which is particularly relevant for this study, Bodenstein et al. [2017] show using a 2-country DSGE model that the domestic response to foreign shocks are greatly amplied by the ZLB. Despite support on theoretical grounds, the empirical evidence has been mixed, with a large majority focusing on the U.S. context. Using high-frequency data, Swanson and Williams [2014a] and Swanson and Williams [2014b] nd that the zero lower bound did not signicantly alter the response of longer-term U.S., U.K., and German yields and exchange rates to macroeconomic news surprises over most periods. By contrast, Datta et al. [2018] show in a high-frequency setting that correlation between oil and U.S. equities strengthened substantially after the 2008 Financial Crisis, attributing this to U.S interest rates hitting the ZLB. This nding is also supported by theoretic rationale of Bodenstein et al. [2013]. When policy rates are constrained at the ZLB, the eect of a de ationary negative demand shock on real interest rates reverses, leading to higher real interest rates when near the ZLB, but lower real interest rates otherwise as the central bank cuts interest rates more than 1-for-1 with in ation. The real interest rate channel, in turn, creates a positive relationship between 109 equity returns and global demand shocks when policy nears the ZLB. My ndings corroborate this view, generalizing to a global setting and to both nancial and macroeconomic dynamics. High-frequency event studies achieve sharper identication in the very short run, but can- not draw longer-run implications for broader macroeconomic eects, which lower-frequency structural VAR analysis achieves (at the cost of stronger identication assumptions). De- bortoli et al. [2020] show that in the United States, both the unconditional volatility of macroeconomic aggregates and their response to shocks did not materially change at the ZLB. Chung et al. [2012] contrast this, using a broad range of structural and statistical models to conclude that the ZLB did have rst-order eects on macroeconomic outcomes even after accounting for unconventional monetary policies. Caggiano et al. [2017] further show using non-linear VAR techniques that uncertainty shocks proxied by the VIX index had a larger impact on the U.S. economy at the ZLB. This paper is most closely aligned with the latter literature using structural VAR tech- niques to evaluate the ZLB irrelevance hypothesis, though departing from it in several ways. My analysis extends the evaluation beyond the U.S. case by considering a panel of 17 ad- vanced economies. Taking a multi-country approach lends much needed statistical power, allows for more general inference and external validation, while also providing a unique set-up which exploits both time-series and cross-sectional variation for identication. I also construct shocks dierently, leveraging world uctuations in commodity prices as a proxy for global demand, taken as given by any individual country. Importantly this measure better serves as an external shock than domestic measures such as nancial conditions or risk premia which themselves are endogenously determined by monetary policy space. Methodologically, I do not assume an interest rate threshold of exactly zero, instead I learn the threshold from the data. Moreover, I test the ZLB irrelevance hypothesis using a threshold-augmented VAR following Chudik et al. [2020], while Debortoli et al. [2020] uses a time-varying coe- cient VAR and Caggiano et al. [2017] estimates an interacted VAR. An important dierence between my approach and the others is that the the threshold-augmented VAR assumes a 110 discrete, instead of smooth, change in the way shocks impact the macroeconomy depending on the policy rate level, which aligns well with structural models. My results provide com- plementary support and extend upon the high-frequency ndings of Datta et al. [2018] who document the structural change between crude oil, an international commodity, and U.S. equity returns near the ZLB. The rest of the paper is organized as follows: Section 3.2 discusses the construction of global demand pressures from world commodity prices. Section 3.3 then explores whether a signicant policy rate threshold exists for the relationship between global demand and do- mestic equity markets. Section 3.4 extends the analysis, estimating a threshold-augmented VAR to test the ZLB irrelevance hypothesis using both nancial and macroeconomic aggre- gates. Section 3.5 concludes. Further details on the data and robustness checks are reported in the Online Supplement. 3.2 Estimating Global Demand Pressure from World Commodity Prices A key empirical challenge to overcome in testing the ZLB irrelevance hypothesis is the endo- geneity of shocks and macroeconomic adjustment at the aggregate, country level. I address this by estimating a measure of global variation in world commodity prices, proxying for global demand pressures. This approach is motivated by evidence that common uctua- tions in commodity prices well proxies for global demand (documented by Delle Chiaie et al. [2018a], Kilian and Zhou [2018a], Alquist et al. [2020] among others). The implicit identication assumption for global demand pressures to be considered ex- ogenous by a country is that no single country on its own in uences global demand. For most small open economies, this assumption is reasonably satised. However, there are ex- ceptions, like China or the United States. To proxy global demand, rst I collect a broad cross-section of 58 commodity prices traded on the world market, and then compute real 111 Figure 3.2: The Common Factor in Real Commodity Price Changes −5.0 −2.5 0.0 2.5 1980−1 1990−1 2000−1 2010−1 2020−1 Standard Deviations Measure is the rst principal component extracted from the cross section of 58 quarterly real commodity price returns, scaled to have mean zero and unit variance. spot commodity returns given by: c kt = ln( C kt C k;t1 ) US;t ; (3.1) where C kt is the USD price for commodity k in quarter t, and US;t is U.S. quarterly in ation over the same period, USt = ln(P US;t =P US;t1 ) where P US;t is the consumer price index of the United States. After scaling each commodity return series to have zero mean and unit variance, I extract the rst principal component as an estimate of the common factor shared across world commodity prices. This is the proxy for global demand uctuations hereby referred to as global demand pressure. The time-series of global demand pressures is shown in Figure 3.2. The sharpest decline in global demand occurded during the 2008 Global Financial Crisis (GFC) and is visually apparent. Interestingly one of the largest increases in global demand pressures followed immediately after, consistent with the broad rally in commodity prices following announcements of unprecedented monetary easing by global monetary authorities. 112 Table C.2 reports the PCA factor loadings on each commodity which makes up global demand pressure. As expected and consistent with related research, the largest loadings are attributed to industrial commodities, like copper, aluminum, rubber, and crude oil. Soybeans and soybean oil also receive large weights, the demand for which are largely driven by emerging markets. Alternative to the PCA-based approach one could. consider a model- free approach which does not rely on information from future periods. A simple yet robust alternative is taking the cross-section average of commodity returns each quarter t. The PCA-based factor and the cross-section average factor are highly correlated (coecient equal to 0.964). Another issue is that not all commodities provide a good proxy for global demand. In Section C.2, I address both of these concerns in a robustness check, by re-constructing the commodity factor using cross-section averages and a subset of commodities which are documented in the literature to be reliable global demand proxies. 3.3 Are Global Demand Spillovers Larger Near the ZLB? 3.3.1 Identifying the interest rate threshold Unfortunately there is no consensus on what interest rate level constitutes a `low' interest rate environment, a level at which the risk of hitting the ZLB becomes material and reshapes the macroeconomic response to shocks. That is, to test the ZLB irrelevance hypothesis, I wish to assess whether the impact of global demand shocks on macroeconomic outcomes become amplied under a risk of hitting the ZLB. I consider equity market returns as a benchmark macroeconomic outcome given its responsiveness to shocks and its reliability as an indicator for a country's economic performance. One approach would be to just consider episodes when the ZLB constraint binds, however this does not capture the fact that the macroeconomy can respond to shocks dierently even 113 before hitting the ZLB so long as there is a risk of hitting the ZLB. Taking a more general approach, I pool data across countries and use a grid search approach to assess whether a universal interest rate threshold can be detected in the data. Consider a simple threshold model measuring the sensitivity of equity market returns to uctuations in global demand: r it = i + 1 y it + 2 it + ( 1 + 2 1[R i;t1 <])g t + it ; (3.2) where r it is the equity market return of countryi over quartert, y it and it are controls for real GDP growth and in ation, respectively, R it is the central bank policy interest rate, and g t measure of global demand uctuations. An unknown interest rate threshold given by dictates the strength of the relationship between uctuations in global demand, g t and equity returns r it . When the interest rate is above , the global demand spillover is given by 1 , and when the interest rate falls below , the spillover is given by 1 + 2 . This formulation easily nests the case of no threshold when 2 is estimated to equal zero. Instead of taking an a priori guess of the threshold , my aim is to identify the threshold from the data, therefore I take a simple approach of estimating Equation 3.2 by running grid search over values of ranging from zero to ten percent (following a methodology similar to Chudik et al. [2020]), selecting the threshold which maximizes the likelihood, or alternatively minimizes the regression standard errors, of the model. In the case that the true threshold is located exactly at the ZLB, this approach nests that possibility, along with the possibility of no threshold. Figure 3.3 shows the regression standard errors (y- axis) over the grid of values from 0 to 20% (x-axis). There is a clear value of which maximizes the likelihood of the model, = 3:2%. It's visually apparent that this threshold is robust in the sense that there are no local minima distributed far away. Rather, the minimum values cluster between the range of 2.5% to 3.5%. The gure also supports the non-linearity of the interest rate threshold. When viewing from right to left, instead of a steadily deceasing regression standard error, there is an abrupt change in direction around. A steadily decreasing function would suggest that global demand spillovers to equity markets 114 increase linearly as interest rates decrease. Figure C.1 charts country-specic interest rate time-series with shaded regions corre- sponding to quarters when the interest rate fell below 3.2%. For the E.U. countries, the interest rates used are short-term rates on the country's respective government bonds up until Q4 1999, with the ECB policy rate used onward from that point. Some countries expe- rienced episodes near the ZLB as early as the 1980's like Japan and Switzerland. Nearly all countries in the sample were near the ZLB after the GFC, or post 2008 period. Given the large degree of clustering in the post-crisis era resulting in limited cross-country variation, we control for the post-crisis period when checking the robustness of the ZLB threshold in subsequent analyses. Figure 3.3: Panel Regression Standard Errors under Varying Policy Rate Thresholds 3.2% Model with No Threshold 10.10 10.15 10.20 10.25 5 10 15 20 Policy Rate Threshold (%) Regression SE Solid horizontal line indicates regression standard errors for model with no policy rate threshold interaction (equal to 10.257). Minimum value equals 10.096 corresponding to threshold of 3.2%. 3.3.2 Time-series and panel evidence of ZLB relevance To further test the relevance of the ZLB for global demand spillovers, I estimate both country- specic and panel threshold regressions with set to 3:2%. Country-specic regressions 115 allow for heterogeneous slope coecients across countries, and any evidence of a signicant threshold eect validates a threshold eect using just variation along the time-series dimen- sion. Meanwhile, panel models pool information from all countries, leading to increased statistical power but at the cost of slope homogeneity. However, the panel approach exploits both the time-series and cross-country dimensions of the data to identify changes in global demand spillovers to equity markets when policy interest rates are near the ZLB. Therefore, these two approaches are complimentary to one another. Individual country equity return regressions are estimated as the following: r it = i + i r i;t1 + 1i y it + 2i it + ( 1i + 2i 1[R i;t1 < 3:2%])g t + it ; (3.3) Table 3.1: Individual Country Threshold Regressions Country i i 1i 2i 1i 2i Australia -0.002 -0.025 0.842 1.305 1.584** -0.566 Austria -0.093 0.164** 1.785* 0.842 1.261 2.301 Belgium 2.287* 0.019 0.547 -1.340 1.522 0.427 Canada 1.059 0.033 1.289 -0.284 0.211 3.444*** Finland 2.630 0.002 1.967* -4.887* 0.824 3.072 France 1.469 0.037 -0.18 0.537 0.848 1.601 Germany 0.748 -0.006 1.375 -0.007 -0.368 3.075* Italy -0.604 0.112 -0.382 1.951** -0.186 3.151 Japan -0.111 0.056 0.403 1.794 -3.330 4.895** Netherlands 1.522 0.052 1.955* -1.352 1.336 0.978 New Zealand 0.612 0.036 1.730* -0.892 0.050 1.613 Norway 2.564 -0.049 -0.475 -0.649 2.901** 2.349 Spain 0.120 -0.099 3.081* -1.054 -0.561 3.956** Sweden 2.090 0.061 0.668 -0.148 -0.398 3.121 Switzerland 1.263 -0.001 2.012** -1.234 -2.601 3.267* United Kingdom 0.579 -0.021 1.191 0.693 -0.178 2.893* United States 2.098** -0.012 1.034 -0.623 -1.700* 4.709*** MG 1.073*** 0.021 1.108*** -0.315 0.071 2.605*** Results from country-specic threshold regressions (Equation 3.3). Mean Group (MG) estimates reported in the bottom row. ***, **, * correspond to 10%, 5%, 1% signicance levels, respectively. 116 where now I include lagged equity returns as a regressor, and all coecients are allowed to be country-specic. Table 3.1 reports country-specic regression coecients which include the key parameter of interest, estimates of 2i the coecient on the threshold term. Mean Group estimates, which are the average of estimates over theN = 17 countries, are reported in the bottom row along with associated standard errors [Pesaran et al., 1999; Chudik and Pesaran, 2019]. The MG estimates imply a signicant positive intercept term which captures the general positive trend observed in global equity prices over the past several decades. Real GDP growth is also signicantly and positively associated with equity returns, with a 1% growth in real output associated with, on average, equity returns of 1.10%. The coecients of interest, those on global demand g t imply a highly signicant change in pass-through to equity markets when interest rates are near zero. In fact, when interest rates are above the threshold of 3.2%, there is no signicant relationship between uctuations in global demand, g t and equity market returns. However, the highly signicant threshold coecient, ^ 2i of 2.605, suggests that when interest rates are near the ZLB, equity market returns are highly sensitive to uctuations in global demand. A 1-standard deviation drop in global demand is associated with roughly 2.6% drop in equity market returns when there is a risk of the ZLB binding. Figure 3.4 charts country-specic sensitivities to global demand uctuations when policy rates are greater than 3.2% (LHS) and when they are below 3.2% (RHS). When policy rates are suciently low, a positive coecient is estimated for every country in the sample. This evidence points toward the zero lower bound being highly relevant in shaping the adjustment to macroeconomic shocks. Moreover, it suggests that nancial markets price ZLB risk. When faced with an exogenous shock, equity market adjustment is amplied since the central bank cannot suciently ease monetary policy in response. Panel regressions with a policy rate threshold exploits both the time-series and cross- section dimension of the data, allowing us to take full advantage of the information across countries. Moreover, it easily allows for robustness checks by controlling for possible con- 117 Figure 3.4: Country-Specic Estimates of Equity Return Sensitivity to Global Demand Above and Below Policy Rate Threshold of 3.2% −2 0 2 JPN CHE USA ESP SWE DEU ITA GBR NZL CAN FIN FRA AUT NLD BEL AUS NOR Coefficient on Global Demand Policy Rate Greater than or Equal to 3.2% 0 1 2 3 4 5 JPN CHE USA ESP SWE DEU ITA GBR NZL CAN FIN FRA AUT NLD BEL AUS NOR Coefficient on Global Demand Policy Rate Less than 3.2% Solid horizontal line indicates regression standard errors for model with no policy rate threshold interaction. Y-axis corresponds to the contemporaneous equity price change (in percent) conditional on a 1-standard deviation increase in global demand pressure. Dashed horizontal line indicates the all-country average. founders which may be correlated with the increasing global demand spillovers near the ZLB. For example, one may wish to control for business cycle boom-busts given the possibility that the transmission of global shocks is regime-driven, depending on the state of the business cycle, which in turn relates to the monetary policy stance. The panel model takes the form: r it = i + 1 y it + 2 it + ( 1 + 2 1[R i;t1 < 3:2%])g t +(X it g t ) + it ; (3.4) where X it includes additional state variables which can possibly alter the transmission of global demand spillovers to equity markets. Specically, I control for episodes of real GDP contractions (y it < 0) with an indicator variable, and I also control for the post Global Financial Crisis (GFC) period with an indicator variable agging the period after Q4 2008, which has been characterized by low interest rates globally and unprecedented levels of unconventional monetary policy. Because Quantitative Easing policies are correlated with the ZLB, an important test is whether global demand spillovers change when policy rates are 118 low during the period in which unconventional policies were absent (pre-GFC). However it should be noted that there still exists the problem of identifying QE policies separately from the post-2008 period, so these results should be considered cursory. 1 Further tests to control for quantitative easing policies beyond excluding the post-2008 period can be undertaken by using information along the entire term structure of interest rates or similarly by considering shadow rates. A signicant 2 estimate after controlling for these factors suggests that global demand spillovers to equity markets are amplied when interest rates are near the ZLB, even in periods before 2008 and after accounting for business cycle volatility. I also include a control for linear time trends to capture the secular decline in interest rates along with an indicator variable for when U.S. interest rates are below 3.2% in X it for robustness. Finally, in the panel regression format, I also test pass-through of global demand uctuations to realized equity volatility, a proxy for economic uncertainty. RV it = i +RV i;t1 + 1 y it + 2 it + 3 r it +( 1 + 2 1[R i;t1 < 3:2%])g t +(X it g t )+ it ; (3.5) where RV it is the realized equity volatility estimated as the standard deviation of daily returns over quarter t for country i. Specically, realized volatility is dened as: RV it = s 1 D(t) 1 X D2t [r id (t) r i (t)] 2 ; (3.6) where r id (t) refers to country i's equity index return on day d in quarter t, r i (t) is the average daily equity return in quarter t, and D(t) is the number of trading days in quarter t. For additional robustness, Equation 3.5 controls for lagged volatility to deal with persistence 2 , and also controls for the contemporaneous equity return. This specically 1 These extensions are a planned work in progress. 2 Adding a lagged dependent variable species Equation 3.5 as a dynamic panel model with xed eects. It is well-known that estimates from this model are biased bias, especially when the time dimension is short. 119 tests for global demand spillovers to the component of equity market volatility which is uncorrelated with returns, or idiosyncratic volatility. Table 3.2: Panel Threshold Regression for Equity Returns Dependent variable: Quarterly Equity Returns (1) (2) (3) (4) (5) (6) (7) (8) Real GDP Growth 1.044 0.975 0.871 0.915 0.905 1.127 0.888 0.892 (0.197) (0.222) (0.189) (0.195) (0.198) (0.128) (0.207) (0.207) In ation 0.365 0.095 0.068 0.077 0.020 0.010 0.058 0.067 (0.362) (0.376) (0.368) (0.367) (0.379) (0.448) (0.397) (0.395) Global Demand 1.346 0.862 1.358 1.451 1.982 2.358 1.969 1.745 (0.257) (0.389) (0.312) (0.304) (0.459) (0.382) (0.499) (0.492) Global Demand Lagged Policy Rate < 3.2% 3.713 3.288 2.905 2.636 3.139 2.717 2.587 (0.357) (0.361) (0.506) (0.569) (0.561) (0.556) (0.589) Global Demand Real GDP Growth < 0% 1.673 1.905 1.779 2.014 1.785 1.640 (0.552) (0.602) (0.617) (0.576) (0.648) (0.611) Global Demand After Q4 2008 0.863 0.632 0.921 0.531 0.527 (0.404) (0.439) (0.166) (0.437) (0.425) Global Demand Linear Time Trend 0.008 0.003 0.008 0.0004 (0.006) (0.006) (0.007) (0.006) Global Demand Lagged U.S. Policy Rate < 3.2% 1.049 (0.597) Country FE Yes Yes Yes Yes Yes Yes Yes Yes Excluding USA No No No No No No Yes Yes Excluding Commodity Exporters No No No No No Yes No No Observations 2,651 2,628 2,628 2,628 2,628 2,039 2,465 2,465 R 2 0.028 0.058 0.064 0.065 0.065 0.069 0.064 0.065 Adjusted R 2 0.021 0.051 0.056 0.057 0.057 0.060 0.055 0.056 *, **, *** correspond to 10%, 5%, and 1% signicance respectively. Driscoll-Kraay standard errors [Driscoll and Kraay, 1998]. Tables 3.2 and 3.3 report panel threshold regression results for equity returns and volatil- ity, respectively. In both tables, the linear eect of global demand spillovers changes sign after including the interaction with the policy rate threshold. This suggests that the correlation between global demand and nancial markets switched signs once interest rates approached the ZLB. Global demand was negatively associated with equity returns when interest rates weren't low. Once interest rates dropped suciently low (i.e. below 3.2%), global demand uctuations became positively associated with equity returns. Similar patterns are seen for equity volatility. These patterns support the ndings of Datta et al. [2018] who document However given the large T nature of this panel (the minimum number of time-series observations a country has in the regression is 116), the bias is negligible [Judson and Owen, 1999]. 120 Table 3.3: Panel Threshold Regression for Realized Equity Volatility Dependent variable: Quarterly Realized Equity Volatility (1) (2) (3) (4) (5) (6) (7) (8) Lagged Volatility 0.557 0.561 0.554 0.543 0.538 0.550 0.542 0.541 (0.031) (0.031) (0.031) (0.032) (0.033) (0.042) (0.033) (0.033) Equity Returns 0.309 0.298 0.291 0.292 0.291 0.276 0.285 0.286 (0.024) (0.027) (0.028) (0.028) (0.027) (0.031) (0.026) (0.026) Real GDP Growth 0.136 0.145 0.242 0.290 0.297 0.310 0.332 0.332 (0.214) (0.202) (0.207) (0.197) (0.197) (0.249) (0.191) (0.191) In ation 0.326 0.166 0.137 0.136 0.010 0.188 0.029 0.032 (0.172) (0.155) (0.158) (0.157) (0.150) (0.224) (0.152) (0.153) Global Demand 1.092 0.073 0.505 0.357 1.595 1.328 1.548 1.608 (0.118) (0.147) (0.135) (0.159) (0.248) (0.206) (0.258) (0.217) Global Demand Lagged Policy Rate < 3.2% 1.785 1.320 1.900 1.309 0.991 1.315 1.349 (0.227) (0.207) (0.320) (0.333) (0.141) (0.347) (0.360) Global Demand Real GDP Growth < 0% 1.928 1.577 1.296 1.749 1.302 1.340 (0.339) (0.369) (0.396) (0.107) (0.415) (0.440) Global Demand After Q4 2008 1.335 1.873 1.319 1.840 1.842 (0.350) (0.369) (0.258) (0.383) (0.384) Global Demand Linear Time Trend 0.018 0.014 0.018 0.020 (0.003) (0.002) (0.003) (0.002) Global Demand Lagged U.S. Policy Rate < 3.2% 0.276 (0.300) Country FE Yes Yes Yes Yes Yes Yes Yes Yes Excluding USA No No No No No No Yes Yes Excluding Commodity Exporters No No No No No Yes No No Observations 2,634 2,612 2,612 2,612 2,612 2,026 2,450 2,450 R 2 0.513 0.530 0.544 0.547 0.551 0.559 0.549 0.549 Adjusted R 2 0.509 0.526 0.539 0.543 0.547 0.554 0.544 0.544 *, **, *** correspond to 10%, 5%, and 1% signicance respectively. Driscoll-Kraay standard errors [Driscoll and Kraay, 1998]. 121 Table 3.4: Panel Threshold Regression for Real Output Growth Dependent variable: Quarterly Real Output Growth (1) (2) (3) (4) (5) (6) (7) (8) Lagged Real GDP Growth 0.067 0.063 0.048 0.047 0.045 0.059 0.035 0.034 (0.059) (0.061) (0.059) (0.060) (0.061) (0.049) (0.061) (0.062) Lagged Global Demand 0.183 0.052 0.00001 0.008 0.116 0.098 0.105 0.068 (0.019) (0.018) (0.022) (0.030) (0.029) (0.029) (0.031) (0.034) Lagged Global Demand Lagged Policy Rate < 3:2% 0.208 0.142 0.284 0.351 0.422 0.363 0.382 (0.038) (0.038) (0.056) (0.066) (0.042) (0.071) (0.080) Lagged Global Demand Lagged Real GDP Growth < 0% 0.206 0.225 0.227 0.252 0.215 0.237 (0.037) (0.031) (0.030) (0.017) (0.035) (0.034) Lagged Global Demand After Q4 2008 0.206 0.150 0.184 0.157 0.135 (0.079) (0.084) (0.101) (0.092) (0.092) Lagged Global Demand Linear Time Trend 0.002 0.002 0.002 0.0003 (0.0004) (0.001) (0.0005) (0.001) Lagged Global Demand Lagged U.S. Policy Rate < 3:2% 0.186 (0.052) Country FE Yes Yes Yes Yes Yes Yes Yes Yes Excluding USA No No No No No No Yes Yes Excluding Commodity Exporters No No No No No Yes No No Observations 2,754 2,706 2,706 2,706 2,706 2,106 2,544 2,544 R 2 0.053 0.067 0.080 0.087 0.090 0.110 0.088 0.091 Adjusted R 2 0.047 0.061 0.073 0.079 0.082 0.102 0.081 0.083 *, **, *** correspond to 10%, 5%, and 1% signicance respectively. Driscoll-Kraay standard errors [Driscoll and Kraay, 1998]. 122 that the correlation between U.S. stock returns and crude oil prices { a commodity which is sensitive to global demand conditions { were slightly negative before before the U.S. zero lower bound period, and then became large and positive afterward. These results go beyond the U.S. context and hold across countries, further validating that this change in global demand spillover to equity markets is driven by a low interest rate environment. After controlling for GDP boom-busts and the post 2008 period, the estimates of 2 remain highly signicant in both the equity return and volatility regressions, and the eect size of the ZLB eect is quantitatively larger than additional global spillovers during GDP busts and the post 2008 period. For equity returns, a 1-standard deviation drop in global demand is associated with a 1.475% drop in equity prices when policy rates are less than = 3:2%, while the same 1-standard deviation drop is associated with 0.50% drop in equity prices under negative real GDP growth, and a 0.73% increase in equity returns in the post 2008 period. However, these marginal eects, going in the same direction, are additive suggesting that global demand spillovers are particularly high when these three states coincide. Despite the benet of high variance in equity market variables, they are subject to in- uence from more than growth expectations { for example, risk-taking appetite and market liquidity, and thus may not always be a reliable proxy for economic activity. Therefore, Table 3.4 reports similar results using real GDP growth as the dependent variable. While there is less variability in real output, it serves as a direct measure of economic performance. A key dierence between the real GDP growth results is that the RHS variables are lagged by one quarter due to the slow-adjusting nature of output compared to asset prices. Regardless, all of the previous results follow through for real GDP: below the policy rate threshold, global demand spillovers to real output become signicantly larger. An important heterogeneous association is also uncovered, where the interaction with the Post-2008 period { which includes the QE policy regime { is associated with signicantly dampening the transmission of global demand spillovers to nancial market volatility (Table 123 3.3). The coecient size indicates that the Post-2008 dampening osets the amplication of demand spillovers to volatility near the ZLB. By contrast, the Post-2008 interaction has a much smaller and insignicant dampening on the transmission of global demand to real GDP growth (Table 3.4). If the Post-2008 indicator is capturing the eects of unconventional monetary policies, it suggests that QE policies may have more eectively reduced nancial market volatility, with a more limited direct impact on real activity. Although it is very possible that QE aected GDP indirectly through reducing volatility, something these base- line regressions do not necessarily capture. All said, these results should be taken as highly suggestive as the Post-2008 indicator does not sharply identify QE policies. In all sets of results, the threshold eect of approaching the ZLB persists after controlling for linear time trends, pointing out that the result is not largely driven by the secular decline in interest rates. Moreover, the results are robust to excluding the United States from the sample, an economically dominant country who's equity market activity may be endogenous with global demand and the results hold after excluding commodity exporting countries. 3 The results in countries excluding the United States are also not driven by low U.S. policy rates, rather low domestic policy rates. Additional issue left for future work is to account for the dramatic rise of China over the past 40 years. China being a granular economy and a major source of consumption like the United States, could very well shape the transmission of global demand spillovers across the world. Relatedly, I later show that global demand spillovers are highly correlated with China's real economic activity. 3.4 Macroeconomic Adjustment Near the ZLB While nancial market adjustment to global demand spillovers is an important indicator of changing economic dynamics near the ZLB, it does not directly provide evidence of chang- ing real economic eects. To fully evaluate the ZLB irrelevance hypothesis, the previous analysis is extended by introducing a threshold-augmented multi-country vector autoregres- 3 These countries are: Australia, Canada, New Zealand, Norway. 124 sion (VAR) in which the global demand shock is identied using order restrictions. The VAR includes both nancial and macroeconomic aggregates, allowing for dynamics to cap- ture potentially asynchronous and persistent macroeconomic adjustment to global demand spillovers. The model is specied as follows: Y it = 0 i Y i;t1 + ( i1 + i2 1[R i;t1 < 3:2%])g t1 +e it ; (3.7) g t = g + g g t1 +e gt ; where Y it = [y it ; it ; hp it ; R it ; r it ;RV it ]: The vector of endogenous variables entering the VAR include variables already intro- duced: real output growth (y it ), in ation ( it ), changes in the policy interest rate, (R it ) equity returns (r it ) and realized volatility (RV it ), but now including changes in the log of real house prices (hp it ). Lagged global demand spillovers g t1 enters the VAR as well, with the eect on Y it depending on whether the policy interest rate is above or below the threshold which is set to 3.2%. Because global spillovers are taken as exogenous, g t follows an AR(1) process, and g t is not in uenced by lagged country-specic variables in Y it (equivalent to a including g t in the VAR but restricting all coecients except its own lag to zero). The reduced form errors are given by e it and e gt . The VAR is allowed to be fully heterogeneous with country-specic coecients to account the fact that countries are likely to response to global innovations dierently because of their distinct economic and policy structures. 125 3.4.1 Estimation and identication The large T dimension of the data allows the model to be estimated country-by-country, estimating country-specic threshold VARs for 17 advanced economies. This estimation procedure is akin to estimating a Global VAR [Pesaran et al., 2004; Chudik and Pesaran, 2016] closely following the threshold-augmented Global VAR approach taken by Chudik et al. [2020] where the authors estimate the impact of COVID-19 induced uncertainty on global economic growth. Non-linear VARs in particular, including threshold-augmented VARs are exible for capturing state-dependencies but pose challenges for estimation and inference. Specically, without imposing restrictions, non-linear VAR estimates often become unstable [Ruge-Murcia, 2014; Caggiano et al., 2017]. Therefore following Caggiano et al. [2017] and Chudik et al. [2020], I consider a parsimonious specication where the interaction operates exclusively through g t1 and I do not interact any other endogenous variables with the interest rate threshold indicator. This choice allows for the possibility of threshold eects in global demand spillovers without risking parameter instability. Structural global demand shocks are identied through recursive ordering using a stan- dard Cholesky decomposition on the reduced form residuals (e it ,e gt ) from each country- specic VAR. Identication assumes that global demand shocks contemporaneously impact equity market returns and volatility, while only aecting real GDP growth, in ation, house prices, and policy rates with a one quarter lag. An additional feature of the model is that I allow the contemporaneous response of equity market returns and volatility to the global demand shock to vary based on the policy rate threshold. Therefore, non-linearities enter the VAR system both through the immediate eect of the shock on nancial market vari- ables and the lagged eect of global demand spillovers on all variables in Y it . Denote igt as the identied structural global demand shock with country i's VAR (recovered from the re- duced form shocke gt ). Then the immediate time 0 response of equity returns and volatility, respectively follow as: 126 e(r) it = ( r i1 + r i2 1[R i;t1 < 3:2%]) igt ; (3.8) e(RV ) it = ( RV i1 + RV i2 1[R i;t1 < 3:2%]) igt ; (3.9) wheree(r) it ande(RV ) it refer to the estimated residuals from country i's VAR (Equa- tion 3.7) for equity returns and realized volatility, respectively, which are contemporaneously impacted by the structural global demand shock, igt . The contemporaneous response to a 1-SD global demand shock takes on a value of ( r i1 ; RV i1 ) when country i's policy rate is above the 3.2% threshold and ( r i1 + r i2 ; RV i1 + RV i2 ) when country i's policy rate falls below 3.2%, for equity returns and volatility respectively. Estimating the average impulse response function (IRF) over the panel is straightforward using the Mean Group (MG) estimator of Pesaran and Smith [1995] and Chudik and Pesaran [2019]. The horizonh mean group, or average, impulse response function for the endogenous variable, denoted Y it , to a 1-SD global demand shock is computed as: MGIRF (h) = 1 N N X i=1 E[Y i;t+h j igt = 1;! t1 ] 1 N N X i=1 E[Y i;t+h j igt = 0;! t1 ] = 1 N N X i=1 E[Y i;t+h j igt = 1;! t1 ]; (3.10) where E[Y i;t+h j igt = 1;! t1 ] is the horizon h impulse response of country i, denoted as the conditional expectation ofY i;t+h given a 1-SD global demand shock, and! t1 denotes the full information set available as of timet 1. The associated non-parametric cross-sectional standard errors computed as: SE(h) = v u u t 1 N 1 N 1 N X i=1 E[Y i;t+h j igt = 1;! t1 ]MGIRF (h) 2 : (3.11) 127 It can be easily seen that the MG IRF is simply the cross-section average of alli country- specic IRFs, each being denoted E[Y i;t+h j igt = 1;! t1 ], at each horizon h. 90% dispersion intervals for each horizon h which I report in the results are equal to MGIRF (h) 1:645SE(h): (3.12) These methods have been applied successfully to large, heterogeneous macroeconomic panel data of similar size to address a variety of research questions. 4 3.4.2 Impulse Response to a Global Demand Shock Figure 3.5 reports the MG IRFs from a 1-standard deviation shock to global demand. The solid lines indicate the response of country-specic variables when the policy rate is above the threshold of 3.2% while the dashed line indicates the response when the policy rate falls below 3.2%. 90% cross-sectional error ands reported. Consistent with the earlier time-series and panel regression evidence, when the policy rate is near the ZLB, both the contemporaneous response in equity returns and volatility are signicantly larger. Specically, the response in equity returns falls from a 0.50% drop to roughly a 2% drop when global demand falls by 1-standard deviation. As expected, the policy interest rate has signicant room to fall when rates are above the threshold, on average falling 40 basis points over ve quarters, while barely moving when policy rates are already near the ZLB. Global demand spillovers are also de ationary, with de ationary forces increasing near the ZLB. Importantly, there is also a large dierence in the real output response. On average, after ve quarters real output growth contracts 0.15% when policy rates are above the threshold, which increases to a contraction of 0.32% when rates are near the ZLB. This larger contraction in real output is accompanied by a positive rather than negative change in the real interest rate following a global demand shock. This results from de ationary forces coupled with the negligible change in the policy rate when near the ZLB. By contrast, in the 4 See for example Chudik et al. [2017], Hernandez-Vega [2019], Cesa-Bianchi et al. [2020b]. 128 Figure 3.5: Impulse Response Functions to a 1-SD Negative Global Demand Shock when Policy Rates are Above (Solid) and Below (Dashed) 3.2% Threshold −0.4 −0.3 −0.2 −0.1 0.0 5 10 Quarters Real GDP (%) −0.3 −0.2 −0.1 0.0 5 10 Quarters Inflation (%) −0.4 −0.2 0.0 0.2 0.4 0.6 5 10 Quarters Real House Price (%) −0.4 −0.2 0.0 5 10 Quarters Policy Rate (%) −4 −2 0 2 5 10 Quarters Equity Return (%) 0.0 0.5 1.0 1.5 2.0 2.5 5 10 Quarters Realized Equity Volatility (%) All IRFs are cumulative except for realized equity volatility. Mean Group IRF estimated from Equation 3.10 along with 90% cross-sectional error bands reported (Equation 3.12). IRFs based on the VAR equations from Equation 3.7. 129 presence of ample monetary space the real interest rate falls in response to a global demand shock as rate cuts tend to exceed the drop in consumer prices. Regardless of the monetary regime, responses in real house prices are statistically indierent from zero. Robustness Further robustness checks are reported in the Online Supplement Section C.2. Specically, the results are robust estimating a measure of global demand spillovers using only a subset of industrial commodity prices such as rubber, coal, oil, and industrial metals. The impulse re- sponses are also reported after estimating the threshold-augmented VAR over the pre-2008 sample when there was no substantial unconventional monetary policies in place. In this shortened sample, three countries do not realize policy rates below the 3.2% threshold (Aus- tralia, New Zealand, United Kingdom). However, the results broadly hold. Importantly, the response of equity returns and real GDP to a negative demand shock are qualitatively similar to those from the full sample estimation, suggesting the post-GFC period, while important, is not the exclusive driver of the non-linearities. The VAR results are also robust to ex- cluding commodity-exporting countries, conrming that the eects are not driven by those countries which are exceptionally sensitive to global commodity prices above and beyond broader global demand eects. 3.4.3 Controlling for other global factors The world commodity factor may be capturing uctuations in economic forces other than global demand, and these omitted variables themselves may impact economic uctations across countries. For instance, g t is signicantly correlated with log changes in the real eective USD exchange rate (REER, correlation of -0.51). The dollar-commodity relationship is well documented, as both tend to respond to similar macroeconomic or nancial shocks. Moreover, world commodities tend to be traded in dollar terms. The U.S. Dollar, an indicator for global liquidity and risk, therefore may be confounded with the eects of global demand 130 pressures. To extract a sharper estimate of the exogenous component of global demand pressures, I condition g t on changes in the USD exchange rate 5 , common uctuations in: advanced economy GDP growth ( y t ), in ation ( t ), and policy rate changes ( R t ), and U.S. GDP growth (y US;t ): g t = g + g g t1 + 2g reer usd;t + 3g y us;t + g y t + 2g t + 3g R t +e gt ; (3.13) where quarter t common uctuations in GDP growth, in ation, and policy rates are approximated with the cross-section average over the set of 17 advanced economies. I then compute the impulse responses using the updated residual e gt from Equation 3.13 which is the component of global demand pressures orthogonal to these other global economic factors. All of the baseline results follow through as shown in Figure 3.6. Notable is that the estimated demand shock, ^ e gt , from Equation 3.13 is signicantly correlated with one-quarter ahead real output growth of China { arguably aside from the United States, one of the few economically dominant countries which drive global demand (Figure C.5). The correlation between 1-quarter lagged global demand shocks and China real output is equal to 0.32, which is roughly three times larger than the correlation with real output of the 17 advanced economies which is on average equal to 0.11. The size and signicance of the correlation is robust when using the baseline measure of global demand spillovers (principal component of commodity prices), and after conditioning China real output growth on its own lagged value. 5 Data on the USD REER are taken from the BIS database. Real eective exchange rates are computed based on trade-weights with 27 countries. Trade-weights are computed on a 3-year rolling average basis using exports and imports. 131 Figure 3.6: Impulse Response Functions to a 1-SD Negative Global Demand Shock after controlling for other Global Factors when Policy Rates are Above (Solid) and Below (Dashed) 3.2% Threshold −0.3 −0.2 −0.1 0.0 5 10 Quarters Real GDP (%) −0.2 −0.1 0.0 5 10 Quarters Inflation (%) −0.25 0.00 0.25 0.50 5 10 Quarters Real House Price (%) −0.4 −0.3 −0.2 −0.1 0.0 5 10 Quarters Policy Rate (%) −4 −2 0 2 5 10 Quarters Equity Return (%) 0.0 0.5 1.0 1.5 2.0 2.5 5 10 Quarters Realized Equity Volatility (%) All IRFs are cumulative except for realized equity volatility. Mean Group IRF estimated from Equation 3.10 along with 90% cross-sectional error bands reported (Equation 3.12). IRFs based on the VAR equations from Equation 3.7. For this set of results, the VAR equation for global demand, g t is Equation 3.13 which controls for contemporaneous changes in the U.S. Dollar REER as a regressor to condition on Dollar uctuations, global uctuations in GDP growth, in ation, monetary policy rates, and U.S. GDP growth when recovering the structural shock. 132 3.4.4 The role of the nancial channel Taking the estimated threshold VAR with demand pressures g t conditioned on global factors as done in the previous section, I next conduct a back-of-the-envelope decomposition to estimate the contribution of nancial market factors in the transmission of global demand shocks to real activity. This excercise is particularly relevant amid the debate on monetary policy trade-os near the ZLB, where low interest rates may stimulate the economy but at the same time promote nancial instabilities. To test the nancial channel contribution, I simply shut down all coecients in the VAR on equity returns and volatility, r it and RV it , respectively, setting them to zero. Addi- tionally, global demand shocks are no longer allowed to contemporaneously aect nancial market variables, therefore transmission of global demand shocks occurs solely through non- nancial market variables in both high and low interest rate regimes, but the coecient estimates on all other variables remain the same. Of course it should be emphasized, this exercise is subject to endogeneity issues as its unlikely that all other VAR coecients remain constant when the coecients on the nancial market variables change, hence this basic exercise should only be referred to as a cursory calculation. 6 Figure 3.7 reports impulse responses after shutting down nancial market VAR coef- cients. Focusing on the response of real GDP to a negative global demand shock when policy rates are greater than the threshold, notice that the cumulative 12-quarter response (-0.11%) is qualitatively similar to that from Figure 3.6 when we leave nancial market co- ecients as is (-0.095%). Meanwhile when near the ZLB, the 12-quarter impact of global demand spillovers on real GDP is roughly -0.14% after shutting down the nancial channels, compared to -0.26%. This back-of-the-envelope computation suggests that nearly half, or 46% of the ZLB amplication of global demand spillovers to real activity transmits through nancial market conditions, and this degree of transmission isn't observed when policy rates 6 A structural model would be necessary to more rigrously address this question. This is left as a future extension. 133 Figure 3.7: Impulse Response Functions to a 1-SD Negative Global Demand Shock after Shutting Down the Financial Channel when Policy Rates are Above (Solid) and Below (Dashed) 3.2% Threshold −0.20 −0.15 −0.10 −0.05 0.00 5 10 Quarters Real GDP (%) −0.3 −0.2 −0.1 0.0 5 10 Quarters Inflation (%) −0.4 −0.2 0.0 0.2 0.4 5 10 Quarters Real House Price (%) −0.3 −0.2 −0.1 0.0 5 10 Quarters Policy Rate (%) −2 −1 0 1 5 10 Quarters Equity Return (%) 0.0 0.5 1.0 5 10 Quarters Realized Equity Volatility (%) All IRFs are cumulative except for realized equity volatility. Mean Group IRF estimated from Equation 3.10 along with 90% cross-sectional error bands reported (Equation 3.12). IRFs based on the VAR equations from Equation 3.7. For this set of results, the VAR equation for global demand, g t is Equation 3.13 which controls for contemporaneous changes in the U.S. Dollar REER as a regressor to condition on Dollar uctuations, global uctuations in GDP growth, in ation, monetary policy rates, and U.S. GDP growth when recovering the structural shock. The nancial channel is shut down in these IRFs by setting all VAR coecients on equity returns and volatility to zero, and not allowing any these variables to respond contemporaneously to a global demand shock. 134 are not near the ZLB. 3.5 Concluding Remarks I empirically evaluate and Zero Lower Bound irrelevance hypothesis from a multi-country perspective, rejecting the view that the impact of shocks on the macroeconomy does not change as policy space nears the ZLB. I construct a measure of external demand pressures from the common factor of world commodity prices, and document a robust threshold such that when policy interest rates fell below 3.2%, the spillover from global demand pressure to domestic equity returns and volatility rose dramatically across a panel of 17 advanced economies. I then extend the analysis using a threshold-augmented VAR, nding that a negative global demand shock induces a signicantly larger contraction in real output when policy rates near the ZLB. This deeper contraction is accompanied by a positive change in real interest rates near the ZLB because the de ationary forces cannot be countered with sucient interest rate cuts. These results support predictions from standard macroeconomic models, and suggest the importance of seeking out new tools for promoting macroeconomic stability in a low-interest rate environment. While this study focuses on the costs associated with the ZLB, complementary research on the eectiveness of alternative policies such as scal, macro-prudential, and quantitative easing policies is worth further exploration, as these policies may oer value when conventional monetary policy capacity is limited and the economy is left particularly vulnerable to shocks. 135 Bibliography Ahmed, R. (2020). Commodity currencies and causality: Some high-frequency evidence. Economics Letters 189, 109016. Ahmed, R. (2021). Monetary policy spillovers under intermediate exchange rate regimes. Journal of International Money and Finance 112, 102342. Ahmed, R., J. Aizenman, and Y. Jinjarak (2019, June). In ation and exchange rate targeting challenges under scal dominance. Working Paper 25996, National Bureau of Economic Research. Aizenman, J. and M. Binici (2016). Exchange market pressure in OECD and emerging economies: Domestic vs. external factors and capital ows in the old and new normal. Journal of International Money and Finance 66 (C), 65{87. Aizenman, J., M. D. Chinn, and H. Ito (2008, December). Assessing the emerging global nancial architecture: Measuring the trilemma's congurations over time. Working Paper 14533, National Bureau of Economic Research. Aizenman, J., M. D. Chinn, and H. Ito (2010, June). The emerging global nancial archi- tecture: Tracing and evaluating new patterns of the trilemma conguration. Journal of International Money and Finance 29 (4), 615{641. Aizenman, J., M. D. Chinn, and H. Ito (2013, August). The \Impossible Trinity" Hypothesis in an Era of Global Imbalances: Measurement and Testing. Review of International Economics 21 (3), 447{458. Aizenman, J., M. D. Chinn, and H. Ito (2016). Monetary policy spillovers and the trilemma in the new normal: Periphery country sensitivity to core country conditions. Journal of International Money and Finance 68, 298 { 330. Aizenman, J., M. Hutchison, and I. Noy (2011, May). In ation Targeting and Real Exchange Rates in Emerging Markets. World Development 39 (5), 712{724. Aizenman, J. and H. Ito (2014, September). The More Divergent, the Better? Lessons on Trilemma Policies and Crises for Asia. Asian Development Review 31 (2), 21{54. Aizenman, J. and J. Lee (2007). International reserves: precautionary versus mercantilist views, theory and evidence. Open Economies Review 18 (2), 191{214. 136 Akinci, O. (2013). Global nancial conditions, country spreads and macroeconomic uctua- tions in emerging countries. Journal of International Economics 91 (2), 358 { 371. Alquist, R., S. Bhattarai, and O. Coibion (2020). Commodity-price comovement and global economic activity. Journal of Monetary Economics 112, 41{56. Angrick, S. (2018, January). Global liquidity and monetary policy autonomy: an examina- tion of open-economy policy constraints. Cambridge Journal of Economics 42 (1), 117{135. Aslam, A., S. Beidas-Strom, R. Bems, O. Celasun, S. K. C elik, and Z. Koczan (2016, Febru- ary). Trading on Their Terms? Commodity Exporters in the Aftermath of the Commodity Boom. IMF Working Papers 2016/027, International Monetary Fund. Avdjiev, S., V. Bruno, C. Koch, and H. S. Shin (2019, March). The Dollar Exchange Rate as a Global Risk Factor: Evidence from Investment. IMF Economic Review 67 (1), 151{173. Baele, L., G. Bekaert, K. Inghelbrecht, and M. Wei (2019, 06). Flights to Safety. The Review of Financial Studies 33 (2), 689{746. Bailey, W. and K. C. Chan (1993). Macroeconomic in uences and the variability of the commodity futures basis. The Journal of Finance 48 (2), 555{573. Bansal, R. and M. Dahlquist (2000). The forward premium puzzle: dierent tales from developed and emerging economies. Journal of international Economics 51 (1), 115{144. Basu, S. and B. Bundick (2017). Uncertainty shocks in a model of eective demand. Econo- metrica 85 (3), 937{958. Beber, A., M. Brandt, and J. Cen (2014, October). Switching Risk O: FX Correlations and Risk Premia. CEPR Discussion Papers 10214, C.E.P.R. Discussion Papers. Beckmann, J., R. L. Czudaj, and V. Arora (2020). The relationship between oil prices and exchange rates: Revisiting theory and evidence. Energy Economics, 104772. Benassy-Quere, A., B. Coeure, and V. Mignon (2006, March). On the identication of de facto currency pegs. Journal of the Japanese and International Economies 20 (1), 112{127. Bernanke, B. S., J. Boivin, and P. Eliasz (2005). Measuring the eects of monetary pol- icy: a factor-augmented vector autoregressive (favar) approach. The Quarterly journal of economics 120 (1), 387{422. Bernanke, B. S., M. Gertler, and S. Gilchrist (1999). The nancial accelerator in a quanti- tative business cycle framework. Handbook of macroeconomics 1, 1341{1393. Bloom, N. (2009). The impact of uncertainty shocks. Econometrica 77 (3), 623{685. Bluedorn, J. C. and C. Bowdler (2010, June). The Empirics of International Monetary Transmission: Identication and the Impossible Trinity. Journal of Money, Credit and Banking 42 (4), 679{713. 137 Bodenstein, M., C. J. Erceg, and L. Guerrieri (2017). The eects of foreign shocks when interest rates are at zero. Canadian Journal of Economics/Revue canadienne d' economique 50 (3), 660{684. Bodenstein, M., L. Guerrieri, and C. J. Gust (2013). Oil shocks and the zero bound on nominal interest rates. Journal of International Money and Finance 32, 941{967. Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics 31 (3), 307 { 327. Borio, C. and B. Hofmann (2017). Is monetary policy less eective when interest rates are persistently low? Caballero, J., A. Fern andez, and J. Park (2019). On corporate borrowing, credit spreads and economic activity in emerging economies: An empirical investigation. Journal of International Economics 118, 160 { 178. Caballero, R. and G. Kamber (2019, March). On the global Impact of risk-o shocks and policy-put frameworks. BIS Working Papers 772, Bank for International Settlements. Caballero, R. J. and A. Simsek (2020a). A model of asset price spirals and aggregate de- mand amplication of a" covid-19" shock. Technical report, National Bureau of Economic Research. Caballero, R. J. and A. Simsek (2020b). A model of ckle capital ows and retrenchment. Journal of Political Economy 128 (6), 2288{2328. Caballero, R. J. and A. Simsek (2020c). A risk-centric model of demand recessions and speculation. The Quarterly Journal of Economics 135 (3), 1493{1566. Caggiano, G., E. Castelnuovo, and G. Pellegrino (2017). Estimating the real eects of uncertainty shocks at the Zero Lower Bound. European Economic Review 100 (C), 257{ 272. Calvo, G. A. and C. M. Reinhart (2002). Fear of oating. The Quarterly Journal of Eco- nomics 117 (2), 379{408. Carstens, A. and H. S. Shin (2019). Emerging markets aren't out of the woods yet. Foreign Aairs. Cesa-Bianchi, A., M. H. Pesaran, and A. Rebucci (2020a). Uncertainty and economic activ- ity: A multicountry perspective. The Review of Financial Studies 33 (8), 3393{3445. Cesa-Bianchi, A., M. H. Pesaran, and A. Rebucci (2020b). Uncertainty and economic activ- ity: A multicountry perspective. The Review of Financial Studies 33 (8), 3393{3445. Chari, A., K. D. Stedman, and C. Lundblad (2020, October). Capital ows in risky times: Risk-on/risk-o and emerging market tail risk. Working Paper 27927, National Bureau of Economic Research. 138 chin Chen, Y. and K. Rogo (2003). Commodity currencies. Journal of International Economics 60 (1), 133 { 160. Emperical Exchange Rate Models. Chinn, M. D. and H. Ito (2006, October). What matters for nancial development? Capital controls, institutions, and interactions. Journal of Development Economics 81 (1), 163{ 192. Christiano, L., M. Eichenbaum, and S. Rebelo (2011). When is the government spending multiplier large? Journal of Political Economy 119 (1), 78{121. Chudik, A., K. Mohaddes, M. H. Pesaran, and M. Raissi (2017, March). Is There a Debt- Threshold Eect on Output Growth? The Review of Economics and Statistics 99 (1), 135{150. Chudik, A., K. Mohaddes, M. H. Pesaran, M. Raissi, and A. Rebucci (2020, September). A counterfactual economic analysis of covid-19 using a threshold augmented multi-country model. Working Paper 27855, National Bureau of Economic Research. Chudik, A. and M. H. Pesaran (2015). Common correlated eects estimation of heteroge- neous dynamic panel data models with weakly exogenous regressors. Journal of Econo- metrics 188 (2), 393{420. Chudik, A. and M. H. Pesaran (2016). Theory and practice of gvar modelling. Journal of Economic Surveys 30 (1), 165{197. Chudik, A. and M. H. Pesaran (2019). Mean group estimation in presence of weakly cross- correlated estimators. Economics Letters 175, 101 { 105. Chung, H., J.-P. Laforte, D. Reifschneider, and J. C. Williams (2012). Have we underesti- mated the likelihood and severity of zero lower bound events? Journal of Money, Credit and Banking 44, 47{82. Converse, N., E. L. Yeyati, and T. Williams (2020, January). How ETFs Amplify the Global Financial Cycle in Emerging Markets. International Finance Discussion Papers 1268, Board of Governors of the Federal Reserve System (U.S.). Datta, D., J. M. Londono, B. Sun, D. O. Beltran, T. RT Ferreira, M. M. Iacoviello, M. R. Jahan-Parvar, C. Li, M. Rodriguez, and J. H. Rogers (2017). Taxonomy of global risk, uncertainty, and volatility measures. Matteo M. and Jahan-Parvar, Mohammad R. and Li, Canlin and Rodriguez, Marius and Rogers, John H., Taxonomy of Global Risk, Un- certainty, and Volatility Measures (2017-11-21). FRB International Finance Discussion Paper (1216). Datta, D. D., B. K. Johannsen, H. Kwon, and R. J. Vigfusson (2018, August). Oil, Equities, and the Zero Lower Bound. Finance and Economics Discussion Series 2018-058, Board of Governors of the Federal Reserve System (U.S.). 139 Davis, J. S., M. B. Devereux, and C. Yu (2020, November). Sudden Stops and Optimal For- eign Exchange Intervention. NBER Working Papers 28079, National Bureau of Economic Research, Inc. De Bock, R. and I. de Carvalho Filho (2015a). The behavior of currencies during risk-o episodes. Journal of International Money and Finance 53, 218{234. De Bock, R. and I. de Carvalho Filho (2015b). The behavior of currencies during risk-o episodes. Journal of International Money and Finance 53 (C), 218{234. Debortoli, D., J. Gal , and L. Gambetti (2020). On the empirical (ir) relevance of the zero lower bound constraint. NBER Macroeconomics Annual 34 (1), 141{170. Dees, S., S. Holly, M. H. Pesaran, and L. V. Smith (2007). Long run macroeconomic relations in the global economy. Economics: The Open-Access, Open-Assessment E-Journal 1 (5), 1{29. Delle Chiaie, S., L. Ferrara, and D. Giannone (2018a). Common factors of commodity prices. Research Bulletin 51. Delle Chiaie, S., L. Ferrara, and D. Giannone (2018b). Common factors of commodity prices. Research Bulletin 51. Devereux, M. B. and C. Engel (1998, December). Fixed vs. Floating Exchange Rates: How Price Setting Aects the Optimal Choice of Exchange-Rate Regime. NBER Working Papers 6867, National Bureau of Economic Research, Inc. Dominguez, K. M., Y. Hashimoto, and T. Ito (2012). International reserves and the global nancial crisis. Journal of International Economics 88 (2), 388{406. Driscoll, J. C. and A. C. Kraay (1998). Consistent covariance matrix estimation with spatially dependent panel data. Review of Economics and Statistics 80 (4), 549{560. Duncombe, G., M. Nigro, and B. Kay (2018). Forecasting factor returns. Two Sigma Investments. Dusak, K. (1973). Futures trading and investor returns: An investigation of commodity market risk premiums. Journal of Political economy 81 (6), 1387{1406. ECB (2019). The International Role of the Euro. Edwards, S. (1996, September). The Determinants of the Choice between Fixed and Flexible Exchange-Rate Regimes. NBER Working Papers 5756, National Bureau of Economic Research, Inc. Eggertsson, G. B. (2011). What scal policy is eective at zero interest rates? NBER Macroeconomics Annual 25 (1), 59{112. Eichengreen, B. (2018). The Open-Economy Trilemma in the Long Run. Korean Economic Review 34, 5{28. 140 Eichengreen, B. and R. Hausmann (1999). Exchange rates and nancial fragility. Technical report, National bureau of economic research. Engel, C. (1996). The forward discount anomaly and the risk premium: A survey of recent evidence. Journal of Empirical Finance 3 (2), 123{192. Engel, C. (2011, October). Currency misalignments and optimal monetary policy: A reex- amination. American Economic Review 101 (6), 2796{2822. Engle, R. (2002). Dynamic conditional correlation: A simple class of multivariate generalized autoregressive conditional heteroskedasticity models. Journal of Business & Economic Statistics 20 (3), 339{350. Erb, C. B. and C. R. Harvey (2006). The strategic and tactical value of commodity futures. Financial Analysts Journal 62 (2), 69{97. Fama, E. F. (1984, November). Forward and spot exchange rates. Journal of Monetary Economics 14 (3), 319{338. Fama, E. F. and K. R. French (1988). Business cycles and the behavior of metals prices. The Journal of Finance 43 (5), 1075{1093. Feldman, B. and H. Till (2006). Backwardation and commodity futures performance: Evi- dence from evolving agricultural markets. The Journal of Alternative Investments 9 (3), 24{39. Fernandez, A., S. Schmitt-Grohe, and M. Uribe (2017). World shocks, world prices, and business cycles: An empirical investigation. Journal of International Economics 108, S2{S14. Fern andez, A., M. W. Klein, A. Rebucci, M. Schindler, and M. Uribe (2016, August). Capital Control Measures: A New Dataset. IMF Economic Review 64 (3), 548{574. Fischer, S. (2001, Spring). Exchange Rate Regimes: Is the Bipolar View Correct? Journal of Economic Perspectives 15 (2), 3{24. Forbes, K. J. and R. Rigobon (2002). No contagion, only interdependence: Measuring stock market comovements. The Journal of Finance 57 (5), 2223{2261. Frankel, J. (2019, April). Systematic Managed Floating. Open Economies Review 30 (2), 255{295. Frankel, J., E. Fajnzylber, S. Schmukler, and L. Serven (2001). Verifying exchange rate regimes. Journal of Development Economics 66 (809), 351{386. Frankel, J., X. Ma, and D. Xie (2019). The impact of exchange rate regimes on economic growth with continuous classication of de facto regimes. 141 Frankel, J., S. Schmukler, and L. Serven (2000, September). Veriability and the vanishing intermediate exchange rate regime. Working Paper 7901, National Bureau of Economic Research. Frankel, J., S. L. Schmukler, and L. Serven (2004, September). Global transmission of interest rates: monetary independence and currency regime. Journal of International Money and Finance 23 (5), 701{733. Frankel, J. A. (1993, January). Is Japan Creating a Yen Bloc in East Asia and the Pacic?, pp. 53{88. The University of Chicago Press. Frankel, J. A. and S.-J. Wei (1992). Yen bloc or dollar bloc: exchange rate policies of the East Asian economies. Technical report. Frankel, J. A. and S.-J. Wei (1993, April). Emerging currency blocs. Working Paper 4335, National Bureau of Economic Research. Franses, P. H., D. Van Dijk, et al. (2000). Non-linear time series models in empirical nance. Cambridge university press. Fratzscher, M. (2012). Capital ows, push versus pull factors and the global nancial crisis. Journal of International Economics 88 (2), 341 { 356. NBER Global. Georgiadis, G. and F. Zhu (2019, July). Monetary policy spillovers, capital controls and exchange rate exibility, and the nancial channel of exchange rates. BIS Working Papers 797, Bank for International Settlements. Gertler, M. and P. Karadi (2015). Monetary policy surprises, credit costs, and economic activity. American Economic Journal: Macroeconomics 7 (1), 44{76. Ghosh, A. R., J. D. Ostry, and M. S. Qureshi (2016). When do capital in ow surges end in tears? American Economic Review 106 (5), 581{85. Girton, L. and D. Roper (1977). A monetary model of exchange market pressure applied to the postwar canadian experience. The American Economic Review 67 (4), 537{548. Glosten, L. R., R. Jagannathan, and D. E. Runkle (1993, December). On the Relation between the Expected Value and the Volatility of the Nominal Excess Return on Stocks. Journal of Finance 48 (5), 1779{1801. Goldberg, L. S. and S. Krogstrup (2018, February). International Capital Flow Pressures. NBER Working Papers 24286, National Bureau of Economic Research, Inc. Goldberg, L. S. and R. Lerman (2019). The U.S. Dollar's Global Roles: Where Do Things Stand? Gopinath, G. (2015). The international price system. Technical report, National Bureau of Economic Research. 142 Gorton, G. and K. G. Rouwenhorst (2006). Facts and fantasies about commodity futures. Financial Analysts Journal 62 (2), 47{68. Gorton, G. B., F. Hayashi, and K. G. Rouwenhorst (2013). The fundamentals of commodity futures returns. Review of Finance 17 (1), 35{105. Gust, C., E. Herbst, D. L opez-Salido, and M. E. Smith (2017). The empirical implications of the interest-rate lower bound. American Economic Review 107 (7), 1971{2006. Haldane, A. G. and S. G. Hall (1991, May). Sterling's Relationship with the Dollar and the Deutschemark: 1976-89. Economic Journal 101 (406), 436{443. Han, X. and S.-J. Wei (2014, December). Policy Choices and Resilience to International Monetary Shocks. Global Economic Review 43 (4), 319{337. Han, X. and S.-J. Wei (2018). International transmissions of monetary shocks: Between a trilemma and a dilemma. Journal of International Economics 110 (C), 205{219. Hastie, T. J. and R. J. Tibshirani (1990). Generalized additive models, Volume 43. CRC press. Haubrich, J., G. Pennacchi, and P. Ritchken (2012). In ation expectations, real rates, and risk premia: Evidence from in ation swaps. The Review of Financial Studies 25 (5), 1588{ 1629. Hernandez-Vega, M. (2019, July). Estimating Capital Flows To Emerging Market Economies With Heterogeneous Panels. Macroeconomic Dynamics 23 (5), 2068{2088. Herwartz, H. and J. Roestel (2017). Mundell's trilemma: Policy trade-os within the middle ground. Journal of International Money and Finance 75 (C), 1{13. Hirshleifer, D. (1988). Residual risk, trading costs, and commodity futures risk premia. The Review of Financial Studies 1 (2), 173{193. Hofmann, B., I. Shim, and H. S. Shin (2019, March). Bond risk premia and the exchange rate. BIS Working Papers 775, Bank for International Settlements. Hong, H. and M. Yogo (2012). What does futures market interest tell us about the macroe- conomy and asset prices? Journal of Financial Economics 105 (3), 473{490. Hossfeld, O. and M. Pramor (2018). Global liquidity and exchange market pressure in emerging market economies. Technical report. Ilzetzki, E., C. M. Reinhart, and K. S. Rogo (2019). Exchange Arrangements Entering the Twenty-First Century: Which Anchor will Hold? The Quarterly Journal of Eco- nomics 134 (2), 599{646. Ito, H. and M. Kawai (2012, September). New Measures of the Trilemma Hypothesis : Implications for Asia. Macroeconomics Working Papers 23331, East Asian Bureau of Economic Research. 143 Ito, H. and M. Kawai (2014, January). Determinants of the Trilemma Policy Combination. ADBI Working Papers 456, Asian Development Bank Institute. Ito, H. and M. Kawai (2016). Trade invoicing in major currencies in the 1970s{1990s: Lessons for renminbi internationalization. Journal of the Japanese and International Economies 42 (C), 123{145. Ito, H. and R. McCauley (2016). Re-scaling global imbalances: key currency zones and renminbi management. Technical report. Ito, H. and R. N. McCauley (2019). A key currency view of global imbalances. Journal of International Money and Finance 94 (C), 97{115. Jaroci nski, M. and P. Karadi (2020). Deconstructing monetary policy surprises|the role of information shocks. American Economic Journal: Macroeconomics 12 (2), 1{43. Jeanne, O. and R. Ranciere (2011). The optimal level of international reserves for emerging market countries: A new formula and some applications. The Economic Journal 121 (555), 905{930. Jeanne, O. and D. Sandri (2020). Global nancial cycle and liquidity management. Technical report, National Bureau of Economic Research. Jord a, O. (2005). Estimation and inference of impulse responses by local projections. Amer- ican economic review 95 (1), 161{182. Jord a, O., M. Schularick, and A. M. Taylor (2020). The eects of quasi-random monetary experiments. Journal of Monetary Economics 112, 22{40. Jord a, , M. Schularick, and A. M. Taylor (2015). Betting the house. Journal of International Economics 96 (S1), 2{18. Judson, R. A. and A. L. Owen (1999). Estimating dynamic panel data models: a guide for macroeconomists. Economics letters 65 (1), 9{15. Kalemli-Ozcan, S. (2019a, September). U.S. Monetary Policy and International Risk Spillovers. NBER Working Papers 26297, National Bureau of Economic Research, Inc. Kalemli-Ozcan, S. (2019b, September). U.s. monetary policy and international risk spillovers. Working Paper 26297, National Bureau of Economic Research. Kilian, L. (2009). Not all oil price shocks are alike: Disentangling demand and supply shocks in the crude oil market. American Economic Review 99 (3), 1053{69. Kilian, L. and X. Zhou (2018a). Modeling uctuations in the global demand for commodities. Journal of International Money and Finance 88 (C), 54{78. Kilian, L. and X. Zhou (2018b). Modeling uctuations in the global demand for commodities. Journal of International Money and Finance 88, 54{78. 144 Klein, M. W. and J. C. Shambaugh (2015, October). Rounding the Corners of the Policy Trilemma: Sources of Monetary Policy Autonomy. American Economic Journal: Macroe- conomics 7 (4), 33{66. Kucher, O. and A. Kurov (2014). Business cycle, storage, and energy prices. Review of Financial Economics 23 (4), 217{226. Kuttner, K. N. (2001). Monetary policy surprises and interest rates: Evidence from the fed funds futures market. Journal of monetary economics 47 (3), 523{544. Ma, G. and R. N. McCauley (2011, March). The evolving renminbi regime and implications for Asian currency stability. Journal of the Japanese and International Economies 25 (1), 23{38. Maggiori, M., B. Neiman, and J. Schreger (2019). The rise of the dollar and fall of the euro as international currencies. In AEA Papers and Proceedings, Volume 109, pp. 521{26. Masson, P. R. (2001). Exchange rate regime transitions. Journal of Development Eco- nomics 64 (2), 571 { 586. McCauley, R. N. and T. Chan (2014, December). Currency movements drive reserve com- position. BIS Quarterly Review. Mendoza, E. G. (2010). Sudden stops, nancial crises, and leverage. American Economic Review 100 (5), 1941{66. Miniane, J. and J. H. Rogers (2007, August). Capital Controls and the International Trans- mission of U.S. Money Shocks. Journal of Money, Credit and Banking 39 (5), 1003{1035. Miranda-Agrippino, S. and H. Rey (2020, 05). U.S. Monetary Policy and the Global Financial Cycle. The Review of Economic Studies. rdaa019. Mohaddes, K. and M. Raissi (2020). Compilation, Revision and Updating of the Global VAR (GVAR) Database, 1979Q2-2019Q4. Mundell, R. A. (1963). Capital mobility and stabilization policy under xed and exible exchange rates. The Canadian Journal of Economics and Political Science/Revue canadi- enne d'Economique et de Science politique 29 (4), 475{485. Nickell, S. J. (1981, November). Biases in Dynamic Models with Fixed Eects. Economet- rica 49 (6), 1417{1426. Obstfeld, M., J. D. Ostry, and M. S. Qureshi (2019a, May). A Tie That Binds: Revisiting the Trilemma in Emerging Market Economies. The Review of Economics and Statistics 101 (2), 279{293. Obstfeld, M., J. D. Ostry, and M. S. Qureshi (2019b). A tie that binds: Revisiting the trilemma in emerging market economies. Review of Economics and Statistics 101 (2), 279{293. 145 Obstfeld, M., J. C. Shambaugh, and A. M. Taylor (2005, August). The Trilemma in History: Tradeos Among Exchange Rates, Monetary Policies, and Capital Mobility. The Review of Economics and Statistics 87 (3), 423{438. Obstfeld, M., J. C. Shambaugh, and A. M. Taylor (2010, April). Financial stability, the trilemma, and international reserves. American Economic Journal: Macroeconomics 2 (2), 57{94. Pesaran, B. and M. H. Pesaran (2010). Conditional volatility and correlations of weekly returns and the var analysis of 2008 stock market crash. Economic Modelling 27 (6), 1398{1416. Pesaran, M. and R. Smith (1995). Estimating long-run relationships from dynamic hetero- geneous panels. Journal of Econometrics 68 (1), 79 { 113. Pesaran, M. H. (2006). Estimation and inference in large heterogeneous panels with a multifactor error structure. Econometrica 74 (4), 967{1012. Pesaran, M. H., T. Schuermann, and S. M. Weiner (2004). Modeling regional interdepen- dencies using a global error-correcting macroeconometric model. Journal of Business Economic Statistics 22 (2), 129{162. Pesaran, M. H., Y. Shin, and R. P. Smith (1999). Pooled mean group estimation of dynamic heterogeneous panels. Journal of the American statistical Association 94 (446), 621{634. Rey, H. (2015, May). Dilemma not Trilemma: The Global Financial Cycle and Monetary Policy Independence. NBER Working Papers 21162, National Bureau of Economic Re- search, Inc. Ruge-Murcia, F. (2014). Indirect inference estimation of nonlinear dynamic general equilib- rium models: with an application to asset pricing under skewness risk. Technical report, Centre interuniversitaire de recherche en economie quantitative, CIREQ. Shambaugh, J. C. (2004). The Eect of Fixed Exchange Rates on Monetary Policy. The Quarterly Journal of Economics 119 (1), 301{352. Shousha, S. (2016). Macroeconomic eects of commodity booms and busts: The role of nancial frictions. Steiner, A. (2017). Central banks and macroeconomic policy choices: Relaxing the trilemma. Journal of Banking Finance 77, 283 { 299. Swanson, E. T. and J. C. Williams (2014a). Measuring the eect of the zero lower bound on medium-and longer-term interest rates. American Economic Review 104 (10), 3154{85. Swanson, E. T. and J. C. Williams (2014b). Measuring the eect of the zero lower bound on yields and exchange rates in the uk and germany. Journal of International Economics 92, S2{S21. 146 Szymanowska, M., F. De Roon, T. Nijman, and R. Van Den Goorbergh (2014). An anatomy of commodity futures risk premia. The Journal of Finance 69 (1), 453{482. Taylor, M. P., D. A. Peel, and L. Sarno (2001, November). Nonlinear Mean-Reversion in Real Exchange Rates: Toward a Solution to the Purchasing Power Parity Puzzles. International Economic Review 42 (4), 1015{1042. Uribe, M. and V. Z. Yue (2006, June). Country spreads and emerging countries: Who drives whom? Journal of International Economics 69 (1), 6{36. Wang, G.-J., C. Xie, M. Lin, and H. E. Stanley (2017). Stock market contagion during the global nancial crisis: A multiscale approach. Finance Research Letters 22 (C), 163{168. Westerlund, J. and J.-P. Urbain (2015). Cross-sectional averages versus principal compo- nents. Journal of Econometrics 185 (2), 372{377. Williamson, J. (2002). The evolution of thought on intermediate exchange rate regimes. The Annals of the American Academy of Political and Social Science 579, 73{86. Wood, S. N. (2017). Generalized additive models: an introduction with R. CRC press. Wu, J. C. and F. D. Xia (2016). Measuring the macroeconomic impact of monetary policy at the zero lower bound. Journal of Money, Credit and Banking 48 (2-3), 253{291. 147 Appendix A Chapter 1 Appendix to \Global Flights-to-Safety and Macroeconomic Adjustment in Emerging Markets" This appendix is organized in four sections. Section A.1 provides detail on relevant data and sources. Section A.2 provides additional results and further robustness checks. Section A.3 describes a method to separate excess risk sentiment from the global demand component embedded in Flight-to-Safety shocks, providing additional results on the eects of ight-to-safety shocks. A.1 Data Data is collected from a variety of sources. To construct global ight-to-safety shocks, the underlying daily data on the VIX Index, Wilshire 5000 index, 10-year Treasury yields, U.S. corporate high yield spreads, and exchange rates are taken from the FRED database. The daily data is collected spans January 2000 to August 2019. The daily data is eventually aggregated to a monthly frequency for further analysis. Monthly average sovereign spreads are measured with J.P. Morgan EMBI indices. The sample contains monthly data on spreads for 34 countries over the period January 2000 to August 2019. All countries have at least 99 observations, Log changes in EMBI spreads are computed as: s it = ln( S it S i;t1 ); 148 whereS it is the average EMBI spread level for countryi over montht. Because the anal- ysis relies on changes in the log EMBI spread, the bulk of summary statistics are reported on s it . Table A.1 reports summary statistics on changes in sovereign spreads across countries. outlier observations of logged EMBI changes greater than +200% or less than -100% are removed. Monthly industrial production data across countries is taken from the World Bank. Year- on-Year changes in log industrial production are computed as: y it = ln( Y it Y i;t12 ); whereY it is the nominal industrial production of countryi in montht. Summary statistics for year-over-year changes in log industrial production are reported in Table A.2. Iraq experienced very large swings in industrial production during the early 2000's when it was invaded and under military occupation. This is visible in her summary statistics. Table A.3 report summary statistics on select commodity and nancial market vari- ables at the monthly frequency. The values are monthly average changes, not end-of-month changes. For emerging markets, country-specic measures of nominal USD exchange rates are from the IMF. These are monthly averages, with changes in log exchange rates interpreted as log returns. Positive changes in denote domestic appreciation vis-a-vis the USD. Country-specic measures of international reserves are taken from the IMF as well. These are denominated in USD. Reserves growth rates are computed as changes in log monthly reserves, where positive monthly growth denotes reserves accumulation. Data on daily equity index prices across 32 countries are taken from Bloomberg to con- struct the global average realized volatility measure,GVOL t , similar to that of Cesa-Bianchi et al. [2020a]. Data on monthly estimates of the global nancial cycle from Miranda- Agrippino and Rey [2020], GFCY t , are available through the authors' website. 149 Table A.1: Summary Statistics for Changes in Log EMBI Spread Country T Min Max Mean Median SD Median Level Argentina 235 -0.730 0.686 0.009 -0.004 0.139 722.793 Belarus 107 -0.248 0.511 -0.007 -0.014 0.116 625.614 Brazil 235 -0.204 0.525 -0.005 -0.019 0.103 270.003 Chile 235 -0.368 0.487 -0.000 -0.002 0.096 139.650 China 235 -0.808 0.659 0.002 0.000 0.127 138.411 Colombia 235 -0.255 0.670 -0.004 -0.020 0.112 216.005 Cote d'Ivoire 235 -0.453 0.305 -0.003 -0.005 0.075 1106.238 Croatia 235 -0.270 0.371 -0.045 -0.070 0.103 257.671 Ecuador 235 -0.769 0.806 -0.007 -0.016 0.139 788.271 Egypt 217 -0.561 0.986 0.010 -0.010 0.187 349.198 El Salvador 208 -0.216 0.550 0.003 -0.006 0.093 376.053 Gabon 140 -0.267 0.646 0.003 -0.006 0.126 425.400 Hungary 235 -0.709 0.823 0.001 -0.002 0.167 123.800 Indonesia 182 -0.300 0.733 -0.004 -0.017 0.113 239.111 Iraq 160 -0.231 0.346 0.000 -0.003 0.095 520.688 Jordan 103 -0.348 0.374 -0.000 0.010 0.081 382.145 Kazakhstan 146 -0.279 0.669 0.001 -0.010 0.133 298.227 Lithuania 117 -0.459 0.395 -0.020 -0.023 0.151 123.726 Malaysia 235 -0.284 0.589 -0.001 -0.007 0.104 141.806 Mexico 235 -0.221 0.584 -0.001 -0.010 0.092 219.976 Pakistan 218 -0.525 0.523 -0.041 -0.024 0.179 512.429 Peru 235 -0.248 0.663 -0.005 -0.019 0.115 194.396 Philippines 235 -0.226 0.561 -0.006 -0.007 0.101 217.405 Poland 235 -0.671 0.582 -0.007 0.008 0.138 109.399 Russia 235 -0.266 0.629 -0.010 -0.025 0.117 241.053 Senegal 99 -0.166 0.213 -0.001 -0.003 0.077 450.697 South Africa 235 -0.261 0.650 0.001 -0.004 0.110 236.514 Sri Lanka 141 -0.285 0.658 -0.001 -0.009 0.115 412.982 Tunisia 207 -0.525 0.481 -0.018 -0.049 0.123 209.755 Turkey 235 -0.241 0.532 0.001 -0.008 0.108 305.410 Ukraine 231 -0.475 0.974 -0.006 -0.012 0.148 620.636 Uruguay 218 -0.340 0.576 -0.002 -0.019 0.114 230.800 Venezuela 235 -0.209 0.605 0.011 0.001 0.109 1038.486 Vietnam 164 -0.283 0.665 -0.002 -0.005 0.137 249.750 Summary statistics for s it (Equation 1.9), monthly changes in the log EMBI spread. Column 8, Median Level, reports the median level of each country's EMBI spread. SD refers to standard deviation. 150 Table A.2: Summary Statistics for Year-over-Year Change in Log Industrial Production Country T Min Max Mean Medan SD Argentina 235 -0.222 0.245 0.022 0.023 0.078 Belarus 151 -0.109 1.997 0.259 0.136 0.436 Brazil 235 -0.170 0.190 0.012 0.013 0.064 Chile 235 -0.131 0.140 0.021 0.027 0.044 China 235 0.038 0.207 0.116 0.114 0.044 Colombia 235 -0.143 0.163 0.025 0.020 0.054 Cote d'Ivoire 195 -0.501 0.581 0.027 0.035 0.163 Croatia 235 -0.142 0.131 0.014 0.017 0.052 Ecuador 235 -0.170 0.491 0.043 0.050 0.078 Egypt 175 -0.145 0.410 0.044 0.034 0.081 El Salvador 235 -0.046 0.079 0.014 0.014 0.023 Gabon 235 -0.377 0.426 -0.005 0.018 0.137 Hungary 235 -0.302 0.291 0.046 0.056 0.087 Indonesia 235 -0.136 0.345 0.040 0.038 0.053 Iraq 235 -0.830 11.500 0.144 0.087 0.860 Jordan 235 -0.229 0.286 0.022 0.015 0.078 Kazakhstan 235 -0.096 0.414 0.072 0.059 0.083 Lithuania 235 -0.260 0.381 0.048 0.050 0.088 Malaysia 235 -0.176 0.234 0.042 0.040 0.063 Mexico 235 -0.177 0.148 0.016 0.022 0.048 Pakistan 235 -0.195 0.319 0.049 0.039 0.084 Peru 235 -0.141 0.222 0.037 0.037 0.073 Philippines 235 -0.287 0.360 0.025 0.025 0.110 Poland 235 -0.153 0.234 0.054 0.055 0.059 Russian 235 -0.170 0.263 0.037 0.040 0.054 Senegal 151 -0.224 0.609 0.060 0.042 0.127 South Africa 235 -0.232 0.100 0.012 0.018 0.051 Sri Lanka 104 -0.143 0.193 0.025 0.020 0.059 Tunisia 235 -0.177 0.165 0.007 0.000 0.050 Turkey 235 -0.240 0.294 0.055 0.065 0.092 Ukraine 200 -0.308 0.221 0.011 0.023 0.107 Uruguay 200 -0.311 0.572 0.048 0.037 0.127 Venezuela 235 -0.648 1.832 -0.045 -0.015 0.229 Vietnam 128 -0.504 0.679 0.104 0.103 0.214 Summary statistics for y it (Equation 1.9). Iraq's large minimum and maximum driven by the war period in the early 2000s. SD refers to standard deviation. 151 Table A.3: Summary Statistics for Select Financial and Commodity Market Variables Market Variable (Monthly) T Mean SD Min Pctl(25) Pctl(75) Max VIX 235 0.001 0.167 0.373 0.098 0.068 0.708 U.S. High Yield Credit Spread 235 0.0005 0.088 0.223 0.059 0.043 0.486 Wilshire 5000 Index 235 0.0002 0.002 0.008 0.001 0.002 0.005 3-month Treasury Yield 235 0.004 0.329 1.738 0.072 0.065 2.025 2-year Treasury Yield 235 0.006 0.124 0.568 0.070 0.061 0.316 5-year Treasury Yield 235 0.006 0.100 0.411 0.058 0.046 0.360 10-year Treasury Yield 235 0.006 0.070 0.378 0.046 0.034 0.194 1-year In ation Expectations 235 0.0001 0.004 0.013 0.002 0.002 0.017 2-year In ation Expectations 235 0.0001 0.002 0.006 0.001 0.001 0.008 10-year In ation Expectations 235 0.007 0.101 0.368 0.067 0.059 0.253 USD/G10 Exchange Rate 235 0.0002 0.019 0.050 0.013 0.013 0.082 Copper Price 235 0.005 0.065 0.354 0.025 0.038 0.230 WTI Crude Oil Price 235 0.003 0.087 0.332 0.045 0.060 0.214 Gold Price 235 0.007 0.037 0.124 0.016 0.032 0.115 In ation expectations are monthly changes (not logged). All others are monthly changes in logs. In ation expectations are estimated using the method of Haubrich et al. [2012]. 152 Table A.4: U.S. Traded ETFs Granting Exposure to an EM Country Country Number of ETFs 1 Argentina 112 2 Belarus 0 3 Brazil 281 4 Chile 118 5 China 571 6 Colombia 101 7 Cote d'Ivoire 0 8 Croatia 1 9 Ecuador 0 10 Egypt 60 11 El Salvador 0 12 Gabon 0 13 Hungary 58 14 Indonesia 127 15 Iraq 3 16 Jordan 3 17 Kazakhstan 6 18 Lithuania 0 19 Malaysia 193 20 Mexico 340 21 Pakistan 47 22 Peru 6 23 Philippines 109 24 Poland 106 25 Russia 162 26 Senegal 0 27 South Africa 231 28 Sri Lanka 2 29 Tunisia 0 30 Turkey 107 31 Ukraine 7 32 Uruguay 0 33 Venezuela 0 34 Vietnam 1 Source: etfdb.com. Data collected as of October 2020. 153 Figure A.1: Response to a 1-Standard Deviation FTS Shock −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 3M US Yield −1.00 −0.75 −0.50 −0.25 0.00 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 2Y US Yield −0.9 −0.6 −0.3 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 5Y US Yield −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 1Y U.S. Infl. Exp. −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 2Y U.S. Infl. Exp. −0.8 −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 10Y U.S. Infl. Exp. Cumulative response (in standard deviations) to a 1-standard deviation structural ight-to-safety (FTS) shock, FTS t . 90% bootstrapped condence bands. A.2 Additional Results and Robustness A.2.1 Robustness 154 Figure A.2: Response to a 1-Standard Deviation FTS Shock −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 3M US Yield −1.00 −0.75 −0.50 −0.25 0.00 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 2Y US Yield −0.9 −0.6 −0.3 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 5Y US Yield −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 1Y U.S. Infl. Exp. −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 2Y U.S. Infl. Exp. −0.8 −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 10Y U.S. Infl. Exp. Cumulative response (in standard deviations) to a 1-standard deviation structural ight-to-safety (FTS) shock, FTS t . 90% bootstrapped condence bands. 155 Figure A.3: Response to a 1-Standard Deviation FTS Shock −1.00 −0.75 −0.50 −0.25 0.00 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Industrial Materials −0.75 −0.50 −0.25 0.00 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Rubber −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Coal −1.00 −0.75 −0.50 −0.25 0.00 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Aluminum −0.3 −0.2 −0.1 0.0 0.1 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Iron −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Soybeans −0.3 −0.2 −0.1 0.0 0.1 0.2 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Coffee −0.4 −0.2 0.0 0.2 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Sugar Cumulative response (in standard deviations) to a 1-standard deviation structural ight-to-safety (FTS) shock, FTS t . 90% bootstrapped condence bands. 156 Figure A.4: Response to a 1-Standard Deviation FTS Shock Orthogonal to Changes in log VIX −0.3 −0.2 −0.1 0.0 0.1 0.2 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 3M US Yield −0.8 −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 2Y US Yield −0.75 −0.50 −0.25 0.00 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 5Y US Yield −0.4 −0.3 −0.2 −0.1 0.0 0.1 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 1Y U.S. Infl. Exp. −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 2Y U.S. Infl. Exp. −0.50 −0.25 0.00 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 10Y U.S. Infl. Exp. −0.2 0.0 0.2 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Gold −0.3 −0.2 −0.1 0.0 0.1 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Silver −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations USD/G10 −0.8 −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Commodities −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Crude Oil −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Copper −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Industrial Materials −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Rubber −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Coal −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Aluminum −0.1 0.0 0.1 0.2 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Iron −0.5 −0.4 −0.3 −0.2 −0.1 0.0 0.1 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Soybeans −0.3 −0.2 −0.1 0.0 0.1 0.2 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Coffee −0.2 0.0 0.2 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Sugar Cumulative response (in standard deviations) to a 1-standard deviation structural ight-to-safety (FTS) shock, FTS t orthogonal to log VIX changes. In a rst-stage, FTS t is regressed on changes in the log VIX index. 90% bootstrapped condence bands. 157 Figure A.5: Setting FTS Condition Threshold c = 1, Response to a 1-Standard Deviation FTS Shock −0.4 −0.3 −0.2 −0.1 0.0 0.1 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 3M US Yield −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 2Y US Yield −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 5Y US Yield −0.4 −0.3 −0.2 −0.1 0.0 0.1 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 1Y U.S. Infl. Exp. −0.5 −0.4 −0.3 −0.2 −0.1 0.0 0.1 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 2Y U.S. Infl. Exp. −0.4 −0.3 −0.2 −0.1 0.0 0.1 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 10Y U.S. Infl. Exp. −0.2 −0.1 0.0 0.1 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Gold −0.5 −0.4 −0.3 −0.2 −0.1 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Silver −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations USD/G10 −0.75 −0.50 −0.25 0.00 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Commodities −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Crude Oil −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Copper −0.8 −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Industrial Materials −0.8 −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Rubber −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Coal −0.8 −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Aluminum −0.4 −0.3 −0.2 −0.1 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Iron −0.5 −0.4 −0.3 −0.2 −0.1 0.0 0.1 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Soybeans −0.1 0.0 0.1 0.2 0.3 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Coffee −0.5 −0.4 −0.3 −0.2 −0.1 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Sugar Cumulative response (in standard deviations) to a 1-standard deviation structural ight-to-safety (FTS) shock, FTS t . 90% bootstrapped condence bands. 158 Figure A.6: Response to a 1-Standard Deviation FTS Shock ordered Last in the Structural VAR −0.2 −0.1 0.0 0.1 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 3M US Yield −0.5 −0.4 −0.3 −0.2 −0.1 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 2Y US Yield −0.5 −0.4 −0.3 −0.2 −0.1 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 5Y US Yield −0.3 −0.2 −0.1 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 1Y U.S. Infl. Exp. −0.4 −0.3 −0.2 −0.1 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 2Y U.S. Infl. Exp. −0.5 −0.4 −0.3 −0.2 −0.1 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 10Y U.S. Infl. Exp. −0.2 −0.1 0.0 0.1 0.2 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Gold −0.3 −0.2 −0.1 0.0 0.1 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Silver −0.3 −0.2 −0.1 0.0 0.1 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations USD/G10 −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Commodities −0.5 −0.4 −0.3 −0.2 −0.1 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Crude Oil −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Copper −0.4 −0.3 −0.2 −0.1 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Industrial Materials −0.4 −0.3 −0.2 −0.1 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Rubber −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Coal −0.4 −0.3 −0.2 −0.1 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Aluminum −0.1 0.0 0.1 0.2 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Iron −0.4 −0.3 −0.2 −0.1 0.0 0.1 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Soybeans −0.1 0.0 0.1 0.2 0.3 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Coffee −0.1 0.0 0.1 0.2 0.3 0.4 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Sugar Cumulative response (in standard deviations) to a 1-standard deviation structural ight-to-safety (FTS) shock, FTS t . 90% bootstrapped condence bands. 159 Figure A.7: Response to a 1-Standard Deviation FTS Shock Orthogonal to Changes in log VIX 0.0 0.1 0.2 0.3 0 10 20 30 Months Standard Deviations Sovereign Spread −0.6 −0.4 −0.2 0.0 0 10 20 30 Months Standard Deviations Industrial Production −1.2 −0.9 −0.6 −0.3 0.0 0 10 20 30 Months % Change Exchange Rate −2.0 −1.5 −1.0 −0.5 0.0 0 10 20 30 Months % Change International Reserves Cumulative MG Response (Equation 3.10) to a 1-standard deviation structural ight-to-safety shock, FTSt. 95% non- parametric dispersion bands as computed in Equation 3.12. Log sovereign spread in monthly changes. Industrial production as year-over-year log change. Negative values imply exchange rate percent depreciation. International reserves in monthly log changes. 160 Figure A.8: Setting FTS Condition threshold c = 1, Response to a 1-Standard Deviation FTS Shock 0.0 0.1 0.2 0.3 0.4 0 10 20 30 Months Standard Deviations Sovereign Spread −0.6 −0.4 −0.2 0.0 0 10 20 30 Months Standard Deviations Industrial Production −1.0 −0.5 0.0 0.5 0 10 20 30 Months % Change Exchange Rate −1.5 −1.0 −0.5 0.0 0 10 20 30 Months % Change International Reserves Cumulative MG Response (Equation 3.10) to a 1-standard deviation structural ight-to-safety shock, FTSt. 95% non- parametric dispersion bands as computed in Equation 3.12. Log sovereign spread in monthly changes. Industrial production as year-over-year log change. Negative values imply exchange rate percent depreciation. International reserves in monthly log changes. 161 Table A.5: Largest Daily Global FTS Shocks, 2000-2019 Description Date FTS d 1. British referendum votes to exit E.U. 2016-06-24 4.89 2. Chinese Correction: Authorities announced plans to curb speculation 2007-02-27 4.74 3. U.S. President Trump controversy 2017-05-17 3.44 4. Lehman Brothers Bankruptcy 2008-09-15 3.33 5. Arab Spring - Instability in the Middle East and North Africa 2011-02-22 3.15 6. Italian political tensions, speculation of E.U. exit 2018-05-29 3.12 7. ECB announces no new emergency support for Greece; Greece calls for bailout referendum 2015-06-29 3.05 8. S&P downgraded Greece's credit rating to 'junk' 2010-04-27 3.04 9.GFC: Congress rejects bank bailout bill 2008-09-29 2.71 10. U.S. - China trade war intensies 2019-08-05 2.61 February 24, 2020 would rank #4 and January 27, 2020 would rank #10 if the index was re- estimated through Feb 28, 2020 to account for the onset of the COVID-19 global pandemic. 162 Table A.6: Largest Daily Percent Wilshire 5000 Declines, 2000-2019 Description Date Change 1. GFC: NBER conrms U.S. recession 2008-12-01 -9.6% 2. 2008 GFC 2008-10-15 -9.4% 3. GFC: Congress rejects bank bailout bill 2008-09-29 -8.75% 4. 2008 GFC 2008-10-09 -7.8% 5. U.S. credit downgrade from AAA to AA+ by S&P 2011-08-08 -7.2% 6. 2008 GFC 2008-11-20 -7.1% 7. Tech Bubble Crash 2000-04-14 -6.6% 8. 2008 GFC 2008-11-19 -6.6% 9. 2008 GFC 2008-10-22 -6.1% 10. GFC: Fed communicates negative outlook 2008-10-07 -5.9% Table A.7: Largest Daily Log VIX (Percent) Changes, 2000-2019 Description Date Change 1. \VIXplosion" 2018-02-05 +76.8% 2. Chinese Correction: Authorities announced plans to curb speculation 2007-02-27 +49.6% 3. U.S. credit downgrade from AAA to AA+ by S&P 2011-08-08 +40.5% 4. British referendum votes to exit E.U. 2016-06-24 +40.1% 5. China slowdown 2015-08-21 +38.1% 6. U.S. President Trump controversy 2017-05-17 +38.1% 7.China introduces new exchange rate mechanism ahead of po- tential Fed hike 2015-08-24 +37.3% 8. N. Korea announces plans to attack the U.S. Naval Base Guam 2017-08-10 +36.7% 9. U.S. China Trade war concerns 2018-10-10 +36.4% 10. Boston Marathon terrorist attack 2013-04-15 +35.9% 163 Table A.8: Domestic Financial Factors and the Impact of FTS shocks on Economic Activity Dependent variable: 18-Month Response of IP Growth (1) (2) (3) (4) (5) (6) Intercept 0.413 0.702 0.626 0.752 0.292 0.307 (0.231) (0.274) (0.311) (0.267) (0.332) (0.355) 6M Spread Response 1.953 2.027 2.198 1.859 1.389 1.537 (0.444) (0.462) (0.591) (0.506) (0.575) (0.671) 6M FX Response 23.964 21.804 17.548 10.725 11.415 (9.254) (10.529) (9.025) (8.135) (8.884) 6M Reserves Response 12.022 7.241 25.492 25.577 (9.114) (7.938) (11.234) (11.796) ln(ETF i + 1) 0.109 0.082 0.080 (0.039) (0.040) (0.040) Commodity Exporter 0.143 (0.210) 6M FX 6M Reserves Response 1,263.024 1,295.828 (514.562) (513.679) Observations 34 34 34 34 34 34 R 2 0.270 0.370 0.432 0.538 0.593 0.601 Adjusted R 2 0.247 0.329 0.375 0.474 0.520 0.513 Robust standard errors. *, **, *** correspond to signicance at the 10, 5, and 1 percent level, respectively. Dependent variable is the cumulative 18-month expected response of IP growth (in SDs) to a 1-SD FTS shock (Dependent and independent variable descriptions found in Equation 1.13). The last independent variable is the interaction of the 6-month cumulative response of country i's exchange rate to a 1-SD FTS shock and the 6-month cumulative response of country i's international reserves. IP growth and changes in log spreads are in units of standard deviations. Exchange rate and reserves are in log changes. Commodity Exporter refers to an indicator variable denoting whether the country is a commodity exporter dened as having greater than 35% of total exports as commodities and greater than 5% of total trade (Aslam et al. [2016]). 164 Table A.9: Domestic Financial Factors and the Impact of VIX shocks on Economic Activity Dependent variable: 18-Month Response of IP Growth (1) (2) (3) (4) (5) (6) Intercept 0.094 0.095 0.010 0.277 0.020 0.025 (0.322) (0.332) (0.312) (0.307) (0.296) (0.291) 6M Spread Response 0.789 0.764 1.107 0.956 0.629 0.719 (0.548) (0.685) (0.637) (0.582) (0.597) (0.648) 6M FX Response 1.552 0.373 4.041 10.140 10.737 (18.582) (19.203) (13.690) (8.099) (8.725) 6M Reserves Response 10.773 4.113 9.865 10.360 (7.616) (8.206) (7.853) (8.549) ln(ETF i + 1) 0.131 0.100 0.099 (0.049) (0.038) (0.038) Commodity Exporter 0.093 (0.180) 6M FX 6M Reserves Response 1,614.727 1,655.827 (646.421) (645.686) Observations 34 34 34 34 34 34 R 2 0.061 0.062 0.139 0.360 0.508 0.513 Adjusted R 2 0.032 0.002 0.053 0.272 0.420 0.405 Robust standard errors. *, **, *** correspond to signicance at the 10, 5, and 1 percent level, respectively. Dependent variable is the cumulative 18-month expected response of IP growth (in SDs) to a 1-SD log VIX shock replacing the FTS shock in Equation 1.9 (Dependent and independent variable descriptions found in Equation 1.13). The last independent variable is the interaction of the 6-month cumulative response of country i's exchange rate to a 1-SD FTS shock and the 6-month cumulative response of country i's international reserves. IP growth and changes in log spreads are in units of standard deviations. Exchange rate and reserves are in log changes. Commodity Exporter refers to an indicator variable denoting whether the country is a commodity exporter dened as having greater than 35% of total exports as commodities and greater than 5% of total trade (Aslam et al. [2016]). 165 Table A.10: Domestic Financial Factors and the Impact of Global Financial Cycle shocks on Economic Activity Dependent variable: 18-Month Response of IP Growth (1) (2) (3) (4) (5) (6) Intercept 0.315 0.190 0.198 0.120 0.196 0.198 (0.289) (0.281) (0.277) (0.287) (0.291) (0.298) 6M Spread Response 0.734 0.628 0.920 0.842 0.798 0.741 (0.457) (0.518) (0.482) (0.441) (0.440) (0.544) 6M FX Response 12.816 11.913 8.052 6.002 5.674 (12.751) (13.507) (11.011) (10.860) (11.204) 6M Reserves Response 9.935 4.141 19.640 19.138 (8.744) (9.175) (11.730) (11.789) ln(ETF i + 1) 0.121 0.079 0.080 (0.055) (0.040) (0.043) Commodity Exporter 0.054 (0.232) 6M FX 6M Reserves Response 1,155.667 1,137.756 (512.268) (520.142) Observations 34 34 34 34 34 34 R 2 0.067 0.125 0.181 0.328 0.436 0.438 Adjusted R 2 0.038 0.069 0.099 0.235 0.336 0.313 Robust standard errors. *, **, *** correspond to signicance at the 10, 5, and 1 percent level, respectively. Dependent variable is the cumulative 18-month expected response of IP growth (in SDs) to a 1-SD log Global Financial Cycle shock of Miranda-Agrippino and Rey [2020] replacing the FTS shock in Equation 1.9 (Dependent and independent variable descriptions found in Equation 1.13). The last independent variable is the interaction of the 6-month cumulative response of countryi's exchange rate to a 1-SD FTS shock and the 6-month cumulative response of countryi's international reserves. IP growth and changes in log spreads are in units of standard deviations. Exchange rate and reserves are in log changes. Commodity Exporter refers to an indicator variable denoting whether the country is a commodity exporter dened as having greater than 35% of total exports as commodities and greater than 5% of total trade (Aslam et al. [2016]). A.2.2 Emerging markets: local projection impulse responses To check whether the results for the panel of emerging markets are robust to model specica- tion, I estimate impulse response functions using Local Projections (Jord a [2005]) rather than the structural VAR approach. I estimate the following local projection regressions for the nancial variables, sovereign spreads, exchange rate returns, and log changes in international 166 reserves: [s i;t+h ; FX i;t+h ; res i;t+h ] = i (h) + 1 X p=0 i1p (h)Y US;tp + 1 X p=0 i2p (h)Y i 0 ;tp + 1 X p=0 i3p (h)y i;tp + 1 X p=0 i4p (h)FTS tp + 2 X p=1 i5p (h)S i 0 ;tp + 2 X p=1 i6p (h)s i;tp + 2 X p=1 i7p (h)FX i;tp + 2 X p=1 i8p (h)res i;tp +e i;t+h : (A.1) To maintain the same recursive identication structure as in the multi-country SVAR, notice that for nancial variables (spreads, exchange rates and reserves), FTS t has a contempora- neous eect on nancial variables as do the measures of economic activity, given in Equation A.1. Meanwhile for the local projections corresponding to industrial production, y i;t+h is given by: y i;t+h = i (h) + 1 X p=0 i1p (h)Y US;tp + 1 X p=0 i2p (h)Y i 0 ;tp + 2 X p=1 i3p (h)y i;tp + 1 X p=0 i4p (h)FTS tp + 2 X p=1 i5p (h)S i 0 ;tp + 2 X p=1 i6p (h)s i;tp + 2 X p=1 i7p (h)FX i;tp + 2 X p=1 i8p (h)res i;tp +e i;t+h : (A.2) In Equation A.2, U.S. economic activity, Y US;t and the EM growth factor Y i 0 ;t enter contemporaneously on the right-hand-side. FTS shocks also enter contemporaneously while the other nancial variables enter with a lag. This way, the impact of the FTS shock will include both it's direct eect on y i;t+h but also the eect through nancial variables. The local projections from a global FTS shock on sovereign spreads, industrial produc- tion, exchange rates and international reserves are then plotted as the Mean Group estimate of i40 estimates, respectively in Figure A.9. Notice that the responses are consistent in direction and closely approximate in size compared to the baseline MG impulse responses estimated form the mutli-country SVAR (Equation 1.9). 167 Figure A.9: Emerging Markets: Average Response to a 1-Standard Deviation FTS Shock Using Local Projections 0.0 0.2 0.4 0.6 0 10 20 30 Months Standard Deviations Sovereign Spread −1.5 −1.0 −0.5 0.0 0 10 20 30 Months Standard Deviations Industrial Production −1.5 −1.0 −0.5 0.0 0 10 20 30 Months % Change Exchange Rate −4 −2 0 2 0 10 20 30 Months % Change International Reserves Cumulative MG Response from local projection estimates (Equations A.1 and A.2) to a 1-standard deviation structural ight-to-safety shock,FTS t . 95% non-parametric dispersion bands as computed in Equation 3.12. Log sovereign spread in monthly changes. Industrial production as year-over-year log change. Negative values imply exchange rate percent depreciation. International reserves in monthly log changes. A.2.3 Asymmetries: Positive versus negative FTS shocks The prevailing VAR model assumes FTS shocks are symmetric { negative (risk-on) and pos- itive (risk-o) movements are treated equally. However, there is good reason to believe that ight-to-safety as risk-o shocks have an asymmetric impact, likely much more severe than the inverse risk-on (or negative change in the FTS index). Introducing non-linearities into a VAR system comes with its own challenges. and does not guarantee parameter stability. A simple yet robust approach to compare the linear IRFs to the IRF when we allow asymmetry is to estimate the MG IRF for the marginal added impact of an FTS shock after allowing for asymmetries, and add that to the standard MG IRF under the linear case. Specically, the asymmetry imposed is that positive FTS shocks (risk-o) can impact the system dierently from negative FTS shocks (risk-on). Implementing this econometric design takes two steps. First, I include in the linear VAR contemporaneous FTS shocks, FTS t . Then, I add an additional equation to the VAR to capture the asymmetry, FTS t 1 FTSt>0 . This series re ects FTS shocks which are strictly positive, setting negative (risk-on) shocks to zero. Because the VAR controls for contemporanousFTS t fully, shocking theFTS t 1 FTSt>0 generates the IRF for the marginal asymmetric eect of a positive FTS shock above and beyond the linear eect. Figure A.10 shows the MG IRF from the standard, linear IRF and compares it to the IRF from a FTS shock once asymmetry is allowed for. When the model allows dierential 168 Figure A.10: Emerging Markets: Response to a 1-Standard Deviation Positive Symmetric FTS Shock (Solid) and After Allowing for Asymmetry (Dashed) 0.0 0.2 0.4 0.6 0.8 0 10 20 30 Months Standard Deviations Sovereign Spread −1.2 −0.9 −0.6 −0.3 0.0 0 10 20 30 Months Standard Deviations Industrial Production −1.5 −1.0 −0.5 0.0 0 10 20 30 Months % Change Exchange Rate −3 −2 −1 0 0 10 20 30 Months % Change International Reserves Cumulative MG Response (Equation 3.10) to a 1-standard deviation structural ight-to-safety shock,FTSt which is symmetric (solid) and after allowing positive (risk-o) shocks to have an asymmetric impact compared to negative (risk-on) FTS shocks (dashed). 95% non-parametric dispersion bands as computed in Equation 3.12. Log sovereign spread in monthly changes. Industrial production as year-over-year log change. Negative values imply exchange rate percent depreciation. International reserves in monthly log changes. eects of positive and negative FTS shocks, the impact of a positive FTS shock strengthens across all four variables of interest. This highlights that treating positive and negative FTS shocks symmetrically tends to downward-bias the impact of positive FTS shocks (risk-o) and induces upward-bias in the impact of negative FTS shocks (risk-on). Indeed, risk-o or positive FTS shocks have a substantially larger absolute impact on emerging market dynamics than negative FTS shocks, or risk-on. A.2.4 A model-free measure of global ights-to-safety The global FTS index, FTS t relies on an estimate of conditional volatility. Therefore, it needs to use the full sample for estimation, which may induce look-ahead biases, and comes with parametric assumptions. This poses an issue if one's primary objective is to forecast. On the other hand, this is less of an issue if one's goal is to combine ex ante and ex post information for explanatory purposes. The latter is the main objective of this paper, and similar full-sample approaches are taken in estimating global nancial shocks from realized volatility in Cesa-Bianchi et al. [2020a] and in constructing the Global Financial Cycle in Miranda-Agrippino and Rey [2020]. 1 1 In Cesa-Bianchi et al. [2020a] to identify nancial shocks, the authors regress global realized volatility (GVOL t ) on global real GDP growth over the full sample period. In Miranda-Agrippino and Rey [2020], the factor model employed to recover the common factor in risky asset prices takes information from the full sample. 169 As an alternative, I present a model-free measure of monthly global FTS shocks using daily changes in the log VIX index. Denote this measure the global FTS-VIX shock series, or FTS t (v), which is dened as: FTS t (v) = X lnVIX d (t) X lnVIX d (t)jd62FTS = X lnVIX d (t)1 d ; (A.3) Where the month t total change in the log VIX index is the sum of two components, FTS t (v) and P lnVIX d (t)jd62FTS. The rst term is the sum of log VIX changes in month t which occurs amid ight-to-safety, or risk-on/risk-o days. The second term is the sum of log VIX changes in the same month which occured on all remaining days. The indicator 1 d , as previously, imposes the ight-to-safety condition, thereby identifying risk-o and risk-on days using the daily returns across the candidate assets, denoted r ad : 1 d 8 > > > < > > > : 1 iffr 1d ;r 5d g>c\fr 2d ;r 3d ;r 4d ;r 6d g<c `Risk-O' 1 iffr 1d ;r 5d g<c\fr 2d ;r 3d ;r 4d ;r 6d g>c `Risk-On' 0 otherwise: (A.4) This way I rst classify each daily change in the VIX as belonging to a risk-on/risk-o event, or not. Then within each of these groups, summing the daily changes to the monthly level I break the total change in the log VIX index over month t into the component that occurred amid risk-on/risk-o, and the remainder. The former is a model-free measure of global FTS, denoted FTS t (v) which uses changes in the VIX index amid ights-to-safety (or risk-on/risk-o). To validate the new measure as a proxy for the baseline measure, the estimated correlation between FTS t and FTS t (v) is 0.87; the model-free measure is highly correlated with the the baseline FTS index. Baseline MG IRFs are reported in Figure A.11 using the model-free FTS shock series, FTS t (v), and the results are largely unchanged compared to the baseline impulse response functions shown in Figure 1.8. A.3 Flight-to-Safety, Excess Risk Sentiment, and Global Demand Large global shocks measured with asset prices re ect both risk sentiment and physical risk (global demand) - the latter referring to changing beliefs over future fundamentals. It's evident that global FTS shocks, a product of asset price movements, exhibits clear links to 170 Figure A.11: Average Response to a 1-Standard Deviation FTS Shock Using the Model Free Measure, FTS t (v) 0.0 0.2 0.4 0 10 20 30 Months Standard Deviations Sovereign Spread −0.8 −0.6 −0.4 −0.2 0.0 0 10 20 30 Months Standard Deviations Industrial Production −1.5 −1.0 −0.5 0.0 0 10 20 30 Months % Change Exchange Rate −2.0 −1.5 −1.0 −0.5 0.0 0 10 20 30 Months % Change International Reserves Cumulative MG Response (Equation 3.10) to a 1-standard deviation structural ight-to-safety shock, FTS t (v) dened in Equation A.3 as the model-free version of the shock series. 95% non-parametric disper- sion bands as computed in Equation 3.12. Log sovereign spread in monthly changes. Industrial production as year-over-year log change. Negative values imply exchange rate percent depreciation. International reserves in monthly log changes. global demand shown by their impact on commodity prices and U.S. in ation expectations and also by the economic relevance of the events triggering them. While the impact of FTS shocks itself is the main focus of this paper, here I separate the eects induced by an excess risk sentiment component and a global demand component of FTS shocks to better understand the two forces. I propose a simple reduced-form separation of FTS shocks into their global demand and excess risk sentiment components. This is accomplished by estimating a principal compo- nents regression (PCR) of global FTS shocks on the common factor in world commodity prices, an established proxy for global demand. The obtained residual then re ects the component of global FTS that is left unexplained by the contemporaneous adjustment in commodity prices, which I refer to as excess risk sentiment. More explicitly, I dene excess risk sentiment as the component of risk aecting nancial asset prices as pure risk premia; it is excess in that it has no causal eect on fundamental global demand and simply serves to compensate risk aversion. Suppose FTS shocks were made up of two orthogonal components, FTS t =G t +V t ; (A.5) whereG t re ects global demand, andV t is the excess risk sentiment. It's `excess' because it is the risk sentiment re ected in asset prices above and beyond whatever eect risk has had 171 Figure A.12: Separating Global Flights-to-Safety Shocks into Excess Risk Sentiment ( b V t ) and Global Demand ( b G t ) −4 −2 0 2 Jan 2000 Jan 2005 Jan 2010 Jan 2015 Jan 2020 Excess Risk Sentiment −7.5 −5.0 −2.5 0.0 2.5 Jan 2000 Jan 2005 Jan 2010 Jan 2015 Jan 2020 Gobal Demand Both series are normalized to have unit standard deviation. on global demand (which is absorbed inG t ). However these two components are unobserved, and therefore must be estimated. Therefore to recover the excess risk sentiment component, I regressFTS t on an estimate for global demand b G t , which I measure as the common factor in commodity prices: FTS t = ^ C t | {z } b Gt + V t ; b V t = V t ; (A.6) where C t = K X k=1 k c k;t : (A.7) C t is aTk matrix of log returns from a broad set of k commodity prices, and coe- cients k as set such 0 C t re ects the rst principal component of the space of commodity returns (i.e. the vector which maximizes the variance across the space of commodities). Specically, I estimate b G t using Principal Components Analysis (PCA) over a broad set of 66 commodity prices. The rst principal component is our estimate of b G t . Each commod- ity price series is log-dierenced and standardized. Then, by then regressing FTS t on the common commodity factor, I dene the obtained residual V t as excess risk sentiment b V t . Figure A.12 shows the decomposition of FTS shocks into risk premia and global de- mand, respectively. Starkly, the 2008-2009 global nancial crisis is identied as a large, negative global demand shock. August, September, and October 2008 re ect deep negative global demand pressure, each exceeding -3 standard deviations. Meanwhile the risk senti- 172 ment component of FTS for these months moved +0.01, +1.83,, +0.29 standard deviations, respectively. September 2008 re ected a particularly disruptive month with a joint adverse global demand and risk sentiment move driving the ight-to-safety. A.3.1 World prices and the excess risk component of FTS shocks Figure A.14 traces IRFs from a 1-SD global demand component (G t ) shock (solid), and also a 1-SD shock to the excess risk sentiment component, V t (dashed) on world prices. The risk sentiment channel drives a signicant portion of the response of U.S. interest rates and in ation expectations to ights-to-safety, while commodities are more sensitive to the global demand component. The U.S. Dollar is also much more sensitive to the global demand component than the risk sentiment component of ights-to-safety. After deconstructing FTS into risk and funadmental components, the signicant appreciative response of gold to adverse risk shocks is apparent, as is the depreciation of gold when hit with an adverse demand shock. These movements highlight the dual nature of gold as both a safe haven asset (hedges risk aversion) and a commodity with industrial use (procyclical). By contrast, another safe haven asset, the U.S. Dollar, appreciates in response to heightened risk sentiment or lower global demand. So while gold may provide a hedge against rising uncertainty (but not weaker global demand), the U.S. Dollar provides a hedge against both greater uncertainty and weaker global demand. A.3.2 Emerging Markets and the excess risk component of FTS shocks Figure A.15 shows the impact of a 1-SD global demand component (solid) shock along with the isolated excess risk component (dashed). Qualitatively, both adverse global demand and risk shocks drive emerging market dynamics in the same direction: tighter nancial conditions followed by economic contractions. However, quantitatively, emerging market dy- namics are much more sensitive to the global demand component compared to a comparably sized risk sentiment shock. A.3.3 Endogeneity and assumptions for separating excess risk sen- timent component of global FTS shocks The reduced-form approach to recover a measure of global excess risk sentiment has the ad- vantage of being convenient, robust and practical. The separation issue, however is subject 173 to complications when taking into account the presence of endogeneity: changing risk per- ceptions themselves can aect global demand (Bloom [2009], Caballero and Simsek [2020c]) and vice versa. Like asset prices, global FTS shocks, therefore, likely contain both a global demand and risk sentiment component, and the two may be correlated with one another. For the principal-components regression approach to consistently estimate true excess risk sentiment, there are a number of underlying conditions that must be satised: 1. The 1st principal component (PC) of commodity price returns re ects global demand. 2. Weak exogeneity of excess risk sentiment. 3. Commodity prices do not pay risk premium on aggregate risk. I discuss these issues here to acknowledge the limitations associated with them and eval- uate how reasonable each assumption may be. The second issue, weak exogeneity of excess risk sentiment implies that global demand is not contemporaneously impacted by excess risk sentiment, but can be impacted with a lag. Point 3 follows from points 1 and 2. If the 1st PC of commodity returns is in fact a proxy for global demand and is additionally not in uenced by excess risk premia the way nancial asset prices are, we should observe that investors in particularly pro-cyclical commodities are not compensated for the aggregate risk they bear. Importantly, point 3 is empirically testable. The 1st PC of commodity price returns re ects global demand The common factor in commodity prices, to proxy global demand,G t , must rst re ect uc- tuations in global demand. Recent and building evidence suggests this condition is validated (Kilian [2009], Kilian and Zhou [2018b] Delle Chiaie et al. [2018b], Alquist et al. [2020]). Importantly, global demand shocks are also not the same as uctuations in global activity. Global demand shocks can exhibit more volatility and move signicantly faster in re ect- ing information than, say, real GDP. This means that controlling for global demand is not the same as regressingFTS t on slow-moving macroeconomic aggregates. Commodity prices exhibit the unique feature of being both tied to the fundamental economy and adjusting at a relatively fast pace (Bailey and Chan [1993], Hong and Yogo [2012]). In fact, some highly nancialized commodity markets, like crude oil, respond to information at the speed of liquid nancial markets. Less liquid commodity markets may exhibit stickier prices, but often these prices still adjust faster than macroeconomic aggregates. 174 Weak exogeneity excess risk sentiment For illustration, suppose FTS shocks can be decomposed into asset price movements re- ecting: global demand G t the component of risk sentiment that aects global demand G t (non-excess risk sentiment), and excess or idiosyncratic risk sentiment component V t , FTS t = G t +V t ; (A.8) G t = e G t G t ; (A.9) where cov(G t ;V t ) = 0; cov( e G t ; G t )< 0; cov( G t ;V t ) = 0: Here, total global demandG t can be decomposed into the "pure" demand eect given by e G t and non-excess rising risk premia G t . Similarly, total risk premia is the sum of G t and excess risk sentiment V t . A crucial condition to satisfy the assumption of weak exogoneity is that non-excess risk sentiment that impacts global demand G t is contemporaneously uncorrelated with excess risk sentiment V t . Why might this condition be satised? Under the rationale that FTS shocks tend source from unique, unusual events. These events are unpredictable. And while the overall " ight-to-safety" signature is similar across these events, the underlying components { global demand, non-excess and excess risk sentiment { driving the ight-to-safety can dier drastically. For example, it may be that the FTS Shock induced by the September 11 terrorist attack was mostly a risk sentiment shock, while FTS during the 2008 Global Financial Crisis were contained a larger global demand shock component. Following the same logic, excess risk sentiment may dier from non-excess risk sentiment from shock to shock in an uncorrelated way. For instance, excess risk sentiment may be more related to technical market conditions or intermediary leverage prior to the FTS shock, while non-excess risk sentiment may be more associated with the degree of macroeconomic uncertainty caused by an unexpected news shock, therefore having a stronger impact on growth. Why might this condition be violated? Excess and non-excess risk sentiment driving asset prices may be correlated over the business cycle. If excess risk sentiment is determined by intermediary leverage, and that leverage varies systematically with the business cycle, the assumption of excess risk premia and non-excess risk premia being uncorrelated would be violated. 175 Commodity prices do not pay risk premium on aggregate risk This condition which follows from the previous assumptions has the advantage of being empirically testable. That is, consistent separation of excess risk sentiment component of FTS shocks from global demand using commodity prices, requires that commodity prices only adjust to changing global demand and not to excess risk premia. This is unlike nancial asset prices, since asset prices adjust to global demand but are also sensitive to investor risk sentiment. Non-excess risk sentiment can impact commodity prices indirectly by causally impacting global demand, but excess changes in risk sentiment do not re ect themselves in commodity prices. To put another way, commodity investors are not compensated for taking on aggregate risk the way it nancial assets compensate holders for bearing the same risk. For this as- sumption to be violated, heightened risk aversion must directly cause changes to commodity prices above and beyond any eect transmitting through risk aversion's eect on global growth prospects. A violation of this assumption would imply that particularly pro-cyclical commodities exhibit excess returns. I argue that considerable evidence suggests that this assumption is reasonably satised. Even at face value, Table A.11 shows annualized returns on commodity ETF investments which invest in futures against the S&P 500 since 2000. Crude oil, copper, and broad commodity prices all exhibit a high degree of procyclical be- havior. Despite this, an investment any of these commodities would have yielded negative annual returns over the past decade. Evidence of no aggregate risk premia applies for broad commodity spot returns too. Figure A.13 shows that for a set of 66 spot commodity returns from 2000-2019, U.S. equity betas are essentially uncorrelated with average returns. If ag- gregate risk premia was priced in the cross-section of commodities, commodities with higher betas would exhibit signicantly higher average returns historically. Table A.11: Commodity Futures Annualized Excess Returns Date Range Commodity Average Return Daily S&P 500 Beta 2007-2020 WTI Crude Oil -19.2% 0.76 2011-2020 Copper -3.5% 0.42 2007-2020 Commodity Basket -3.9% 0.43 2007-2020 S&P 500 6.16% 1 Daily log returns, annualized. Data taken from ETFs: USO, CPER, DBC, respectively. More rigorous evidence that commodity investments do not compensate for taking on aggregate risk has been documented over several decades (Dusak [1973], Feldman and Till [2006], Erb and Harvey [2006]). Rather, commodity risk premia has been linked to pro- 176 Figure A.13: Cross Section of Monthly Commodity Spot Return Betas, 2000-2019 R = 0.14 , p = 0.27 -3 0 3 6 9 -5 0 5 10 Monthly Beta Vs. U.S. Equities Annualized Monthly Return (%) Returns are annualized. U.S. Equity index used is the Wilshire 5000. ducer hedging demand 2 , which is an idiosyncratic supply-side phenomena and other factors like momentum (Hirshleifer [1988], Gorton and Rouwenhorst [2006], Gorton et al. [2013], Szymanowska et al. [2014]). Some commodities like energy and metals are more sensitive to global economic conditions than others (e.g. agriculture). There is some evidence of positive excess returns among energy and metals, but not related to associated aggregate risk. Rather, these commodities have higher expected returns during business cycle peaks when inventory is low, supportive of the producer hedging theory (Fama and French [1988], Kucher and Kurov [2014], Duncombe et al. [2018]). This goes in the opposite direction of what standard asset pricing theory would imply. 2 This comes from The Theory of Storage: in the face of low inventories, commodity prices and volatility rise due to risk of 'stock-out'. As a result, consumers of the commodity store supply at elevated levels. To hedge their production, risk-averse producers must provide additional compensation to counterparties as incentive to enter into commodity futures contracts. 177 Figure A.14: Risk and Fundamental Components of FTS: Response to a 1-Standard De- viation Global Demand Component Shock (Solid) and Excess Risk Sentiment Component Shock (Dashed) −0.50 −0.25 0.00 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 3M US Yield −1.00 −0.75 −0.50 −0.25 0.00 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 2Y US Yield −1.00 −0.75 −0.50 −0.25 0.00 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 5Y US Yield −0.8 −0.6 −0.4 −0.2 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 1Y U.S. Infl. Exp. −0.75 −0.50 −0.25 0.00 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 2Y U.S. Infl. Exp. −0.75 −0.50 −0.25 0.00 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations 10Y U.S. Infl. Exp. −0.75 −0.50 −0.25 0.00 0.25 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Gold −1.00 −0.75 −0.50 −0.25 0.00 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Silver −1.0 −0.5 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations USD/G10 −2.0 −1.5 −1.0 −0.5 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Commodities −1.5 −1.0 −0.5 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Crude Oil −1.5 −1.0 −0.5 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Copper −1.5 −1.0 −0.5 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Industrial Materials −1.5 −1.0 −0.5 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Rubber −1.5 −1.0 −0.5 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Coal −1.5 −1.0 −0.5 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Aluminum −0.50 −0.25 0.00 0.25 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Iron −0.8 −0.4 0.0 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Soybeans −0.6 −0.3 0.0 0.3 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Coffee −0.75 −0.50 −0.25 0.00 0.25 0 1 2 3 4 5 6 7 8 9 101112 Months Standard Deviations Sugar Cumulative Response (in standard deviations) to a 1-standard deviation risk and fundamental componenet of FTS: Component attributed to the excess risk sentiment component of FTS, ^ V t (dashed) and global demand component, ^ G t (solid) decomposed via Equation A.6. 90% bootstrapped condence bands. 178 Figure A.15: Emerging Markets and the Risk and Fundamental Components of FTS: Re- sponse to a 1-Standard Deviation Global Demand Component Shock (Solid) and Excess Risk Sentiment Component Shock (Dashed) 0.0 0.2 0.4 0.6 0 10 20 30 Months Standard Deviations Sovereign Spread −1.00 −0.75 −0.50 −0.25 0.00 0 10 20 30 Months Standard Deviations Industrial Production −2 −1 0 0 10 20 30 Months % Change Exchange Rate −3 −2 −1 0 0 10 20 30 Months % Change International Reserves Cumulative MG Response (Equation 3.10) to a 1-standard deviation risk and fundamental component of FTS: Component attributed to the excess risk sentiment component of FTS, ^ V t (dashed) and global demand component, ^ G t (solid) decomposed via Equation A.6. 95% dispersion intervals as computed in Equation 3.12. Sovereign spreads and Industrial Production response in standard deviations. Exchange rate and international reserves response in percent. Negative exchange rate movement is local depreciation vis-a-vis USD. 179 Appendix B Chapter 2 Appendix to \Monetary Policy Spillovers under Intermediate Exchange Rate Regimes" This appendix contains supplemental results from the chapter \Monetary Policy Spillovers under Intermediate Exchange Rate Regimes". Section B.1 and B.2 reports supplementary gures and tables, respectively. 180 B.1 Figures Figure B.1: Distribution ofR 2 for Quarterly regressions where peg intensity estimates equal 1 Density plot for R 2 statistics across all country-quarters which have ^ W b it = 1 for one b (xed exchange rate regimes). Estimated from Equation 2.1. 181 Figure B.2: Change in Peg Intensity from 2000 to 2018 by currency Peg intensities estimated from Equation 2.1, top panel reports change in intensity from 2000 to 2018 with respect to USD peg intensity. Bottom panel reports change in intensity from 2000 to 2018 with respect to EUR intensity. 182 Figure B.3: Peg intensities over time, selected countries Peg intensities estimated from Equation 2.1, with triangle (solid) points denoting peg inten- sity with respect to the EUR and upside-down triangle (clear) points denoting peg intensity with respect to the USD. 183 Figure B.4: Peg intensities over time, cross-country average Each period point refers to the cross-country average of peg intensities. Peg intensities estimated from Equation 2.1, with triangle (solid) points denoting peg intensity with respect to the EUR and upside-down triangle (clear) points denoting peg intensity with respect to the USD. 184 Figure B.5: GAM Estimates: E.U. spillover estimates by Peg Intensity Spillover estimate is under free capital controls (K it = 1). Estimates are from Equation 2.9. Shaded areas are 95% credible intervals. Number of knots selected: 10 via REML. Red dashed line is the implied linear spillover under Equation 2.6. 185 Figure B.6: GAM Estimates: E.U. spillover estimates by Peg Intensity, with knot number set to 5 Spillover estimate is under free capital controls (K it = 1). Estimates are from Equation 2.9. Shaded areas are 95% credible intervals. Number of knots selected: 5. Red dashed line is the implied linear spillover under Equation 2.6. 186 B.2 Tables Table B.1: Country Summary Country T Type 1 Albania 75 EME 2 Argentina 62 EME 3 Australia 75 AE 4 Bahrain 75 EME 5 Brazil 73 EME 6 Bulgaria 75 EME 7 Canada 75 AE 8 Chile 74 EME 9 China 75 EME 10 Colombia 75 EME 11 Croatia 70 EME 12 Czech.Republic 75 EME 13 Denmark 75 AE 14 Hong Kong 75 AE 15 Hungary 75 EME 16 Iceland 74 AE 17 India 75 EME 18 Indonesia 75 EME 19 Israel 75 EME 20 Japan 75 AE 21 Kazakhstan 75 EME 22 Macedonia 72 EME 23 Malaysia 75 EME 24 Mauritius 75 EME 25 Mexico 74 EME 26 Nepal 75 EME 27 New Zealand 75 AE 28 Norway 75 AE 29 Peru 74 EME 30 Philippines 75 EME 31 Poland 75 EME 32 Qatar 75 EME 33 Romania 71 EME 34 Russia 74 EME 35 Saudi Arabia 75 EME 36 Serbia 72 EME 37 Singapore 75 AE 38 South Africa 75 EME 39 South Korea 75 EME 40 Sweden 75 AE 41 Switzerland 75 AE 42 Thailand 75 EME 43 Trinidad and Tobago 74 EME 44 Turkey 66 EME 45 United Kingdom 75 AE 46 Uruguay 65 EME Summary of countries in the main panel. Type denotes Advanced (AE) or Emerging Market Economy (EME), respectively. Column T refers to country sample size of interest rate change observatoins, R it . 187 Table B.2: Peg Intensities to Base Currencies # Country/FX USD (2000) USD (2018) EUR (2000) EUR (2018) JPY (2000) JPY (2018) 1 AED 1.00 1.00 0.00 0.00 0.00 0.00 2 ALL 0.30 0.75 0.09 0.84 0.00 0.08 3 ARS 0.96 0.50 0.03 0.58 0.02 0.16 4 AUD 0.22 0.09 0.30 0.65 0.22 0.05 5 BGN 0.04 0.00 1.00 1.00 0.01 0.00 6 BHD 1.00 1.00 0.00 0.00 0.00 0.00 7 BND 0.74 0.06 0.03 0.00 0.03 0.07 8 BRL 0.99 0.25 0.05 0.22 0.02 0.51 9 CAD 0.69 0.17 0.06 0.11 0.12 0.00 10 CHF 0.13 0.07 0.88 0.52 0.03 0.24 11 CLP 0.68 0.42 0.08 0.00 0.02 0.12 12 CNY 1.00 0.10 0.00 0.00 0.00 0.00 13 COP 0.50 0.30 0.00 0.21 0.09 0.18 14 CZK 0.08 0.00 0.76 0.79 0.04 0.00 15 DKK 0.01 0.00 0.99 0.98 0.00 0.00 16 DZD 0.27 0.45 0.63 0.27 0.06 0.00 17 GBP 0.09 0.00 0.15 0.07 0.00 0.01 18 HKD 0.96 0.93 0.03 0.02 0.03 0.00 19 HRK 0.00 0.01 0.98 0.98 0.02 0.00 20 HUF 0.06 0.00 0.99 0.90 0.00 0.01 21 IDR 0.50 0.29 0.24 0.30 0.32 0.00 22 ILS 0.95 0.37 0.06 0.42 0.04 0.08 23 INR 0.92 0.49 0.02 0.15 0.03 0.01 24 ISK 0.25 0.11 0.45 0.26 0.00 0.18 25 KRW 0.96 0.05 0.02 0.17 0.19 0.05 Estimates of W b it from Equation 2.1 for 2000 and 2018 (quarterly averages), respectively. Base countries b considered are the U.S., E.U., and Japan. 188 Table B.3: Peg Intensities to Base Currencies (cont.) # Country/FX USD (2000) USD (2018) EUR (2000) EUR (2018) JPY (2000) JPY (2018) 26 KWD 0.85 0.82 0.04 0.03 0.04 0.02 27 LKR 0.87 0.74 0.09 0.01 0.04 0.06 28 MKD 0.39 0.25 0.15 0.05 0.01 0.08 29 MXN 0.98 0.27 0.14 0.53 0.09 0.23 30 MYR 0.93 0.48 0.07 0.10 0.03 0.01 31 NOK 0.30 0.00 0.74 0.47 0.09 0.01 32 NPR 0.99 0.82 0.06 0.41 0.01 0.18 33 NZD 0.32 0.16 0.31 0.62 0.09 0.10 34 OMR 1.00 1.00 0.00 0.00 0.00 0.00 35 PEN 0.95 0.68 0.05 0.16 0.07 0.02 36 PHP 0.83 0.95 0.20 0.32 0.12 0.05 37 PKR 0.87 0.96 0.00 0.24 0.09 0.11 38 PLN 0.34 0.00 0.33 0.94 0.00 0.03 39 QAR 1.00 1.00 0.00 0.00 0.00 0.00 40 RON 1.00 0.04 0.11 0.95 0.03 0.03 41 RSD 0.47 0.00 0.14 0.33 0.14 0.13 42 RUB 0.99 0.43 0.08 0.27 0.12 0.00 43 SAR 1.00 1.00 0.00 0.00 0.00 0.00 44 SEK 0.20 0.00 0.68 0.44 0.04 0.08 45 SGD 0.73 0.14 0.06 0.17 0.08 0.04 46 THB 0.76 0.06 0.07 0.08 0.16 0.02 47 TRY 0.59 0.67 0.44 0.60 0.07 0.00 48 TTD 0.96 0.88 0.03 0.03 0.04 0.02 49 TWD 0.92 0.32 0.10 0.04 0.06 0.06 50 UAH 0.96 0.69 0.00 0.31 0.00 0.05 51 UYU 1.00 0.76 0.05 0.33 0.04 0.14 52 ZAR 0.44 0.03 0.39 0.49 0.08 0.26 Estimates of W b it from Equation 2.1 for 2000 and 2018 (quarterly averages), respectively. Base countries b considered are the U.S., E.U., and Japan. 189 Table B.4: Spillover Eects across Peg Intensity Bins: Pooled Model (Equation 2.6) with Exchange Rate Regime Dummies width=1 Bin 1 2 3 4 5 6 ^ W US it (RA, 2) [0,0.1] (0.1,.30] (0.30,.50] (0.50,0.70] (0.70,0.90] (0.90,1] All Countries ^ US -0.011 0.182 -0.046 0.022 0.127* 0.389*** (0.065) (0.160) (0.157) (0.213) (0.065) (0.141) ^ EU 0.285* 0.257 0.614*** 0.998** -0.060 0.418*** (0.156) (0.212) (0.235) (0.404) (0.248) (0.119) Advanced Economies ^ US 0.080* 0.098 0.168*** 0.305*** 0.376*** 0.771*** (0.041) (0.104) (0.056) ( 0.066) (0.050) (0.211) ^ EU 0.444*** 0.303*** 0.399** 1.039*** 0.332*** 0.566*** (0.081) (0.066) (0.171) (0.199) (0.096) (0.085) Emerging Markets ^ US -0.210 0.271 -0.198 -0.064 0.076 0.292** ( 0.159) (0.552) ( 0.348) ( 0.304) (0.073) (0.135) ^ EU 0.198 0.136 0.773** 0.309 -0.771 0.401** ( 0.209) (0.336) (0.386) (1.252) (0.701) (0.171) ***,**,* refer to signicance at the 1%, 5% and 10% level, respectively. Robust standard errors clustered at the Country level. Regression specication of Equation 2.6, using dummy variables for values of ^ W US it (RA, 2). Estimation period Q2 2000 - Q4 2018. Country Fixed Eects included. . 190 Table B.5: Spillover Eects across IRR (2019) Ilzetzki et al. [2019] Exchange Rate Regimes: Pooled Model (Equation 2.13) with Exchange Rate Regime Dummies Floating Fixed IRR Classication 1 2 3 4 5 All Countries ^ US (IRR) 0.019 0.208*** -0.126 0.474*** 0.516** (0.107) (0.069) (0.084) (0.162) (0.211) Advanced Economies ^ US (IRR) 0.194*** 0.328*** 0.768*** 0.268*** 1.069*** (0.056) (0.047) (0.216) (0.075) (0.025) Emerging Markets ^ US (IRR) -0.235 0.160 -0.213*** 0.862*** 0.282 (0.214) (0.102) (0.006) (0.276) (0.181) ***,**,* refer to signicance at the 1%, 5% and 10% level, respectively. Robust standard errors clustered at the Country level. Estimation period Q2 2000 - Q4 2016. Country Fixed Eects included. 191 Table B.6: Imputing Shadow Rates at the ZLB (1) (2) (3) All Advanced Emerging Countries Economies Markets ^ US 0.319*** 0.643*** 0.220** (0.087) (0.116) (0.089) ^ EU 0.189*** 0.233*** 0.142 (0.067) (0.079) (0.110) Adj. R 2 0.14 0.39 0.13 F-Statistic 57.11 50.85 39.03 NT 2,532 644 1,887 Country FE Y Y Y Time FE N N N ***,**,* refer to signicance at the 1%, 5% and 10% level, respectively. Robust standard errors clustered at the Country level. Regression specication of Equation 2.6, withR bt values for U.S. and E.U. at the ZLB imputed using Wu and Xia [2016] shadow rates. Estimation period: Q2 2000 - Q4 2018. Peg intensity estimate used is ^ W b it (RA, 2). Within adjusted R-squared reported. Table B.7: FOMC Monetary Policy Shocks (1) (2) (3) All Advanced Emerging Countries Economies Markets ^ US 0.944** 1.049*** 0.867 (0.392) (0.281) (0.534) ^ EU 0.535*** 0.817*** 0.239 (0.128) (0.117) (0.168) Adj. R 2 0.13 0.33 0.13 F-Statistic 54.85 42.61 38.88 NT 2,532 644 1,887 Country FE Y Y Y Time FE N N N ***,**,* refer to signicance at the 1%, 5% and 10% level, respectively. Robust standard errors clustered at the Country level. Regression specication of Equation 2.14. Estimation period: Q2 2000 - Q4 2018. Peg intensity estimate used is ^ W b it (RA, 2). FOMC monetary policy shocks are implied yield changes from front month Fed Funds Futures contracts over the day of an FOMC announcement. Changes are aggregated to quarterly frequency. Within adjusted R-squared reported. 192 Table B.8: Omitting the 2008 Global Financial Crisis (1) (2) (3) All Advanced Emerging Countries Economies Markets ^ US 0.522*** 0.616*** 0.474*** (0.127) (0.121) (0.166) ^ EU 0.398*** 0.575*** 0.183 (0.152) (0.102) (0.343) Adj. R 2 0.12 0.39 0.11 F-Statistic 42.88 44.78 28.91 NT 2,120 539 1,580 Country FE Y Y Y Time FE N N N ***,**,* refer to signicance at the 1%, 5% and 10% level, respectively. Robust standard errors clustered at the Country level. Regression specication of Equation 2.6. Estimation period: Q2 2000 - Q4 2018 but omitting crisis window of Q1 2008 - Q4 2010. Peg intensity estimate used is ^ W b it (RA, 2). Within adjusted R-squared reported. Table B.9: Before the Yuan entered the SDR (Pre-2016) (1) (2) (3) All Advanced Emerging Countries Economies Markets ^ US 0.367*** 0.754*** 0.255** (0.126) (0.217) (0.166) ^ EU 0.497*** 0.766*** 0.186 (0.135) (0.115) (0.343) Adj. R 2 0.13 0.46 0.14 F-Statistic 48.55 56.83 32.10 NT 2,157 556 1,600 Country FE Y Y Y Time FE N N N ***,**,* refer to signicance at the 1%, 5% and 10% level, respectively. Robust standard errors clustered at the Country level. Regression specication of Equation 2.6. Estimation period: Q2 2000 - Q4 2015, omitting period with the Yuan entering the SDR Basket (as of 2016). Peg intensity estimate used is ^ W b it (RA, 2). Within adjusted R-squared reported. 193 Table B.10: Including lower-dimension interaction terms (Equation 2.15) (1) (2) (3) All Countries Advanced Economies Emerging Markets ^ 1;US -0.608** -0.535*** -0.534 (0.274) (0.118) (0.346) ^ 4;US 0.592* 0.408 0.538 (0.315) (1.488) (0.382) ^ 5;US 0.635** 0.562*** 0.407 (0.301) (0.120) (0.427) ^ 7;US -0.293 0.213 -0.151 (0.398) (1.570) (0.475) ^ 1;EU 0.127 -0.053 0.171 (0.203) ( 0.388) (0.210) ^ 4;EU 0.209 -0.570 0.286 (0.395) (1.556) (0.475) ^ 5;EU 0.162 0.460 -0.004 (0.278) (0.411) (0.327) ^ 7;EU -0.119 0.825 -0.328 (0.504) (1.613) (0.691) Adj. R 2 0.15 0.46 0.133 F-Statistic 26.98 30.25 17.90 NT 2,532 644 1,887 Country FE Y Y Y Time FE N N N ***,**,* refer to signicance at the 1%, 5% and 10% level, respectively. Robust standard errors clustered at the Country level. Regression specication of Equation 2.15. Estimation period: Q2 2000 - Q4 2018. For peg intensities using 2-quarter rolling average, ^ W b it (RA, 2). Estimated peg intensities are from Equation 2.1. Within adjusted R-squared reported. 194 Appendix C Chapter 3 Appendix to \Global Demand Spillovers and Financial Stability near the Zero Lower Bound" This appendix is organized as follows: Section C.1 discusses the procurement and cleaning of the data used in the analysis. Section C.2 discusses further results from robustness checks. Finally supplementary tables and gures are provided. C.1 Data The data to construct the index of global demand pressure comes from the World Bank's commodity market "Pink Sheet" database which reports monthly prices across a broad range of commodities since 1960. I convert monthly prices to quarterly by taking the last monthly value within each quarter-year. A list of commodities can be found in Table C.2. The data primarily used in the analysis is composed of a cross-country unbalanced panel from Q1 1979 to Q4 2019 covering N = 17 advanced economies. These countries are: Australia, Austria, Belgium, Canada, Finland, France, Germany, Italy, Japan, Netherlands, New Zealand, Norway, Spain, Sweden, Switzerland, United Kingdom, and United States. Under full data availability a country has T = 163 observations, though some variables are not populated for each country since the beginning of the sample period, resulting in an unbalanced panel. Quarterly short-term interest rates, real GDP growth, and in ation are from the GVAR Database [Mohaddes and Raissi, 2020]. Daily country equity index prices are from Datastream to construct quarterly equity returns and realized volatility. growth rates and returns correspond to quarterly changes in natural logarithms. Central bank policy 195 Table C.1: Summary Statistics Statistic NT Mean St. Dev. Min Pctl(25) Pctl(75) Max Short Term Rate 2,771 5.260 4.740 0.904 1.453 8.453 25.783 Equity Returns 2,651 1.782 10.430 53.782 2.938 7.471 64.918 Equity Realized Volatility 2,651 15.876 8.404 2.906 10.383 18.758 77.436 Central Bank Policy Rate 2,051 4.097 4.507 0.750 1.000 5.500 50.000 Real GDP Growth 2,771 0.502 0.878 6.011 0.080 0.973 4.679 In ation 2,771 0.780 0.868 2.158 0.276 1.000 8.543 Real House Price Growth 2,483 0.443 2.232 10.733 0.768 1.725 11.557 Summary statistics of raw data, pre-interpolation. All measures are in percentage units (%). Equity returns, GDP growth, in ation, and real house price growth are quarterly log-dierences. Equity realized volatility is annualized. Maximum time period: Q2 1979 - Q4 2019. Countries in the panel: Australia, Austria, Belgium, Canada, Finland, France, Germany, Italy, Japan, Netherlands, New Zealand, Norway, Spain, Sweden, Switzerland, United Kingdom, and United States. rates and real residential house prices are taken from the Bank for International Settlements (BIS) website. For Japan's policy rate values were missing from 2001 to 2006 and 2013 to 2016, these missing values were interpolated using the last available value. For the EU countries, policy rates are substituted with each country's short-term interest rate for the pre-ECB period (from the beginning of the sample through Q4 1998). C.2 Robustness Checks C.2.1 Industrial commodity price factor For commodity prices to make good proxies for global demand, certain criteria must be met as discussed in Alquist et al. [2020]. Critically, some requirements include that vertical inte- gration should be minimal so idiosyncratic shocks in one commodity market does not aect others, and commodities which may substitute for nancial assets (e.g. gold) should also be excluded. As a robustness check, I compute an second proxy for global demand spillovers more aligned with these requirements using the average return across only industrial com- modity prices: crude oil, rubber, coal, aluminum, iron ore, copper, lead, tin, nickel, zinc. The correlation between this index and the original index for global demand spillovers is 0.85. Using the same 3.2% policy rate threshold, the MG IRFs are reported in Figure C.2, showing that the baseline results are broadly unchanged, with the dierential response in real output growth slightly widening. 196 C.2.2 Pre-2008 period To explore the extent the post-GFC period is driving the results of shock amplication near the ZLB, I re-estimate the threshold-augmented var only using data up to Q4 2007. I leave the same 3.2% interest rate threshold, but three countries did not see policy rates fall below this level before the GFC: Australia, New Zealand and the United Kingdom, and most countries did not hit the ZLB at this point. Therefore, the sample size for the `near ZLB' regime IRFs is reduced to 14 rather than 17 countries. Figure C.3 reports the MG IRFs for the estimates using just the pre-2008 sample. The results weaken somewhat as expected given many country samples do hit the ZLB, but still the response of real output to a negative demand shock was much larger when policy rates neared the ZLB. After ve quarters, real output contracts on average -0.10% when policy rates were above the threshold, and -0.24% when rates fell below 3.2%. The response of equity market returns continue to dier sharply depending on the policy rate. A negative demand shock still leads to signicantly lower stock market prices when near the ZLB, while stock prices rise in response otherwise. Policy rates continue to fall more when not constrained by ZLB risk. Compared to the full sample results, the policy rate when near the ZLB falls more during the pre-2008 period, a result driven by the fact that while below 3.2%, they still had room to fall (i.e. they are closer to 3.2% than zero). By contrast, the response of in ation changes compared to the full-sample estimation: in ation rose in response to a negative demand shock near the ZLB, though only by about 0.10% after ve quarters. C.2.3 Commodity Exporters Commodity exporters may behave very dierently than other advanced economies, especially to demand measures based on commodity price uctuations which aect these countries directly by altering the terms-of-trade. In Figure C.4 I report the MG IRFs but exclude the four commodity exporters in the sample: Australia, Canada, New Zealand and Norway. All of the baseline results broadly hold, though reducing the sample size leads to wider condence intervals. The overall results of a policy rate threshold for global demand spillovers are not driven by the subgroup of commodity exporting countries. 197 Figure C.1: Policy Interest Rates and Periods Below Threshold of 3.2% 0 10 20 30 40 1980−1 1990−1 2000−1 2010−1 2020−1 Australia 0 3 6 9 12 1980−1 1990−1 2000−1 2010−1 2020−1 Austria 0 5 10 15 1980−1 1990−1 2000−1 2010−1 2020−1 Belgium 0 5 10 15 20 1980−1 1990−1 2000−1 2010−1 2020−1 Canada 0 5 10 15 1980−1 1990−1 2000−1 2010−1 2020−1 Finland 0 5 10 15 1980−1 1990−1 2000−1 2010−1 2020−1 France 0.0 2.5 5.0 7.5 10.0 12.5 1980−1 1990−1 2000−1 2010−1 2020−1 Germany 0 5 10 15 20 1980−1 1990−1 2000−1 2010−1 2020−1 Italy 0.0 2.5 5.0 7.5 1980−1 1990−1 2000−1 2010−1 2020−1 Japan 0.0 2.5 5.0 7.5 10.0 12.5 1980−1 1990−1 2000−1 2010−1 2020−1 Netherlands 0 10 20 30 40 50 1980−1 1990−1 2000−1 2010−1 2020−1 New Zealand 0 5 10 15 1980−1 1990−1 2000−1 2010−1 2020−1 Norway 0 5 10 15 20 1980−1 1990−1 2000−1 2010−1 2020−1 Spain 0 10 20 30 40 1980−1 1990−1 2000−1 2010−1 2020−1 Sweden 0 2 4 6 1980−1 1990−1 2000−1 2010−1 2020−1 Switzerland 0 5 10 15 1980−1 1990−1 2000−1 2010−1 2020−1 United Kingdom 0 5 10 15 20 1980−1 1990−1 2000−1 2010−1 2020−1 United States Shaded area indicates policy rates below 3.2%. For E.U. countries, policy rate is replaced with the country's respective short-term interest government interest rate for the period preceding the monetary union. 198 Figure C.2: Impulse Response Functions to a 1-SD Negative Global Demand Shock (in- dustrial commodity prices) when Policy Rates are Above (Solid) and Below (Dashed) 3.2% Threshold −0.3 −0.2 −0.1 0.0 5 10 Quarters Real GDP (%) −0.2 −0.1 0.0 5 10 Quarters Inflation (%) −0.2 0.0 0.2 0.4 0.6 5 10 Quarters Real House Price (%) −0.4 −0.2 0.0 5 10 Quarters Policy Rate (%) −4 −2 0 2 5 10 Quarters Equity Return (%) 0 1 2 5 10 Quarters Realized Equity Volatility (%) All IRFs are cumulative except for realized equity volatility. Mean Group IRF estimated from Equation 3.10 along with 90% cross-sectional error bands reported (Equation 3.12). IRFs based on the VAR equations from Equation 3.7. Global demand spillovers estimated as the average real return of USD-denominated industrial commodity prices (crude oil, rubber, coal, aluminum, iron ore, copper, lead, tin, nickel, zinc). 199 Figure C.3: Impulse Response Functions to a 1-SD Negative Global Demand Shock (pre-2008 sample) when Policy Rates are Above (Solid) and Below (Dashed) 3.2% Threshold −0.4 −0.3 −0.2 −0.1 0.0 5 10 Quarters Real GDP (%) −0.1 0.0 0.1 0.2 5 10 Quarters Inflation (%) −1.5 −1.0 −0.5 0.0 5 10 Quarters Real House Price (%) −0.3 −0.2 −0.1 0.0 5 10 Quarters Policy Rate (%) −2 0 2 5 10 Quarters Equity Return (%) −0.25 0.00 0.25 0.50 5 10 Quarters Realized Equity Volatility (%) All IRFs are cumulative except for realized equity volatility. Mean Group IRF estimated from Equation 3.10 along with 90% cross-sectional error bands reported (Equation 3.12). IRFs based on the VAR equations from Equation 3.7. 3 of the 17 countries did not have policy rates under 3.2% in the pre-2008 sample: Australia, New Zealand, United Kingdom. 200 Figure C.4: Impulse Response Functions to a 1-SD Negative Global Demand Shock (exclud- ing commodity exporters) when Policy Rates are Above (Solid) and Below (Dashed) 3.2% Threshold −0.4 −0.3 −0.2 −0.1 0.0 5 10 Quarters Real GDP (%) −0.4 −0.3 −0.2 −0.1 0.0 5 10 Quarters Inflation (%) −0.50 −0.25 0.00 0.25 0.50 5 10 Quarters Real House Price (%) −0.4 −0.2 0.0 5 10 Quarters Policy Rate (%) −4 −2 0 2 5 10 Quarters Equity Return (%) 0 1 2 5 10 Quarters Realized Equity Volatility (%) All IRFs are cumulative except for realized equity volatility. Mean Group IRF estimated from Equation 3.10 along with 90% cross-sectional error bands reported (Equation 3.12). IRFs based on the VAR equations from Equation 3.7. 4 of the 17 countries are classied as commodity exporters and excluded in the estimation: Australia, Canada, New Zealand, Norway. 201 Figure C.5: Timet Estimated Global Demand Shock (Solid), Timet+1 Real Output Growth of China (Dashed) −5.0 −2.5 0.0 2.5 1980−1 1990−1 2000−1 2010−1 2020−1 Standard Deviations Correlation between the two series equals 0.32 (t=4.32). Global demand shock is the estimated residual ^ e gt from Equation 3.13. 202 Table C.2: Global Demand Pressure: PCA Weights on Commodity Components Commodity PCA Loading Phosphate Rock -0.08 Tobacco -0.02 Natural Gas (Europe) -0.00 Natural Gas (Japan) -0.00 Tea (Kolkata) 0.00 Shrimp 0.01 Oranges 0.01 Plywood 0.02 Logs 0.02 Potash 0.03 Chicken 0.03 Bananas 0.04 Natural Gas (Index) 0.04 Natural Gas (U.S.) 0.05 Sawnwood 0.06 Sugar 0.07 Fish Meal 0.08 Tea (Colombo) 0.08 Coee (Arabica) 0.08 Coee (Robusta) 0.08 Beef 0.09 Ground Nut Oil 0.09 Sugar (World) 0.09 Tea (Average) 0.09 Cocoa 0.09 Urea 0.10 Tea (Mombasa) 0.10 Rice 0.10 Gold 0.11 Logs (Cameroon) 0.12 Fertilizer (DAP) 0.13 Iron Ore 0.15 Wheat (SRW) 0.15 Coal 0.15 Cotton 0.15 Coconut Oil 0.15 Sugar (Europe) 0.15 Fertilizer (TSP) 0.16 Wheat (HRW) 0.16 Zinc 0.16 Silver 0.17 Nickel 0.17 Barley 0.17 Crude Oil (Dubai) 0.17 Lead 0.17 Soybean Meal 0.17 Maize 0.18 Crude Oil (WTI) 0.18 Crude Oil (Brent) 0.18 Platinum 0.18 Palm Oil 0.18 Tin 0.19 Sorghum 0.19 Rubber 0.20 Soybean Oil 0.21 Aluminum 0.21 Soybean 0.21 Copper 0.21 203

Abstract (if available)

Abstract This thesis brings together three research papers which investigate empirically the measurement and impact of global financial, commodity, and monetary shocks across advanced and emerging economies. I put particular emphasis on how macroeconomic policiesㅡdomestic and internationalㅡshape the transmission of these shocks to financial markets and the real economy. ❧ The first paper studies the international dimension of global financial shocks known as 'flights-to-safety'. Financial market imperfections point toward large macroeconomic costs associated with flights-to-safety in the absence of policy intervention. I investigate this implication empirically by developing a measure of global flights-to-safety and modeling their impact on emerging markets. Defined as joint tail realizations across developed market risky and safe asset returns, large flights-to-safety map to unexpected tail events and shape future world commodity prices, interest rates and U.S. Dollar fluctuations. In emerging markets, a global flight-to-safety induces a sharp rise in sovereign risk and exchange market pressure followed by a protracted drop in economic activity. These effects are substantially larger than those of U.S. monetary policy shocks and domestic financial shocks. Heterogeneity in adjustment patterns across countries suggest financial disruption as a key transmission channel but also a role for policy intervention: The impact of flights-to-safety on economic activity is amplified in countries realizing sharper adjustment in financial conditions, four times larger in emerging markets with U.S. exchange traded funds, and mitigated through 'leaning against the wind' with international reserves. ❧ The second paper explores whether intermediate exchange rate regimes such as managed floats grant the degree of monetary independence implied by the international Trilemma. Testing the international Trilemma traditionally relies on discretely classified exchange rate regimes. This simplification limits the implications drawn for middle-ground policies like managed floats or basket pegs, and inhibits inference on the empirical shape of the exchange rate stability-monetary autonomy trade-off. To address these issues, this paper proposes a continuous measure of exchange rate flexibility for estimating monetary policy spillovers along the entire spectrum of peg intensities. Monetary spillovers generally increase with exchange rate stability, even within middle ground policies, and basket pegs diversify such spillovers. I then estimate the empirical shape of the trade-off using machine learning techniques, finding that the relationship between monetary autonomy and exchange rate stability is significantly non-linear in both advanced economies and emerging markets. Specifically, partially targeting the exchange rate translates to disproportionately smaller or larger monetary spillovers along middle-ground exchange rate regimes. For emerging markets in particular, active reserves management is a key mechanism associated with these non-linearities. ❧ The third and final paper is preliminary work revisiting the Zero Lower Bound (ZLB) irrelevance hypothesis, which states that the ZLB constraint on monetary policy does not amplify the macroeconomic adjustment to shocks. This contradicts a wide class of macroeconomic models predicting that shocks are amplified at the ZLB because the effects of these shocks cannot be offset by lowering interest rates, thereby causing real interest rates to increase rather than decrease - potentially triggering a deflationary spiral. Much of the literature empirically testing this hypothesis has found that the ZLB is irrelevant: that macroeconomic adjustment to recessionary shocks are not amplified near the ZLB. I argue, however, that these studies are subject to the Lucas Critique because the size of a domestic shock is endogenously determined in part by the available monetary policy space. To overcome this issue, I evaluate the ZLB irrelevance hypothesis from a multi-country perspective, investigating country-specific adjustment to world shocks across a panel of 17 advanced economies from 1979 to 2019. Contrasting the prevailing literature, preliminary results point toward a rejection of the ZLB irrelevance hypothesis. This work will set the foundation for a full research agenda over the coming years.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Three essays in international macroeconomics and finance

PDF

Essay on monetary policy, macroprudential policy, and financial integration

PDF

Emerging market financial crises, investors and monetary policy

PDF

Early-warning systems for crisis risk

PDF

Essays on interest rate determination in open economies

PDF

Essays on business cycle volatility and global trade

PDF

Essays on innovation, human capital, and COVID-19 related policies

PDF

Essays on narrative economics, Climate macrofinance and migration

PDF

Three essays on heterogeneous responses to macroeconomic shocks

Asset Metadata

Creator Ahmed, Rashad (author)

Core Title Essays on monetary policy and international spillovers

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Publication Date 03/26/2021

Defense Date 03/11/2021

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag capital flows,commodity cycles,exchange rates,financial spillovers,financial stability,flight to safety,global shocks,interest rates,international reserves,OAI-PMH Harvest,risk-off,sovereign risk,tail risk,zero lower bound

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Aizenman, Joshua (committee chair), Pesaran, M. Hashem (committee member), Ramcharan, Rodney (committee member), Ranciere, Romain (committee member), Zeke, David (committee member)

Creator Email rashad.ahmed334@gmail.com,rashadah@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-432254

Unique identifier UC11666549

Identifier etd-AhmedRasha-9366.pdf (filename),usctheses-c89-432254 (legacy record id)

Legacy Identifier etd-AhmedRasha-9366.pdf

Dmrecord 432254

Document Type Dissertation

Rights Ahmed, Rashad

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA