U.S. Real Interest Rates and Default Risk in Emerging Economies

Board of Governors of the Federal Reserve System International Finance Discussion Papers Number 1051 August 2012 US real interest rates and default risk in emerging economies Nathan Foley-Fisher and Bernardo Guimaraes NOTE: International Finance Discussion Papers are preliminary materials circulated to stimulate discussion and critical comment. References to International Finance Discussion Papers (other than an acknowledgement that the writer has had access to unpublished material) should be cleared with the author or authors. Recent IFDPs are available on the Web at www.federalreserve.gov/pubs/ifdp/. This paper can be downloaded without charge from Social Science Research Network electronic library at www.ssrn.com.

US real interest rates and default risk in emerging economies ∗ Nathan Foley-Fisher† Bernardo Guimaraes‡ August 2012 Abstract This paper empirically investigates the impact of changes in US real interest rates on sovereign default risk in emerging economies using the method of identification through heteroskedasticity. Policy-induced increases in US interest rates starkly raise default risk in emerging marketeconomies. However,the overallcorrelationbetween US realinterest rates andtheriskofdefaultisnegative,demonstratingthattheeffectsofothervariablesdominate the anterior relationship. Keywords: real interest rates; default risk; sovereign debt; identification through heteroskedasticity. JEL Classification: F34, G15 ∗Theviewsinthispaperaresolelytheresponsibilityoftheauthorsandshouldnotbeinterpretedasreflecting the views of the Board of Governors of the Federal Reserve System or of any other person associated with the Federal ReserveSystem. We thankseminar participants at the EBRD and two anonymousreferees for helpful suggestions. †International Finance Division, Federal Reserve Board, Washington, D.C., USA; phone: 1-202-452-2350; email: nathan.c.foley-fisher@frb.gov ‡bernardo.guimaraes@fgv.br, Sao Paulo School of Economics - FGV

1 Introduction The theoretical economic effect of changes in US real interest rates on default risk in emerging economies has been studied by, amongst others, Guimaraes (2011) and the channelisoftencitedasanon-domesticdriver ofcountryriskpremia(Neumeyer and Perri 2005). The mechanism runs that when US real interest rates rise, the opportunity costs to those who buy emerging economies’ debt increase, which raises interest rates in emerging economies. This direct effect increases the debt burden on emerging economies, raising the risk that they will default on their debt and requiring emerging economies to offer even higher interest rates in compensation. Anecdotal evidence from the Latin American debt crisis of the 1980’s and the Mexican crisis in 1994, both of which were preceded by sharp interest rate hikes in the US, suggests that this theoretical channel might be an important empirical one. Empirically identifying this theoretical relationship is not trivial, however, owing to the usual problems of reverse causality and common omitted variables. The latter is especially problematic because US real interest rates and default risk in emerging economies are both affected by variables that cannot be easily measured, such as global market factors, risk appetite, and expectations about economic performance and the political scenario. This paper identifies the effects of changes in US real interest rates on default risk in emerging economies using the method of identification through heteroskedasticity as set out by Rigobon (2003) and Rigobon and Sack (2004). As discussed in detail in Section 2, we take data on US real interest rates from inflation-indexed Treasury bonds, and proxy default risk using J.P. Morgan’s Emerging Markets Bond Index Plus (EMBI+) premia in emerging economies over the period between 1998 and 2008. The idea behind the identification method is that there is a greater variance of changes in real interest rates on dates when the Federal Open Market Committee (FOMC) meets. The meetings of the FOMC can be seen as an extra shock to US interest rates, which have an impact on the EMBI+ premia. The key identifying assumption is that the timing of FOMC meetings does not affect the EMBI+ premia through any channel other than the changes in real interest rates. Other shocks that directly affect the EMBI+ premia are assumed to be uncorrelated with thetimingofFOMCmeetings. Thisassumptionresemblesthedesiredcharacteristicsofan instrument inIVregressions. However, thetiming ofFOMC meetings affectsthevariance, not the level of shocks, so a usual IV strategy cannot be employed. The methodology of identificationthroughheteroskedasticity yieldsasyntheticinstrument basedondifferences 2

in the covariance matrices of our data between dates when the FOMC does and does not meet. Our findings are presented in Section 3, where we show that unexpected policyinduced increases in interest rates lead to greater EMBI+ premia and, by implication, default risk in emerging economies. A 1 basis-point increase in 10-year US real interest rates raises EMBI+ premia by around 1 basis point, which means that the cost of borrowing in emerging economies rises substantially more than in the US. This confirms the hypothesised theoretical relationship between changes in US real interest rates and the risk of default and suggests that more attention ought to be paid to this relationship in the literature on default risk. Apositive correlationbetween defaultriskandUSrealinterest rateswouldimplythat emerging economies should issue debt contingent on US real interest rates because such a contingency would negate the increased default risk not associated with fundamental changes in emerging economies. Note, however, that this policy prescription depends not onthecausal relationshipbetween USrealinterest ratesandtheEMBI+ premium, but on the correlation between both. Omitted variables that significantly affect this correlation would also affect the performance of debt contracts contingent on US real interest rates. In actuality, on dates when the FOMC does not meet, we observe a significant correlation with the opposite sign: changes in real interest rates are negatively related to changes in EMBI+ premia. Moreover, the overall correlation between real interest rates and the EMBI+ premium is negative: a 2 bp increase in the 10-year US real rate is on average related to a 1 bp decrease in the EMBI+. The results suggest that high real interest rates reflect favourable external conditions for emerging markets, which reduce the risk of default. This finding resonates with that of Longstaff et al. (2011), where global risk factors (proxied by US markets) are shown to be the major determinant of sovereign credit risk premia. Regardless of the precise reason for the negative correlation, the policy implication is clear: emerging economies should not issue debt contingent on US real interest rates. Previous academic work has attempted to establish the nature of the relationship between US real interest rates and sovereign default risk by applying different methods to deal with the aforementioned endogeneity problems. Some of this work has relied on structural assumptions in vector autoregressions to identify the relationship (e.g., Uribe and Yue 2006). For our purposes, high-frequency data on financial prices can provide more information and allow for a cleaner identification strategy.1 1Uribeand Yue(2006)alsostudytheeffectofinterestratesandtheEMBI+premiumonvariableslikeoutput, 3

An alternative to structural assumptions are ‘traditional’ instruments in IV strategies, such as in Zettelmeyer (2004), where changes in the policy rate are employed as instruments for longer-term real interest rates. This methodology also needs to assume that changes in the instrument do not affect EMBI+ premia through alternative channels. Moreover, the instruments themselves must be exogenous, which is a stronger, and therefore less desirable, assumption than that employed in this paper. Additional studies investigate the direct effect of changes in the US federal funds target rate on emerging market spreads (Arora and Cerisola 2001). However, the theoretical relationship of interest is between default risk and the longer-term real interest rate, not the short-term nominal rate, which cannot be assumed to be endogenous. Moreover, even changes in the target rate might not be truly exogenous (see Rigobon and Sack 2004). In a more closely related exercise, Robitaille and Roush (2006) employ an event study approach using Brazilian data and find similar results to those of our paper. 2 Data and empirical methodology Our measure of the interest rate, i, is from 10-year inflation-indexed Treasury bonds.2 To quantify the risk of default, e, we use J.P. Morgan’s Emerging Markets Bond Index Plus (EMBI+), which is comprised of medium-term debt of more than one year to maturity.3 All data are obtained from the Global Financial Database (www.globalfinancialdata.com). We want to obtain long data series with minimal concern for events that might obfuscate a potential relationship. For this reason we select emerging economies that have not defaulted, and use daily data running from January 1998 to December 2008. We are interested in how a change in the interest rate affects the EMBI+ premia, so our sample consists of values of ∆e = e e and ∆i = i i and is divided in two: the t t+1 t−1 t t+1 t−1 − − sub-sample C corresponds to the dates of monetary policy shocks, and the sub-sample N corresponds to dates with no shocks.4, 5 There are two endogeneity concerns that mean a simple ordinary least squares regression will not identify the effect of changes in US real interest rates on the risk of default (EMBI+ premia). First, changes in the EMBI+ premia can cause changes in the interest and in that case ourmethodology cannot beapplied. 2Our analysis is robust to the use of alternative measures of the real interest rate based on inflation-adjusted nominal Treasury rates of 3 monthsand 10 years. SeeAppendixA. 3EMBI+ tracks total returns for traded US dollar- and other external currency-denominated Brady bonds, loans, Eurobondsand local market instruments. 4For additional justification for using datain differencesrather than levels, see AppendixB. 5Sub-sample C contains the dates of scheduled and unscheduled FOMC meetings and the Federal Reserve Chairman’s semi-annual monetary policy testimony to Congress. For a full list of these dates, see http://www.federalreserve.gov/monetarypolicy/fomccalendars.htm 4

rate, for example, when default risk falls and in response investors switch demand from safe Treasury assets to emerging market debt. Second, and more importantly, the interest rate and the exchange rate are influenced by other common omitted variables. The following system of equations is a simple representation of both endogeneity issues6: ∆e = α∆i +z +η (1) t t t t ∆i = β∆e +γz +ε (2) t t t t Where ∆i is the change in US real interest rate; ∆e the change in the EMBI+ premium; t t z a vector of omitted variables including, for example, external market conditions; ε a t t monetary policy shock; and η a shock to EMBI+. t The objective is to identify α in Equation 1. Our identification strategy is borrowed from Rigobon and Sack (2004), who show that the impact of monetary policy shocks on asset prices can be identified because the variance of shocks is substantially larger on the days in sub-sample C. Their paper used the identification strategy to establish a significant response of 10-year Treasury yields to monetary policy shocks. That monetary policy shocks can influence 10-year real interest rates means that the variance of changes in these rates is significantly larger on the days in sub-sample C. This effect isnot large, but islargeenoughtosignificantly affect thevarianceof∆i . Weexploit t this effect by combining it with the assumption that the policy shock to real interest rates neither affects EMBI+ through z nor η , but only through its effect on ∆i. t t In sum, we assume that the variance of interest rate shocks (ε ) in sub-sample C is t higher than the variance in sub-sample N; whilst the variances of η and z are the same t t across both sub-samples. As is usual in other identification strategies for our underlying system ofequations, weassumez , ε andη havenoserialcorrelationandareuncorrelated t t t with each other. Our assumptions can be written in terms of the second moments of the shocks in the two sub-samples C and N in the following way: σC > σN ε ε σC = σN η η σC = σN z z Tohelpjustifytheunderlying assumptions, Table1shows theincreaseinthevariation in the US real interest rate and the change in covariance between the real interest rate 6Weshow in AppendixC that allowing for a richer lag structuredoes not materially affect theresults. 5

and EMBI+ premia over the sub-samples. The fact that the standard deviations of EMBI+ premia appear to decrease from sub-sample N to sub-sample C, when we expect mild increases, suggests that we require a more accurate statistical test of whether our assumptions on the variance of shocks over the two sub-samples are valid.7 Applying the test set out in Levene (1960), reported in Table 2, we established that the standard deviationoftherealinterestrateincreasessignificantlyinsub-sampleC,whilethevariance of EMBI+ does not significantly change because the effect of the variance increase in Equation 2 only weakly effects the variance of EMBI+ through the interest rate.8 Table 1: Data descriptives Standard Covariance with deviation US real rate Sub-sample C Sub-sample N Sub-sample C Sub-sample N US real rate 0.093 0.063 . . Emerging Market 24.491 29.020 0.198 -0.211 Latin America 25.017 32.317 0.278 -0.253 Brazil 30.249 48.318 0.357 -0.278 Bulgaria 24.476 27.181 0.175 -0.117 Mexico 19.221 21.876 0.066 -0.214 Panama 12.486 14.849 0.028 -0.208 Peru 20.892 20.939 0.128 -0.185 Venezuela 43.545 50.526 0.852 -0.263 Note: 131 observations in sub-sampleC, 2,604 daysin sub-sampleN. We are not assuming that the FOMC ignores factors that affect emerging market default risk, nor are we supposing that FOMC decisions have no impact on emerging market prices – that is actually the effect we are estimating. We are precisely assuming that FOMC decisions do not directly reveal important information about emerging markets that might otherwise affect EMBI+ premia, they are only affecting EMBI+ premia through changes in US real interest rates. The underlying view is that the Committee might have private information about how it will react to movements in emerging markets and how it plans to conduct monetary policies in general but does not know more than the market about emerging economies. 7Wecannotapplystandardtestsofvarianceequality,becausetheyrequirethattheunderlyingdatabenormally distributed. Asisreportedin AppendixD,demonstrated throughplotsofeach variables’quantilesagainst those ofthenormaldistributionandempiricaltestsofskewnessandkurtosis,noneofourseriesarenormallydistributed. 8Althoughthetestresultsarepresentedusingthesamplemeanof thedata,similar resultsareobtainedwhen using the50th percentile or 10% trimmed mean. 6

Table 2: Levene (1960) test of equal variance Test statistic p-value based on mean US real rate 12.371 0.000 Emerging Market 0.215 0.643 Latin America 0.458 0.499 Brazil 2.273 0.132 Bulgaria 0.000 0.977 Mexico 0.031 0.860 Panama 0.021 0.884 Peru 0.908 0.341 Venezuela 0.635 0.801 Note: Null hypothesisis equalvariance Now, consider the following variables: ∆i′ ∆i′ ′ ∆I C , N ≡ (cid:20)√T √T (cid:21) C N ∆e′ ∆e′ ′ ∆E C , N ≡ (cid:20)√T √T (cid:21) C N ∆i′ ∆i′ ′ w C , − N ≡ (cid:20)√T √T (cid:21) C N A major result in Rigobon and Sack (2004) is that α can be consistently estimated by a standard instrumental variables approach with the novel instrument, w, which is correlated with the dependent variable, ∆I, but is neither correlated with z nor η . It t t is correlated with ∆I because the greater variance in sub-sample C implies the positive correlation between ∆i′ /√T and ∆i′ /√T more than outweighs the negative cor- C C C C relation between ∆(cid:0)i′ /√T a(cid:1)nd (cid:0)∆i′ /√T (cid:1). It is neither correlated with z nor η N N − N N t t because the positi(cid:0)ve and nega(cid:1)tive c(cid:0)orrelation of e(cid:1)ach part of the vector cancel each other out. The usual assumption in IV regressions is that the instrument affects the dependent variable only through the regressor. The key difference here is that instead of having a variable assumed to be correlated with ε and uncorrelated with any of the other variables, we assume that the variance of ε is larger on the days in sub-sample C and the variances of other variables are the same in both sub-samples. 7

3 Results Table 3 presents the results from implementing our identification strategy, which reveals that policy shocks to real interest rates are positively correlated with emerging economies’ EMBI+. This coincides with our original intuition that when the US tightens monetary policy, it is harder for emerging economies to borrow, and the risk of default proxied by EMBI+ increases. Table 3: The response of EMBI+ premia to interest rate shocks Co-eff Std Err T-stat Emerging Market 0.868 0.179 4.840 Latin America 1.115 0.195 5.717 Brazil 1.334 0.269 4.969 Bulgaria 0.649 0.170 3.808 Mexico 0.607 0.138 4.394 Panama 0.496 0.094 5.264 Peru 0.659 0.140 4.697 Venezuela 2.279 0.318 7.162 Note: Each estimation uses 2,735 observations. The magnitude of the response is large: an unexpected increase in the 10-year real interest rate of one basis point leads to an increase in the EMBI+ premium of a similar order of magnitude. Table 4 shows the results from analysis of the relationship between US real interest rates andEMBI+ premia in each separate sub-sample (theresults across bothsamples are in Table 5). Crucially, the ‘normal’ correlation between ∆E and ∆I is actually negative (and smaller in absolute value) in sub-sample N. Our interpretation is that increases in US real interest rates are correlated with other things that are good for emerging markets and thus decrease their cost of borrowing. Future research ought to investigate which aspects of international financial markets, correlated with US real interest rates, are most important to the risk of emerging market default. The results in Table 3 are substantially different from the OLS estimates using only the sub-sample C presented in Table 4. While the former shows a strong positive relation, the latter shows a mild and insignificant effect. Rosa (2011) has noted that, in some applications, the results from employing the identification through heteroskedasticity methodology are not much different from a simple OLS using the subsample where the FOMC meets. That is not the case here since we are using the long-term interest rates, where endogeneity is likely to be much more important than when the policy rate is used, 8

Table 4: Separate analysis of sub-samples Sub-sample C Sub-sample N Coeff Std Err T-Stat Coeff Std Err T-stat Emerging Market 0.230 0.224 1.029 -0.494 0.087 -5.700 Latin America 0.317 0.228 1.390 -0.591 0.096 -6.131 Brazil 0.406 0.275 1.474 -0.649 0.145 -4.492 Bulgaria 0.217 0.226 0.960 -0.274 0.081 -3.363 Mexico 0.089 0.177 0.503 -0.500 0.065 -7.692 Panama 0.036 0.114 0.311 -0.487 0.044 -11.186 Peru 0.146 0.191 0.766 -0.430 0.062 -6.937 Venezuela 0.924 0.389 2.371 -0.617 0.151 -4.076 Note: 131 observations in sub-sampleC, 2,604 daysin sub-sampleN. Table 5: Full sample analysis Co-eff Std Err T-stat Emerging Market -0.423 0.082 -5.174 Latin America -0.503 0.091 -5.535 Brazil -0.547 0.135 -4.038 Bulgaria -0.226 0.077 -2.934 Mexico -0.443 0.062 -7.194 Panama -0.437 0.041 -10.586 Peru -0.375 0.059 -6.347 Venezuela -0.467 0.143 -3.266 Note: Each estimation uses 2,735 observations. and the correlation between variables in the N sample is different from the causal effect. 4 Concluding remarks The strong and positive relation between exogenous changes in US real interest rates and the EMBI+ premium highlights the importance of US interest rate shocks. The fact that the overall correlation between US rates and the EMBI+ premium is negative highlights the importance of other aspects of international financial markets, such as favourable external conditions to emerging economy borrowing. Froma policy perspective, our result has implications for proposals to issue debt that is contingent on exogenous factors that affect the ability to repay. One of these ideas is that a higher US real interest rate makes it more difficult for emerging market economies to repay, so reducing emerging market debt payments when US interest rates increase would be welfare improving. Our finding that the overall correlation is negative implies that making emerging market sovereign debt contingent on US real interest rates would have an opposite result from the desired 9

effect. Research on sovereign default should note that shocks affecting foreignreal interest rates might have very different effects on emerging market default risk. References Arora, V. and Cerisola, M. (2001). How Does U.S. Monetary Policy Influence Sovereign Spreads in Emerging Markets?, IMF Staff Papers 48: 474–498. Guimaraes, B. (2011). Sovereign default: which shocks matter?, Review of Economic Dynamics 14(4): 553–576. Levene, H. (1960). Robust tests for equality of variances, in I. O. et al. (ed.), In Contributions to Probability and Statistics: Essays in Honor of Harold Hotelling, Stanford University Press, pp. 278–292. Longstaff, F. A., Pan, J., Pedersen, L. H. and Singleton, K. J. (2011). How sovereign is sovereign credit risk?, American Economic Journal: Macroeconomics 3(2): 75–103. Neumeyer, P. and Perri, F. (2005). Business cycles in emerging economies: the role of interest rates, Journal of Monetary Economics 52(2): 345–380. Rigobon, R. (2003). Identification through heteroskedasticity, Review of Economics and Statistics 85(4): 777–792. Rigobon, R. and Sack, B. (2004). The impact of monetary policy on asset prices, Journal of Monetary Economics 51(8): 1553–1575. Robitaille, P. and Roush, J. (2006). How Do FOMC Actions and U.S. Macroeconomic Data Announcements Move BrazilianSovereign Yield Spreads and Stock Prices?, Board of Governors of the Federal Reserve System, International Finance Discussion Paper No. 868 . Rosa, C. (2011). The validity of the Event-Study Approach: Evidence from the Impact of the Fed’s Monetary Policy on U.S. and Foreign Asset Prices, Economica(311): 429–439. Uribe, M. and Yue, V. (2006). Country spreads and emerging countries: Who drives whom?, Journal of International Economics 69(1): 6–36. Zettelmeyer, J. (2004). The impact of monetary policy on exchange rates: evidence from three small open economies, Journal of Monetary Economics 51(3): 635–652. 10

A Appendix - alternative US interest rates In this appendix we prepare four alternative estimates of US real interest rates which are then used in place of the real rates reported in the main text as a robustness exercise. We obtaintwo nominal interest rateseries andtwo inflationmeasures fromtheGlobal Financial Database (www.globalfinancialdata.com). Both interest rate series are constant maturity, consistent with the data in the main text. We use a 3 month T-Bill rate consistent with existing quantitative studies in the literature, and a 10 year Treasury Bond rate consistent with the data in the main text because we maintain that long term rates are a more appropriate measure of the opportunity cost to investors in emerging market sovereign debt. The first measure of inflation is based on the Bureau of Labor Statistics monthly Consumer Price Index. We obtain the annual inflation rate in the year prior to each month, and average over the previous three months’ annual inflation rates to obtain a monthly estimate of future inflation. The second measure is the University of Michigan survey of annual CPI inflation expectations, which are also reported monthly. Both monthly series are assigned to the last working day of the month and subsequently cubic splined to obtain interpolated daily series of annual expected inflation. Each gross interest rate is divided by both gross expected inflation measures and netted. Figure 1 below shows the comparison of rates over time, and Table 6 shows the cross-correlations between the series. Tables 7 – 16 show the results from repeating the analysis described in the main text with the full sample, individual sub-samples (FOMC and non-FOMC meeting days) and applying the method of identification through heteroscedasticity for the four alternative measures of real interest rates. When using T-Bill rates, the standard errors are generally lower but the coefficients are much smaller. There are fewer significant coefficients and the magnitudes appear to be lower (no statistical tests of differences were run). Running the analysis separately on the sub-samples shows that the coefficients on the days when the FOMC meet are again insignificant, but those days when the FOMC do not meet appear to be of smaller magnitude although they remain significantly negative. When using T-Bond rates, the coefficients are generally of comparable magnitudes but the standard errors are much larger resulting in fewer significant positive coefficients. This is probably reflecting the fact that our measures of expected inflation are noisy when applied to daily data. All coefficients that are significant are positive. 11

Figure 1: US Real Interest Rates setaR tseretnI laeR SU 5 0 5− 1998 2000 2002 2004 2006 2008 Inflation Indexed 10−year Bond Yield 3m TBill real rate using BLS CPI 10Yr Bond real rate using BLS CPI 3m TBill real rate using UMICH CPI 10Yr Bond real rate using UMICH CPI Table 6: Correlation between real interest rate measures TIPS T-Bill & T-Bill & T-Bond & Yield BLS CPI UMICH CPI BLS CPI T-Bill & BLS CPI 0.753 . T-Bill & UMICH CPI 0.766 0.936 . T-Bond & BLS CPI 0.763 0.769 0.620 . T-Bond & UMICH CPI 0.846 0.710 0.745 0.862 12

A.1 T-Bill rates and BLS CPI inflation expectations Table 7: Full sample analysis (T-Bill & BLS CPI) Co-eff StdErr T-stat Emerging Market -0.301 0.060 -4.986 Latin America -0.332 0.067 -4.950 Brazil -0.325 0.100 -3.248 Bulgaria -0.147 0.058 -2.552 Mexico -0.211 0.045 -4.688 Panama -0.207 0.031 -6.678 Peru -0.237 0.044 -5.419 Venezuela -0.479 0.106 -4.539 Table 8: The response of EMBI+ premia to interest rate changes (T-Bill & BLS CPI) Co-eff StdErr T-stat Emerging Market 0.227 0.107 2.122 Latin America 0.165 0.115 1.443 Brazil 0.199 0.160 1.244 Bulgaria 0.211 0.105 2.014 Mexico 0.177 0.081 2.174 Panama 0.114 0.055 2.078 Peru 0.224 0.084 2.677 Venezuela 0.137 0.189 0.726 Table 9: Separate analysis of FOMC and non-FOMC meeting days (T-Bill & BLS CPI) Sub-sample C Sub-sample N Co-eff StdErr T-stat Co-eff StdErr T-stat Emerging Market -0.007 0.146 -0.045 -0.339 0.065 -5.247 Latin America -0.055 0.148 -0.369 -0.368 0.072 -5.105 Brazil -0.033 0.178 -0.184 -0.364 0.108 -3.358 Bulgaria 0.053 0.152 0.347 -0.173 0.062 -2.812 Mexico 0.005 0.115 0.047 -0.239 0.048 -4.970 Panama -0.028 0.074 -0.373 -0.230 0.033 -6.934 Peru 0.019 0.125 0.154 -0.271 0.047 -5.809 Venezuela -0.135 0.269 -0.502 -0.524 0.113 -4.639 13

A.2 T-Bill rates and Univ. of Michigan CPI inflation expectations Table 10: Full sample analysis (T-Bill & UMICH CPI exp.) Co-eff StdErr T-stat Emerging Market -0.241 0.059 -4.117 Latin America -0.248 0.065 -3.810 Brazil -0.301 0.097 -3.098 Bulgaria -0.146 0.056 -2.617 Mexico -0.203 0.044 -4.661 Panama -0.191 0.030 -6.343 Peru -0.238 0.043 -5.595 Venezuela -0.476 0.102 -4.651 Table 11: The response of EMBI+ premia to interest rate changes (T-Bill & UMICH CPI exp.) Co-eff StdErr T-stat Emerging Market 0.282 0.112 2.513 Latin America 0.186 0.120 1.553 Brazil 0.238 0.168 1.416 Bulgaria 0.280 0.110 2.541 Mexico 0.235 0.086 2.747 Panama 0.153 0.058 2.646 Peru 0.281 0.088 3.203 Venezuela 0.224 0.199 1.126 Table 12: Separate analysis of FOMC and non-FOMC meeting days (T-Bill & UMICH CPI exp.) Sub-sample C Sub-sample N Co-eff StdErr T-stat Co-eff StdErr T-stat Emerging Market 0.038 0.145 0.263 -0.276 0.063 -4.405 Latin America -0.016 0.147 -0.109 -0.277 0.070 -3.964 Brazil -0.013 0.177 -0.074 -0.336 0.105 -3.214 Bulgaria 0.081 0.151 0.539 -0.174 0.059 -2.929 Mexico 0.031 0.115 0.271 -0.232 0.046 -4.994 Panama -0.007 0.074 -0.094 -0.213 0.032 -6.642 Peru 0.040 0.124 0.318 -0.272 0.045 -6.033 Venezuela -0.102 0.268 -0.380 -0.522 0.109 -4.783 14

A.3 10Yr Bond rates and BLS CPI inflation expectations Table 13: Full sample analysis (T-Bond & BLS CPI) Co-eff StdErr T-stat Emerging Market -0.865 0.064 -13.484 Latin America -0.962 0.071 -13.497 Brazil -0.975 0.108 -8.997 Bulgaria -0.377 0.063 -6.014 Mexico -0.873 0.046 -18.825 Panama -0.577 0.032 -17.893 Peru -0.576 0.047 -12.267 Venezuela -1.027 0.114 -8.986 Table 14: The response of EMBI+ premia to interest rate changes (T-Bond & BLS CPI) Co-eff StdErr T-stat Emerging Market 1.316 0.509 2.586 Latin America 1.909 0.591 3.233 Brazil 2.944 0.825 3.570 Bulgaria 1.317 0.478 2.755 Mexico 0.686 0.373 1.840 Panama 0.679 0.268 2.533 Peru 0.684 0.372 1.837 Venezuela 2.682 0.905 2.963 Table 15: Separate analysis of FOMC and non-FOMC meeting days (T-Bond & BLS CPI) Sub-sample C Sub-sample N Co-eff StdErr T-stat Co-eff StdErr T-stat Emerging Market -0.378 0.210 -1.801 -0.899 0.067 -13.453 Latin America -0.322 0.213 -1.511 -1.007 0.074 -13.521 Brazil -0.101 0.259 -0.389 -1.036 0.114 -9.101 Bulgaria 0.000 0.221 0.002 -0.404 0.065 -6.194 Mexico -0.526 0.161 -3.259 -0.898 0.048 -18.615 Panama -0.297 0.105 -2.828 -0.597 0.034 -17.758 Peru -0.298 0.180 -1.657 -0.595 0.049 -12.251 Venezuela -0.200 0.392 -0.510 -1.085 0.119 -9.128 15

A.4 10Yr Bond rates and Univ. of Michigan CPI inflation expectations Table 16: Full sample analysis (T-Bond & UMICH CPI exp.) Co-eff StdErr T-stat Emerging Market -0.761 0.063 -12.149 Latin America -0.824 0.070 -11.813 Brazil -0.910 0.105 -8.643 Bulgaria -0.364 0.061 -5.976 Mexico -0.830 0.045 -18.373 Panama -0.537 0.031 -17.056 Peru -0.559 0.046 -12.244 Venezuela -0.994 0.111 -8.960 Table 17: The response of EMBI+ premia to interest rate changes (T-Bond & UMICH CPI exp.) Co-eff StdErr T-stat Emerging Market 2.144 0.807 2.656 Latin America 2.722 0.928 2.934 Brazil 4.267 1.332 3.203 Bulgaria 2.227 0.767 2.904 Mexico 1.334 0.599 2.227 Panama 1.180 0.441 2.677 Peru 1.294 0.579 2.237 Venezuela 4.266 1.458 2.925 Table 18: Separate analysis of FOMC and non-FOMC meeting days (T-Bond & UMICH CPI exp.) Sub-sample C Sub-sample N Co-eff StdErr T-stat Co-eff StdErr T-stat Emerging Market -0.282 0.213 -1.327 -0.793 0.065 -12.166 Latin America -0.240 0.216 -1.114 -0.862 0.073 -11.852 Brazil -0.056 0.261 -0.215 -0.966 0.110 -8.747 Bulgaria 0.063 0.222 0.284 -0.392 0.063 -6.206 Mexico -0.473 0.164 -2.890 -0.853 0.047 -18.217 Panama -0.254 0.107 -2.380 -0.555 0.033 -16.975 Peru -0.254 0.181 -1.399 -0.579 0.047 -12.279 Venezuela -0.127 0.395 -0.321 -1.051 0.115 -9.127 16

B Appendix - estimation in levels The analysis presented in this appendix is intended to justify time-differencing the data in the paper. We show that (i) there is no significant increase in the variance of the levels of the US real interest rate on the dates the FOMC meets, which is inconsistent with the fundamental assumption underpinning the methodologyof identification through heteroskedasticity; and (ii) the data we use are highly persistent over time, and as a result the usual tests cannot reject a unit root. An analysis in levels would be subject to the critique that any results were spurious. The fundamental assumption underpinning the methodology of identification is not directly testable because we cannot identify the shocks. But the best available evidence we have suggests that it is appropriate to apply the methodology in differences, but not in levels. Table 19 shows the descriptive statistics for our variables in levels using data defined to capture the level of each variable on the day after the FOMC meeting dates. The analysis is repeated in Table 20 using the level of variables on the same day as the FOMC meeting. In both cases, and similar to Table 1, there is no significant difference in the standard deviation of EMBI+ variables on the days when the FOMC meets from the days when it does not. In Table 19 there is a (weakly) significant reduction in the standard deviation of the US real interest rate on the days when the FOMC meets, and in Table 20 there is no significant change. This is not consistent with the assumption that the variance of the interest rate would significantly increase on FOMC meeting days. Table 19: Data descriptives (levels) Standard Covariance with Levene (1960) test deviation US real rate of equal variance FOMC No FOMC FOMC No FOMC mean test p-value US real rate 0.885 0.898 . . 2.717 0.066 Emerging Market 314.595 319.333 194.756 202.617 0.103 0.749 Latin America 296.342 295.865 124.524 127.392 0.037 0.847 Brazil 421.413 418.673 153.673 156.705 0.078 0.780 Bulgaria 302.345 313.245 223.860 238.411 0.449 0.503 Mexico 180.834 184.149 114.845 120.151 0.183 0.668 Panama 119.021 120.635 61.070 63.281 0.009 0.924 Peru 217.518 213.622 124.616 123.782 0.032 0.858 Venezuela 381.318 386.208 130.639 152.410 0.008 0.927 Notes: Levene(1960) test statistic based on mean; nullhypothesisis equalvariance FOMC means theset of daysimmediately after FOMC meetings Table 21 shows the results from tests of stationarity on the variables in levels and 17

Table 20: Data descriptives (levels) Standard Covariance with Levene (1960) test deviation US real rate of equal variance FOMC No FOMC FOMC No FOMC mean test p-value US Real Rate 0.883 0.898 . . 0.668 0.414 Emerging Market 314.952 319.314 194.621 202.629 0.070 0.792 Latin America 298.070 295.774 123.374 127.455 0.054 0.816 Brazil 426.906 418.392 151.228 156.836 0.091 0.764 Bulgaria 309.204 312.911 227.779 238.222 0.096 0.757 Mexico 182.428 184.070 115.595 120.116 0.147 0.702 Panama 120.181 120.577 61.350 63.272 0.018 0.893 Peru 216.637 213.664 123.413 123.843 0.018 0.894 Venezuela 379.584 386.308 130.098 152.424 0.003 0.953 Notes: Levene(1960) test statistic based on mean; nullhypothesisis equalvariance FOMC means theset of dayson which FOMC meetings are held first differences. Both tests include a constant but no trend term; the Phillips-Perron specification includes seven Newey-West lags. The variables in levels are all non-stationary. Identical specifications for the differenced time-series employed in the paper show they are stationary. We conclude that it is more appropriate to specify the model in terms of differences than in levels. Table 21: Stationarity test statistics Levels First Differences Phillips-Perron Dickey-Fuller Phillips-Perron Dickey-Fuller US real rate -1.32 -1.28 -24.36 -25.14 Emerging Market -1.11 -1.03 -22.70 -23.83 Latin America -1.49 -1.46 -23.32 -24.66 Brazil -2.84 -2.80 -22.36 -23.47 Bulgaria -1.60 -1.59 -25.08 -25.26 Mexico -1.51 -1.50 -23.05 -24.31 Panama -0.88 -0.59 -24.74 -25.27 Peru -1.95 -1.92 -25.21 -25.58 Venezuela -0.44 -0.29 -25.50 -26.43 Notes: Nullhypothesisis stationarity in all unit root tests Phillips-Perron specifications use seven Newey-Westlags Critical valuesare -3.43 (1%);-2.86 (5%);-2.57 (10%) 18

C Appendix - dynamic model This appendix reports the results from a dynamic specification of the model, as an investigation of dynamic effects, for example overshooting, in the reaction of the EMBI+ spread to changes in US real interest rates9. We maintain the definition of the variables as in the main text, i.e. ∆X X X , but re-specify the model as follows: t t+1 t−1 ≡ − Table 22: ∆E = α ∆I +α ∆I +α ∆E t 1 t 2 t−2 3 t−2 instrumented Table 23: ∆e t = α 1 ∆|i{ t z+}α 2 ∆i t−2 +α 3 ∆e t−2 The Tables below should be compared with Tables 3 and 5 in the main text. Following the notationin the maintext, the instruments employed in the 2SLSestimates of dynamic model in Table 22 are w, ∆I , and ∆E . t−2 t−2 We find that in general the coefficients on the lags in both specifications were statistically insignificant and conclude that there is no systematic evidence of dynamic effects present in the data. Table 22: Identification via heteroscedasticity dynamic analysis Co-efficients Standard Error T-statistic US RR L.US RR L.DV US RR L.US RR L.DV US RR L.US RR L.DV E. Market 0.96 0.00 7.34 0.17 0.00 6.50 5.57 2.56 1.13 L Am. 1.23 0.00 4.26 0.19 0.00 7.08 6.52 0.37 0.60 Brazil 1.48 0.00 2.49 0.26 0.00 9.81 5.69 0.17 0.25 Bulgaria 0.69 0.00 8.56 0.16 0.00 6.21 4.17 0.48 1.38 Mexico 0.68 0.00 6.18 0.13 0.00 4.99 5.15 1.55 1.24 Panama 0.54 -0.00 2.57 0.09 0.00 3.46 5.91 -0.01 0.74 Peru 0.73 0.00 5.22 0.13 0.00 5.05 5.43 1.66 1.04 Venezuela 2.28 0.00 -1.24 0.31 0.00 11.76 7.33 4.82 -0.11 Notes: DV – dependentvariable– EMBI+ premium. USRR – US real interest rate. Each estimation uses 2,611 observations. 9Wegratefully acknowledge thisfollows thesuggestion of an anonymousreferee. 19

Table 23: Full sample dynamic analysis Co-efficients Standard Error T-statistic US RR L.US RR L.DV US RR L.US RR L.DV US RR L.US RR L.DV E. Market -0.39 -0.02 0.04 0.08 0.08 0.02 -4.64 -0.25 2.00 L. Am. -0.47 -0.06 0.00 0.09 0.09 0.02 -5.00 -0.68 0.06 Brazil -0.49 -0.16 0.00 0.14 0.14 0.02 -3.53 -1.11 0.13 Bulgaria -0.20 0.17 -0.03 0.08 0.08 0.02 -2.49 2.12 -1.65 Mexico -0.41 -0.03 0.00 0.06 0.06 0.02 -6.63 -0.47 0.19 Panama -0.42 -0.06 0.01 0.04 0.04 0.02 -9.91 -1.39 0.75 Peru -0.36 0.03 0.07 0.06 0.06 0.02 -6.02 0.51 3.67 Venezuela -0.47 -0.13 0.03 0.15 0.15 0.02 -3.18 -0.88 1.27 Notes: DV – dependentvariable– EMBI+ premium. USRR – US real interest rate. Each estimation uses 2,611 observations. D Appendix - tests of variance The increase in the variation in the US real interest rate and the change in covariance between the real interest rate and EMBI+ premia over the sub-samples are apparent from Table 1 in the main text, but the fact that the standard deviations of EMBI+ premia appear to decrease from sub-sample N to sub-sample C, when we expect mild increases, suggests we require a more accurate statistical test of whether our assumptions on the variance of shocks over the two sub-samples are valid. Importantly, however, we cannot apply standard tests of variance equality, because they require that the underlying data be normally distributed. As the plots of each variables’ quantiles against those of the normal distribution in Figure 2 demonstrate, and the empirical tests of skewness and kurtosis confirm in Table 24, none of our series are normally distributed. Levene (1960)provides a test where thenull isequal variancewhen samples aredrawn from a distribution that is not Gaussian normal. The results from this test are presented in Table 25, and show that the variance of the US real interest rate significantly increases, but the variance of all EMBI+ premia does not change significantly.10 On the basis of these results, we conclude that the standard deviation of the real interest rate increases significantly on the days when the variance of interest rate movements is greater. We cannot reject the null that the variance of EMBI+ is the same in both sub-samples. According to our assumptions, the policy shocks should yield only 10Theresultsarepresentedusingthesamplemeanofthedata,similarresultsareobtainedwhenusingthe50th percentile or 10% trimmed mean. 20

Figure 2: Q-Q plots of each variable quantiles against normal distribution quantiles tekraM gnigremE 004 002 0 002− 004− −100 −50 0 50 100 Inverse Normal aciremA nitaL 004 002 0 002− 004− −100 −50 0 50 100 Inverse Normal lizarB 005 0 005− −200 −100 0 100 200 Inverse Normal airagluB 006 004 002 0 002−004− −100 −50 0 50 100 Inverse Normal ocixeM 003 002 001 0 001−002− −100 −50 0 50 100 Inverse Normal amanaP 002 001 0 001− −50 0 50 Inverse Normal ureP 051 001 05 0 05−001− −100 −50 0 50 100 Inverse Normal aleuzeneV 0001 005 0 005− −200 −100 0 100 200 Inverse Normal 21

Table 24: Test of skewness and kurtosis skewness kurtosis p-value p-value US real rate 0.000 0.000 Emerging Market 0.000 0.000 Latin America 0.000 0.000 Brazil 0.000 0.000 Bulgaria 0.000 0.000 Mexico 0.000 0.000 Panama 0.000 0.000 Peru 0.000 0.000 Venezuela 0.000 0.000 Note: Nullhypothesisis normal distribution Table 25: Levene (1960) test of equal variance Test statistic p-value based on mean US real rate 12.371 0.000 Emerging Market 0.215 0.643 Latin America 0.458 0.499 Brazil 2.273 0.132 Bulgaria 0.000 0.977 Mexico 0.031 0.860 Panama 0.021 0.884 Peru 0.908 0.341 Venezuela 0.635 0.801 Note: Null hypothesisis equalvariance small increases in the variance of EMBI+, as the unexpected policy shocks to US real interest rates are only a small part of the variation of emerging market default risk, so the results of the tests on variances in both sub-samples, albeit not conclusive, are not at odds with the identifying assumptions. 22