International Bank Lending Channel of Monetary Policy

How does domestic monetary policy in systemic countries spillover to the rest of the world? This paper examines the transmission channel of domestic monetary policy in the cross-border context. We use exogenous shocks to monetary policy in systemically important economies, including the U.S., and local projections to estimate the dynamic effect of monetary policy shocks on bilateral cross-border bank lending. We find robust evidence that an increase in funding costs following an exogenous monetary tightening leads to a statistically and economically significant decline in cross-border bank lending. The effect is weakened during periods of high uncertainty. In contrast, the effect is found to not vary according to the degree of borrower country riskiness, further weakening support for the international portfolio rebalancing channel.


I. INTRODUCTION
The reason for focusing on cross-border banking flows is threefold. First, while most previous studies focused on net capital flows, the rapid expansion of gross international asset and liability positions calls for a deeper understanding of the spillovers through gross flows that better reflect the impact on national balance sheets (Milesi-Ferretti and Tille, 2011;Broner et al., 2013). Second, to the extent that cross-border banking flows have meaningful implications for economic and financial conditions in recipient countries, as suggested by the recent empirical studies (Popov and Udell, 2012;Schnabl, 2012;Bruno and Shin, 2015a;Bräuning and Ivashina, 2019;Morais et al., 2019), examining the effect of monetary policy shocks on these flows helps identify the transmission channel of monetary policy spillovers. Third, the bilateral nature of cross-border banking flow data permits a cleaner identification of the international transmission channel of monetary policy since it allows controlling for credit demand factors in a recipient country (Cetorelli and Goldberg, 2011;Correa et al., 2017;Avdjiev et al., 2018).

©International Monetary Fund. Not for Redistribution
While the monetary policy in systemically-important economies, such as the U.S., has proven to be a robust driver of international capital flows and risky asset prices across the globe , there is no consensus on the sign of the effect of monetary policy actions on cross-border banking flows.
In principle, the credit channel of the monetary policy suggested by Bernanke and Gertler (1995) amplifies the effect of monetary policy shocks through frictions in the liability-side (a bank lending channel) or the asset-side (a portfolio rebalancing channel) of financial intermediaries or both. In either case, monetary policy tightening would result in a decline in bank lending. Additionally, Borio and Zhu (2012) and Coimbra and Rey (2017) show that monetary policy influences agents' risk-taking behavior, thereby increasing the credit supply during periods of easing. Bruno and Shin (2015b) model this risktaking channel in an international context, specifically looking at the banking sector, and show that U.S. expansionary monetary policy increases cross-border bank capital flows through higher leverage of international banks.
However, monetary policy actions do not always have the same effect on bank lending in the domestic and international context. For example, according to the portfolio rebalancing channel (Den Haan et al., 2007;Dell'Ariccia et al., 2017), domestic monetary policy tightening may increase crossborder bank lending by eroding the net worth and collateral value of domestic borrowers and thus leading to a reallocation of lending toward relatively safer borrowers abroad (Correa et al., 2017;Argimon et al., 2019).
The existing empirical evidence on the effect of monetary policy on cross-border bank lending is mixed. Using data from the Bank for International Settlements (BIS)' Locational Banking Statistics (LBS) for the period 1995-2007, Bruno and Shin (2015a find that higher U.S. interest rates have a negative impact on cross-border bank lending, which is consistent with an international bank lending channel of monetary policy. Temesvary et al. (2018) and Bräuning and Ivashina (2019) corroborate this finding using banks-firms matched loan-level data. Miranda-Agrippino and Rey (2019) find evidence for the so-called global risk aversion channel as a tightening in U.S. monetary policy triggers a surge of aggregate risk aversion and consequent retrenchments of international credit flows, particularly in the banking sector.
In contrast, Cerutti at al. (2017), also using the data from the BIS LBS, find that higher U.S.
short-term interest rates are associated with an increase in cross-border bank lending. Correa et al. (2017) and Avdjiev et al. (2018) extend this finding to a large sample of lender countries, providing supporting evidence for the international portfolio rebalancing channel. Using Dutch, Spanish, and U.S. confidential supervisory data, Argimon et al. (2019) also find that global banks increase cross-border ©International Monetary Fund. Not for Redistribution lending in response to domestic monetary tightening. Avdjiev and Hale (2019) find mixed evidence about the effect of monetary policy on international bank lending depending on the prevailing international capital flow regimes (high vs. low bank lending growth) and on the drivers of the monetary policy rate (macroeconomic fundamentals vs. monetary policy stance). 2 We show that this lack of empirical consensus is mostly due to the use of short-term policy rates in testing the contrasting theoretical channels and argue that our empirical framework can reconcile the mixed evidence in the literature. Because monetary policy is typically guided by a rule, the largest part of the variation in monetary policy actions is due to the systematic component of monetary policy-that is, the response of the central bank to the current and expected future state of the economy. As discussed by Ramey (2016), identifying the causal effect of monetary policy requires looking at the exogenous deviations from the monetary rule. Earlier studies often identified exogenous shocks to monetary policy when investigating the credit channel of monetary policy. Surprisingly, however, most of the existing studies on cross-border bank lending have not properly addressed this issue. 3 Rather, they have typically examined the effect of an increase in the policy rate, which is confounded by the endogenous response of monetary policy to underlying economic conditions.
To address the endogeneity concern, we employ exogenous monetary policy shocks in the U.S.-the shocks identified by a narrative approach of  and those identified by external instruments using high-frequency data of Gürkaynak et al. (2005) and Gertler and Karadi (2015) 4 -and, in other eight advanced economies, the exogenous shocks series constructed by Furceri et al. (2018) 5 and apply the local projection method (Jordà, 2005) in estimating the dynamic effect of monetary policy on cross-border bank lending. We pay particular attention to the effect of U.S. monetary policy shocks given the dominance of U.S. monetary policy in shaping international capital flows and the special role of U.S. dollars in the international financial system. Thus, our analysis 2 Avdjiev and Hale (2019) decompose changes in the Federal funds rate into what is predicted by the Taylor rule (the "macro fundamentals component") and the difference between the Federal funds rate and the one implied by the Taylor rule (the "monetary policy stance component"). However, they find that an increase in both components has positive effects on crossborder bank lending from U.S. banks, especially when lending to advanced economies, which is in sharp contrast to our findings.
complements the voluminous literature on the real, monetary, and financial spillover effects of U.S. monetary policy (Kim, 2001;Canova, 2005;Bruno and Shin, 2015b;Bräuning and Ivashina, 2019;. Given the ample empirical evidence on the nonlinear effect of monetary policy shocks on economic and financial activity (Cover, 1992;Tenreyro and Thwaites, 2016;Castelnuovo and Pellegrino, 2018), we further investigate whether the effect of monetary policy shocks on cross-border banking flows depends on the underlying state of business cycles (expansions vs. recessions) in the source economy, global financial risks or uncertainty (low vs. high), and the sign of the shocks (tightening vs. easing). In addition, we analyze whether the riskiness of a recipient country strengthens or weakens the international bank lending channel of monetary policy, which bears significant policy implications.
The key results of the paper are the following: • Exogenous monetary policy tightening in systemically important source economies leads to an economically and statistically significant decline in cross-border bank lending. This holds for the U.S. as well as the other advanced economies analyzed. These results sharply contrast with the evidence presented in previous studies using similar data but relying on the level of policy rates as a measure of monetary policy actions (Correa et al., 2017;Avdjiev et al., 2018;Argimon et al., 2019).
• U.S. monetary policy shocks have a statistically and economically significant effect on crossborder bank lending even when controlling for global financial risks or uncertainty (proxied by the VIX) or liquidity risks (proxied by the LIBOR-OIS spread), implying that U.S. monetary policy is an independent source of the so-called "global financial cycle." 6 • The effect tends to be larger during periods of lower global uncertainty (proxied by the VIX index of implied volatility on the U.S. equity options), consistent with the monetary policy ineffectiveness during the period of high uncertainty (Aastveit et al., 2017;Castelnuovo and Pellegrino, 2018).
• We find that the effect of monetary policy tightening on cross-border bank lending does not depend on the riskiness of borrower countries (proxied by political risk index or income status), against the predictions of the international portfolio rebalancing channel of monetary policy.
The remainder of the paper is organized as follows. Section II describes the data on crossborder banking and exogenous measures of monetary policy shocks. Section III illustrates the empirical methodology and provides a thorough analysis of the international bank lending channel of monetary policy, including various robustness tests and additional exercises. Section IV concludes.

A. Cross-border banking flows
We use data on cross-border claims from the BIS' LBS to test the international bank lending channel of monetary policy in systemically important countries. This dataset provides a geographical breakdown of reporting banks' counterparties and the information about the currency composition of their balance sheets. The LBS dataset captures outstanding claims and liabilities of internationally active banks located in reporting countries against their counterparties residing in more than 200 countries. The data is compiled following the residency principle that is consistent with the BoP (Balance of Payments) statistics.
As the conventional bank lending channel relates to conditions in the location where banks make funding and lending activities, the residency principle has a conceptual appeal over the nationality principle used in Consolidated Banking Statistics (CBS). Moreover, to the extent to which foreign affiliates are subject to host-country regulation or have access to local bank liquidity facilities (Avdjiev et al., 2018), the residency principle is more appropriate than the nationality principle in identifying the international bank lending channel of monetary policy.
The major advantage of the BIS LBS data, compared to the banking flows collected from the BoP statistics, is the detailed breakdown of the reported series by recipient countries. 7 Banks record their positions on an unconsolidated basis, including intragroup positions between offices of the same banking group. Currently, banking offices located in 46 countries, including many offshore financial centers, report the LBS. The LBS dataset captures around 95 percent of all cross-border interbank business (Bank for International Settlement, 2017). The bulk of cross-border bank claims and liabilities takes a form of loans and securities of the domestic banking sector vis-à-vis all counterparty sectors (including banks and non-banks, and the private and public sector).
Another advantage of the BIS LBS dataset is that the currency composition of cross-border claims and liabilities is available so that cross-border banking flows expressed in U.S. dollars are adjusted for movements in exchange rates. The adjustment for exchange rate movements turns out to be crucial in our setup since fluctuations in the exchange rate, which are influenced by monetary policy shocks, also affect cross-border bank lending. The availability of a currency breakdown enables the BIS to calculate break-and exchange rate-adjusted changes in amounts outstanding. Such adjusted changes approximate underlying flows during each quarter. 8 The adjusted change is calculated by first converting U.S. dollar-equivalent amounts outstanding into their original currency using the end-of-period exchange rates, then calculating the difference in amounts outstanding in the original currency, and finally converting the difference into a U.S. dollar-equivalent change using the average period exchange rates. 9 As the BIS LBS only report the exchange rate-adjusted flows, we construct the exchange rate-adjusted stock of the cross-border claims from country i to country j as the cumulated sum of the exchange rate-adjusted flows, where the initial value of the exchange rate-adjusted stock is set equal to the exchange rate-unadjusted claimsdirectly available from the BIS LBS. 10 The time-series coverage of LBS database varies significantly across countries. Some advanced economies such as the U.S. have reported these statistics since 1977, while some emerging market economies started reporting the statistics only after the 2000s. We first analyze the international bank lending channel of U.S. monetary policy using data from 1990Q1 to 2012Q4 due to its extensive availability of cross-border banking flow data and well-identified exogenous monetary policy shock as 8 Adjusted changes in amounts outstanding are calculated as an approximation for flows. In addition to exchange rate fluctuations, the quarterly flows in the locational datasets are corrected for breaks in the reporting population. 9 Nevertheless, the adjustment practice by the BIS cannot eliminate the possibility of under-or overestimation of actual flows. Adjusted changes could still be affected by changes in valuations, writedowns, the underreporting of breaks, and differences between the exchange rate on the transaction date and the quarterly average exchange rate used for conversion. See Avdjiev and Hale (2019) for further details. 10 Figure A.1 in the appendix shows the exchange rate-unadjusted and adjusted U.S. cross-border claims. While the two series co-move very closely (the correlation is 0.75) over the entire sample, accounting for the valuation effect results in a more pronounced decline in cross-border bank lending from the U.S. during the global financial crisis (GFC), suggesting that the appreciation of U.S. dollars during this period partially offsets a larger decline in "real" cross-border bank lending originally denominated in U.S. dollars. The correlation between the exchange rate-adjusted and unadjusted series in the other eight advanced economies is about 0.6. well as its dominance in shaping international capital flows and the special role of U.S. dollars in the international financial system. 11 Throughout the analysis, we drop offshore financial centers using the IMF classification from our sample because their behaviors might differ substantially from the rest of the sample. Following Correa et al. (2017) and Choi and Furceri (2019), we further drop observations with the size of crossborder positions less than $5 million, or with negative total outstanding claims. Observations of the dependent variable in the upper and lower one percentile of the distribution are excluded from the sample to reduce the influence of outliers. 12 Table A.1 in the appendix lists the final sample of countries used in the analysis, together with their income status, an indicator whether they belong to the euro area, and their average status regarding the exchange rate regime, monetary policy independence, and capital account openness during the sample period using the trilemma index constructed by . 13 The details on this index will be discussed in the online appendix B. We use these country-specific characteristics to investigate factors affecting the international bank lending channel of monetary policy.
To provide a first look at the pattern of cross-border bank lending, we present the size of total cross-border claims and liabilities as a share of the GDP in 2010Q4 for the 9 reporting countries in Table A.2 in the appendix. When normalized to the size of domestic GDP, the predominant role of European countries in the cross-border banking system is apparent. Cross-border claims of global banks in advanced economies are, on average, larger than liabilities, suggesting that they are net lenders in this market.
We further illustrate the bilateral structure of the data by presenting examples of bilateral crossborder claims between the U.S. and six countries in Figure A.2 in the appendix. Given the confidentiality of the data, we do not reveal the identity of the individual counterparties, but the first four recipient countries are advanced economies and the last two recipient countries are emerging market economies. Some observations stand out from the figure. First, the different scales of the y-axis 11 Data before 1990 are sparse and with gaps. 12 The qualitative results are robust to the inclusion of the observations less than $5 million. They are also robust to (i) dropping the dependent variables at the top/bottom 2.5 percentile, (ii) winsorizing the dependent variables at the 1 or 2.5 percentile of the distribution, and (iii) including all the observations. in these graphs re-emphasize the dominance of advanced economies in accounting for these flows.
Second, the pattern of cross-border lending is quite heterogeneous across countries.

B. Identification of monetary policy shocks
As discussed earlier, the proper identification of the causal effect of monetary policy actions on cross-border bank lending requires using monetary policy actions that are orthogonal to current and expected future macroeconomic conditions (Ramey, 2016). This is the main novelty of the paper compared to many existing studies on the spillover effects of monetary policy through cross-border banking flows.

Measures of exogenous U.S. monetary policy shocks.
In the baseline analysis, we use the exogenous monetary policy shock series constructed by Coibion (2012) who extends the monetary policy shocks identified by  using a narrative approach.
where m denotes the FOMC meeting, is the target federal funds rate going into the FOMC meeting, , is the Greenbook forecast from meeting m of real output growth in quarters around meeting m, , is the Greenbook forecast of GDP deflator inflation, and 0 is the Greenbook forecast of the current quarter's average unemployment rate.
The estimated residuals m are then defined as exogenous monetary policy shocks, purged of anticipatory effects related to economic conditions. Our quarterly measure of monetary policy shocks comes from summing the orthogonalized innovations to the Federal funds rate from each meeting within a quarter. As a robustness check, we also use the identification strategy of Gertler and Karadi (2015) based on high-frequency data used as external instruments (see next section).

Measures of exogenous monetary policy shocks in other advanced economies.
We follow the methodology of Furceri et al. (2018) to construct quarterly measures of exogenous monetary policy ©International Monetary Fund. Not for Redistribution shocks for other eight advanced economies. They identify monetary policy shocks in two steps, which closely follow the work by Auerbach and Gorodnichenko (2013) in identifying fiscal shocks in advanced economies. First, they compute the unexpected changes in policy rates (proxied by short-term rates) using the forecast errors of the policy rates provided by Consensus Economics. 14 Second, they regress for each country the forecast errors of the policy rates on similarly-computed forecast errors of inflation and output growth and identify the shocks as the residuals of this regression. We follow this approach, but we further purify these surprises of any predictable components by projecting it on current and lagged GDP growth and inflation, in order to eliminate any remaining endogeneity issue.
Specifically, we estimate where , ( = , ∆ , ) is the unexpected changes in policy rates (proxied by short-term rates), real GDP growth, and the inflation rate, respectively-defined as the difference between the actual value at the end of the quarter and the value expected by analysts as of the beginning of the last quarter for each country i. ∆ , and , are corresponding actual real GDP growth and the inflation rate. The estimated residuals î,t are then defined as exogenous monetary policy shocks for advanced economies other than the U.S. 15 As discussed by Furceri et al. (2018), this methodology has two main advantages. First, it overcomes the issue of "policy foresight" (Leeper et al., 2013) where economic agents receive news about possible changes in monetary policy and alter their behavior before the actual changes in policy happen. Changes in actual policy rate cannot capture this policy foresight of economic agents, leading to inconsistent estimates of the effect of monetary policy shocks on cross-border bank lending. Our approach, on the contrary, is free from this issue since it uses forecast errors which already reflect policy foresight by its construction. Second, this methodology reduces endogeneity issues as the shocks are orthogonal to unexpected changes in economic activity as well as to current and lagged endogenous variables.
14 The Consensus Economics publications report forecasts for short-term (3-months) rates at the end of the next three months.  Table A.3 in the appendix shows the standard deviation of the country-specific exogenous monetary policy shock series, together with its correlation with the U.S. monetary policy shocks constructed by Furceri et al. (2018) and Coibion (2012), respectively. The correlation between the U.S. monetary policy shock series from Furceri et al. (2018) and that from Coibion (2012) is 0.62 for the overlapped sample. The correlation between the country-specific monetary policy shock series with the U.S. is typically small except for Canada, ensuring that the identification of international bank lending channel of monetary policy from other advanced economies is unlikely confounded by the effect of U.S. monetary policy on the short-term rates in these economies. 16

A. Local projection method
We use Jordà (2005)'s local projection method to estimate the dynamic effect of monetary policy shocks on cross-border bank lending. The local projection method has been advocated by Auerbach and Gorodnichencko (2012) and Ramey and Zubairy (2018), among others, as a flexible alternative to VAR specifications without imposing the pattern generated by structural VARs. In the bilateral panel data setting, we adopt the local projection method over commonly used VAR models for the following specific reasons. 17 First, the exogenous shocks we use are already orthogonalized to contemporaneous and expected future macroeconomic conditions. For this reason, we do not need to further identify monetary policy shocks using restrictions in VAR models-a common approach in many empirical analyses in both domestic and international setups.
Second, our estimation entails a large international panel dataset with a constellation of the fixed effects, which makes a direct application of standard VAR models more difficult. In addition, the local projection method obviates the need to estimate the equations for dependent variables other than the variable of interest, thereby significantly economizing on the number of estimated parameters. 16 The high correlation between the monetary policy shocks in the U.S. and Canada reflects the close economic ties between the two countries and is consistent with the recent finding that Canada is the only country among a group of advanced economies experiencing a near-complete passthrough of conventional U.S. monetary policy for short-term interest rate.
Third, the local projection method is particularly suited to estimating nonlinearities (for example, how the effect of monetary policy shocks differs during expansions and recessions in the source economy), as its application is much more straightforward compared to non-linear structural VAR models, such as Markov-switching or threshold-VAR models. Moreover, it allows for incorporating various time-varying features of source (recipient) economies directly and allow for their endogenous response to monetary policy shocks.
Lastly, the error term in the following panel estimations is likely to be correlated across countries. This correlation would be difficult to address in the context of VAR models, but it is easy to handle in the local projection method by either clustering standard errors by time period or using the Driscoll-Kraay standard errors allowing for arbitrary correlations of the errors across countries and time (Driscoll and Kraay, 1998).
Despite the advantages mentioned above, the local projection method has some drawbacks compared to structural VARs. First, since the iterated VAR method produces more efficient parameter estimates than the local projection method, the impulse response function estimated by local projections is often associated with large confidence intervals. This problem of less precise estimates is exacerbated as a forecast horizon increases due to the decreasing sample size in each estimation. Thus, we report both 68% and 90% confidence intervals in the following analyses.
Second, compared to a single equation framework in the local projection method, structural VARs allow tracing the dynamic endogenous response of various macroeconomic variables in the system to monetary policy shocks, which in turn can also affect the dynamics of cross-border bank lending. We enhance the credibility of the identified shock by analyzing separately the effect of U.S.
monetary policy shocks on domestic economic variables including domestic bank lending. Figure B.1 in the online appendix confirms that output and investment decline, while the nominal exchange rate appreciates in response to the exogenous monetary policy tightening. Although we find a weak price puzzle, the effect on CPI is not statistically significant. Most importantly, we also find that domestic bank lending decreases following an increase in bank funding costs, consistent with the bank lending channel of monetary policy in the domestic context.

B. International bank lending channel of U.S. monetary policy shocks
The local projection method simply requires the estimation of a series of regressions for each horizon h for each variable. Following Auerbach and Gorodnichencko (2012) and Ramey and Zubairy (2018), we run a series of regressions for different horizons, ℎ =, 1, 2, … as follows: where , is the log of exchange rate-adjusted cross-border bank claims from the U.S. to borrowers in countries j in time t; ℎ is a recipient country-fixed effect, which controls unobserved time-invariant characteristics specific to a country j; ℎ is the measure of exogenous U.S. monetary policy shocks; , is a set of control variables including lags of the dependent variable and the monetary policy shocks, and various control variables in the recipient country j (for example, real GDP growth, the short-term interest rate, inflation, and the nominal exchange rate growth vis-à-vis the U.S.) and their lags.
While the impulse responses generated by the local projections are not an estimate of the total effects of U.S. monetary policy-due to a constellation of fixed effects-, the exogeneity of the monetary policy shock and controlling for the demand-side factors allow us to investigate the spillover channel of monetary policy. In the baseline analysis, we use four lags of control variables in , (i.e., = 4), however, the selection of the lag length does not affect our findings. 18 We estimate equation (3) using OLS, which would result in the inconsistency of the leastsquares parameter estimates due to the combination of lagged dependent variables and fixed effects (Nickell, 1981). However, because the time-series dimension of the panel dataset is quite large, the inconsistency is unlikely a major concern. Following Auerbach and Gorodnichencko (2012), standard errors are clustered by time to account for the fact that the shock is identical to all recipient countries in any given period. Equation (3) is estimated for h=0, 1, 2,…, 7 so that we trace the dynamic effect of monetary policy shocks over two years. After dropping outliers and missing observations following the criterion explained above, our baseline estimation of the U.S. monetary policy shocks covers crossborder lending to 45 recipient countries.
Baseline results. Figure 1 presents the dynamic response of cross-border bank lending to exogenous U.S. monetary policy shocks. The results provide evidence of a significant negative effect which is consistent with the cross-border bank lending and the risk-taking channels of monetary policy. In particular, a 100 basis-point (bp) exogenous tightening is found to lead to more than a 10 percent decline in cross-border bank lending after two quarters, which is not only statistically but also economically significant. 19 Table 1 summarizes the full estimation results using exogenous U.S. monetary policy shocks above. 20 The coefficients on the lagged dependent variable are negative and highly statistically significant, suggesting that the growth rate of the cross-border bank lending is mean-reverting. The coefficients on a recipient country's real GDP growth are positive, although they are not statistically significant in most cases. 21 The coefficients on the recipient country's short-term interest rate are not statistically significant. While this finding is in contrast to Bruno and Shin (2015a), who find that a higher interest rate in a recipient country increases cross-border bank lending toward this country, it is mostly driven by the emerging market recipient economies in the sample where the interest rates have been typically countercyclical. Indeed, when we restrict the set of recipient countries to advanced economies, we find a positive and statistically significant coefficient on the recipient country's short-term interest rate, consistent with the finding that interest rate differentials are a strong pull factor of cross-border banking flows.
The results on the exchange rate suggest that a depreciation of the local currency is associated with a decline in cross-border bank lending toward the recipient country, which is consistent with the 20 While we control for the four lags of the variables, we only report the estimation results up to one lag to save space here. The results are available upon request. 21 Once we drop the exchange rate growth in the estimation, the coefficients on a recipient country's real GDP growth become statistically significant without any material change in the coefficients of the monetary policy shock. This finding is consistent with the stylized fact that local economic growth is a pull factor of international capital flows. The results are available upon request. bank lending on the lagged level of the Federal funds rate, which follows closely the static specification in Correa et al. (2017). 22 Column (I) in Table 2 summarizes the estimation results. Consistent with many existing studies, an increase in the U.S. monetary policy rate is associated with an increase in crossborder bank lending. Column (II) provides the estimation results using the changes in the Federal funds rate instead of its lagged level, which deliver similar results although the estimated coefficient of interest is not statistically significant. Interestingly, the sign of the estimated coefficient of interest switches its sign when employing the exogenous U.S. monetary policy shocks (column III). 23 Second, we examine the dynamic effect by re-estimating equation (3)  better captured by a dynamic framework than a static one mixing short-and medium-term effects.

High-frequency identification with external instruments.
We verify the validity of our baseline results by further exploiting an alternative identification strategy based on high-frequency data used as external instruments (Gertler and Karadi, 2015;Cloyne et al., 2018, Miranda-Agrippino and. This hybrid approach, proposed by Gertler and Karadi (2015), combines the high-frequency identification widely used in the finance literature (Kuttner, 2001;Gürkaynak et al., 2005) with external instrument methods (Mertens and Ravn, 2013). The main idea behind this approach is that changes in the Federal funds futures in a narrow window (e.g., 30 minutes) around the FOMC monetary policy announcements capture the unexpected Fed policy actions. The key assumption is that these financial market surprises are uncorrelated with shocks other than the monetary policy ones. Since these changes are a noisy measure of the monetary policy structural shock, Gertler and Karadi (2015) used them as instruments in a proxy-SVAR framework. 25 One drawback of this approach is that it is not immune to policy foresight issues coming from the mismatch of the information set of the private agents and the policymakers. In other words, the FOMC may have additional information about the future path of the economy. Without accounting for these different information sets in the VAR, the shock may incorporate the endogenous response of the policy instrument to the expected future path of macroeconomic variables (Ramey, 2016). Furthermore, these shocks can be correlated with the Fed's real-time forecasts of relevant macroeconomic variables, and therefore not exogenous. 26 In contrast, the narrative approach used in the baseline analysis does not suffer from these problems, as the shocks are orthogonal to the Fed's real-time forecasts of relevant macroeconomic variables.
We estimate Gertler and Karadi (2015)'s monthly reduced-form VAR over the period 1990M1-2012M12. The VAR includes U.S. industrial production and the consumer price index (both in logarithm), the government bond yields at different maturities (the policy indicator), and the excess bond premium constructed by Gilchrist and Zakrajšek (2012), which is an indicator for refinancing conditions on the secondary corporate bond markets. We then apply Ramey (2016)'s three-step approach to extract the structural monetary policy shocks, and, following Cloyne et al. (2018), we aggregate these surprises at a quarterly frequency and use them as instruments in our analysis of crossborder banking flows. Specifically, we modify equation (3), substituting the  monetary policy shock with the policy instrument (Treasury bond yield) instrumented with the VARestimated monetary policy shock: In terms of covariates, , consists of the same set of control variables as in equation (3), namely the lags of the dependent variable, the lags of the estimated monetary policy shocks, and other control variables in the recipient country j. As any other instrumental variable framework, additionally to the exogeneity condition (i.e., the external instrument must be uncorrelated with the other structural shocks), the instrument must satisfy the relevance condition-namely, it should be contemporaneously and highly correlated with the instrumented variable. Following Gertler and Karadi (2015), we test different combinations of policy indicators (three and six-month government bond yields as well as one/two/three/five/seven/ten-year government bond yields) and instruments (six/nine-month and one- year ahead on three-month Eurodollar deposits, current/one/two-month Fed funds futures). Based on the F-test of the joint model (the F-statistic on the first stage) and the Stock and Yogo (2005) critical values, we chose the combinations that pass the weak instrument test. Second, we have used four lags of the dependent variable and the control variables in the baseline analysis. We demonstrate that our findings do not depend critically on the selection of lags.
Panel B in Figure A.7 shows that our results hardly change with the selection of eight lags.
Third, while we have clustered standard errors at the time level in the baseline specification, we test the robustness of our findings by clustering standard errors at the recipient country level or at the recipient country-time level. We also compute Driscoll-Kraay standard errors that allow arbitrary correlations of the errors across countries and time. We only report the results from using Driscoll-Kraay standard errors in Panel C in Figure A.7 to save space, but the results obtained using standard errors clustered at the recipient country level and at the recipient country-time level are similar to those clustered at the time level. In sum, the statistical significance of our findings does not hinge on the way we account for the correlations in the error term.
The use of the recipient country-fixed effects and a recipient country's macroeconomic variables cannot fully control for potential time-varying factors affecting cross-border banking flows at the bilateral level. One obvious candidate for such factors is bilateral trade flows between the U.S. and its counterpart countries. This variable is particularly relevant for the study of international capital flows, as the current account and the financial account are tightly related by the accounting identity.
While banking flows correspond to only a subset of total capital flows, therefore mitigating this problem, we test the robustness of our findings by controlling for bilateral trade flows. 29  where is an indicator of the state of the economy normalized to have zero mean and unit variance.
The estimated parameters depend on the average behavior of the economy in the historical sample between t and t+h, given the shock, the initial state, and the control variables.
Since the parameter estimates on the control variables incorporate the average tendency of the economy evolving between the states, the estimates incorporate both natural transitions and endogenous transitions from one state to the other that occur in the data. is the five-quarter moving average of real GDP growth and ( ) is a smooth transition function used to estimate the effect of monetary policy shocks in expansions vs. recessions. 32 We choose = 1.5 following Auerbach and Gorodnichencko (2012) so that the economy spends about 20 percent of the time in a recessionary regime. 33 As shown in Figure A.9 in the appendix, the probability of a recession regime we estimate using a smooth transition function captures well the official NBER recession dates.
This approach is equivalent to the smooth transition autoregressive model developed by Granger and Teräsvirta (1993) and has the following advantages. The effect of U.S. monetary policy on cross-border banking flows may also depend on global financial conditions. To test this hypothesis, we repeat a similar exercise to the one for the state of the business cycles but identifying global financial condition regimes based on the VIX. 34 Figure 5 shows that the international bank lending channel of U.S. monetary policy tends to be weaker during the high- 32 The results, available upon request, are similar when considering a measure of output gap. 33 Our results hardly change when using alternative values of the parameter , between 1 and 6. uncertainty period (high-VIX period), consistent with the recent findings that heightened uncertainty reduces the effectiveness of monetary policy (Aastveit et al., 2017;Castelnuovo and Pellegrino, 2018).
Monetary policy tightening vs. easing. The effect of monetary policy on economic activity is found to be larger for monetary policy tightening than easing (Cover, 1992;Tenreyro and Thwaites, 2016).
To test for a similar asymmetry on its effects on cross-border banking flows, we estimate the following specification: , +ℎ − , −1 = ℎ + + ℎ ℎ where is a dummy variable that takes a value of one for monetary policy tightening and zeroes otherwise, and + ℎ and − ℎ capture the effect of a monetary tightening and easing, respectively. The results presented in the top panel of Figure

Borrower country riskiness and international portfolio rebalancing channel of U.S. monetary policy.
In this section, we investigate whether borrower country riskiness affects the effect of U.S. monetary policy on cross-border bank lending by utilizing the heterogeneity in the pool of foreign borrowers. borrowers abroad. If this mechanism is at work, we would observe a larger decline in cross-border bank lending toward relatively riskier recipient countries. Since measuring ex-ante riskiness of borrowers at a country level is a challenging task we take two complementary approaches to enhance the credibility of our results. First, we adopt the Political Risk Index of ICRG (International Country Risk Guide) and separate the recipient countries into two groups based on their sample average of the total ICRG index.
A country with higher than the above median value is considered as a safe borrower, and vice versa.

©International Monetary Fund. Not for Redistribution
Second, we simply group the recipient countries into advanced and emerging market economies based on their IMF classification. Figure 6 shows the results using the ICRG index as a proxy for borrower riskiness. The effects on cross-border bank lending are negative and statistically significant in both cases. Moreover, their magnitudes are similar and the difference is not statistically significant. Figure A.11 in the appendix also confirms that the effects are similar when using alternative measures of riskiness. 35 Although detailed counterparty-level data is required to draw a more complete picture of the underlying mechanism, the evidence from the recipient-country level data does not support the prediction of the international portfolio rebalancing channel.

C. International bank lending channel of monetary policy in other advanced economies
So far, we have focused on the international bank lending channel of U.S. monetary policy.
However, a natural question is whether we can generalize the U.S. results to other systemically important advanced economies. Despite its paramount importance in policymaking, the existing studies have not reached a consensus on this issue. To shed light on this issue, we extend equation (3) to incorporate the bilateral panel structure of the data as follows: 35 We also test whether there is an asymmetry in cross-border lending toward borrowers in euro vs. non-euro area in response to U.S. monetary policy shocks. Figure A.12 in the appendix shows that the spillover tends to be stronger toward euro-area borrowers.
, , +ℎ − , , −1 = , ℎ + , ℎ + ℎ ℎ , + ∑ ℎ , , − =1 + , , +ℎ , where i and j indicate source and recipient country, respectively; , , is the log of cross-border lending from global banks located in country i to borrowers in country j in time t; , ℎ is a source-recipient fixed effect; , ℎ is a recipient-time fixed effect; ℎ , is the measure of exogenous monetary policy shocks in country i described earlier; , , is a set of control variables, including four lags of the dependent variable and of the monetary policy shocks.

IV. CONCLUSION
We Taken together, our findings bear significant implications for both policymakers (central banking policies and international monetary and financial coordination) and academics.
Stretching somewhat further, we make some methodological innovation, which is useful for future applied works. The dynamic estimation framework of local projections applied to the bilateral dataset allows estimating impulse response functions, which are not straightforward using a large bilateral dataset. The impulse response functions we estimate are also consistent with the spirit of earlier works on the domestic bank lending channel of monetary policy using VARs. Our findings suggest that a static estimation framework adopted in the existing studies using the LBS may not be adequate to identify the channel of monetary policy spillovers. To our best knowledge, this paper is one of the first kind to apply such a dynamic estimation framework to a large international bilateral dataset, thereby advancing an econometric framework for empirical researchers.
Two areas of future research we believe are important. First, we should enhance our understanding of the international transmission channel of monetary policy by further disentangling the bank lending and the risk-taking channel. Second, we should test whether the effect of unconventional monetary policy has been different than that of conventional monetary policy presented in this paper and in previous studies.  (3). Autocorrelation and heteroskedasticity-consistent standard errors are clustered at the time levels. *** denotes 1% significant level, ** denotes 5% significance level, and * denotes 10% significance level. While we control for the four lags of the variables, we only report the estimation results up to one lag to save space here. Note: The dependent variables are the growth rate of exchange rate-adjusted U.S. bilateral cross-border claims. The measure of monetary policy shocks is the lagged Federal funds rate in column (I) and (IV), the changes in the Federal funds rate in column (II) and (V), and the exogenous monetary policy shocks from Coibion (2012) in column (III) and (VI). Heteroskedasticity-robust standard errors are clustered at the time levels. *** denotes 1% significant level, ** denotes 5% significance level, and * denotes 10% significance level.          We compute the time-series average of the status regarding the exchange rate regime, capital account openness, and monetary policy independence. A country with "*" denotes that it is also a source country where monetary policy shocks are originated in the second part of the analysis.        and Coibion (2012), proxy-SVAR identified shock using two-year treasury yield, instrumented with threemonth ahead Fed funds futures.

Appendix B. Estimation results from additional exercises
In this online appendix, we provide the estimation results from several additional exercises.
While these results are not first-order important, they still provide some insights on the interpretation of the main results.
Domestic effect of U.S. monetary policy shocks. We enhance the credibility of the identified shocks by analyzing separately the effect of monetary policy shocks on domestic economic variables and check whether the responses are consistent with the theoretical prediction. As shown in Figure B.1, the results are generally consistent with  and Coibion (2012) despite some differences in the sample period. Consistent with the theoretical prediction, output and investment decline, while the nominal exchange rate appreciates in response to exogenous monetary policy tightening. Although we find a weak price puzzle on impact, the increase in CPI is not statistically significant.
We also analyze the response of domestic bank lending to the monetary policy shock to deepen the understanding of our findings. Under the bank lending channel of monetary policy, domestic bank lending must also decrease due to an increase in a bank's funding costs. We download domestic bank lending data "Bank credit to the private non-financial sector" from BIS and deflate them using U.S.
CPI to obtain a real measure of bank lending.
Consistent with the prediction of the bank lending channel of monetary policy, we find a significant decline in domestic bank lending following monetary policy tightening. However, one should note that the magnitude of the response cannot be directly compared to the case of cross-border bank lending in the main text for two reasons. First, counterparty entities for domestic claims are the non-financial private sector, which is narrower than those in cross-border claims. Second, the estimation results in the main text are derived after controlling for a variety of fixed effects and covariates in recipient countries.

Econometric issues in the analysis of regressions with generated regressors. Econometric issues
regarding the consistency in the estimated standard errors may arise because the shocks used in the baseline regression is an estimate (generated regressors). However, one should note that the shocks we use are the residuals, not the predictor from the first-stage regression. When the residuals are used as a shock, even an OLS estimate of the standard error is consistent (see Pagan, 1984 for detailed discussions on the issue of generated regressors). Nevertheless, we test the robustness of our findings by instrumenting changes in the federal funds rate using  where , is an indicator variable regarding the trilemma status of each recipient country j in time t and , includes the four lags of , in addition to the previous control variables.
As emphasized by , to the extent that the exchange rate regime or capital account openness varies over time, using the time-invariant characteristics-a common practice in VAR studies-could bias the results toward finding less stark difference driven by monetary policy shocks across country groups. In this case, the results in Rey (2013)  We use the trilemma index constructed by  to test how the Mundellian trilemma characterizes the degree of spillovers of U.S. monetary policy through the international bank lending channel. Their index quantifies the degree of achievement along the three dimensions of the trilemma hypothesis: exchange rate stability, monetary policy independence, and financial openness, thereby providing a comprehensive and consistent overview of an individual recipient country's trilemma status. Here, we describe each of the three indices only briefly. See  for further details about the construction of the index and some caveats in its interpretation.
In principle, annual standard deviations of the monthly exchange rate between the home country and the base country are calculated to measure exchange rate stability and the index is normalized between zero and one. 37 The extent of monetary independence is measured as the reciprocal of the annual correlation of the monthly money market rates between the home country and the base country and normalized between zero and one. We use the updated version of the Chinn-Ito index (KAOPEN) to measure capital account openness (Chinn and Ito, 2008). Since KAOPEN is based upon reported restrictions, it is necessarily a de jure index of capital account openness. 38 Since a recipient country fixed effect will absorb any time-invariant recipient country characteristic in our specification, it is important to note that what we identify is the within variation in the time-varying trilemma index. The base country is defined as the country that a home country's monetary policy is most closely linked with as in Shambaugh (2004). Since we are interested in the cross-border spillovers of U.S. monetary policy shocks, we use its base country's value with respect to the U.S. when the base country of a sample country is not the U.S. For example, since Belgium's base country is Germany, the Belgian exchange rate regime is floating vis-à-vis the U.S., although it is pegged to Germany. 38 We focus on the KAOPEN measure of capital controls in Chinn and Ito (2008), updated in July 2017. KAOPEN is based on the four binary dummy variables that codify the tabulation of restrictions on cross-border financial transactions reported in the IMF's Annual Report on Exchange Arrangements and Exchange Restrictions: (i) capital account openness; (ii) current account openness; (iii) the stringency of requirements for the repatriation and/or surrender of export proceeds; and (iv) the existence of multiple exchange rates for capital account transactions. KAOPEN index's main merit is that it attempts to measure the intensity of capital controls insofar as the intensity is correlated with the existence of other restrictions on international transactions. 39 We also use the updated version of binary regime classification by Shambaugh (2004) to sort out de facto pegged and floating exchange rate regimes for robustness checks. We use the most basic measure of the exchange rate regime employed in Shambaugh (2004). In Shambaugh's classification, a country is classified as pegged if its official nominal exchange rate stays within ±2 percentage bands over the course of the year against the base country. Non-pegs are also assigned a base determined by the country they peg to when they are pegging at other times in the sample. The floating regime does not necessarily include pure floats only but includes all sorts of non-pegged regimes. Probably due to the binary nature, we find even less stark difference in this case. meaningful difference between the regime emerges when monetary policy independence is considered.
Panel C shows that the spillover tends to be stronger when the recipient country maintains monetary policy independence (i.e., not increasing the interest rate in response to U.S. monetary policy tightening).
The insignificant difference across the exchange rate regime and the degree of capital controls might be driven by the high correlation between the two. Financial market development may lead a country to become more prepared to adopt greater exchange rate flexibility (i.e., abandoning peg) and open its capital markets to international investors. Indeed, the correlation between the average exchange rate stability index and the capital openness index is -0.54 (p-value of 0.005). Such a strong negative relationship is evidence of the so-called "binding" trilemma  and suggests that ignoring the mutual dependence would bias the estimation results.
Thanks to the flexibility provided by local projections, we can consider the effect of the exchange rate regime and capital account openness jointly. To sharpen the identification of the trilemma, we construct a two-by-two regime based on the interaction between the exchange rate stability index and capital account openness index. Figure   Note: The graph shows the response of cross-border bank lending to a 100 bp increase in the Federal funds rate using the exogenous variation from  as an instrument (left panel) and a 100 bp increase in policy rates in other advanced economies using the exogenous variation from Furceri et al. (2018) as an instrument (right panel). Horizon h=0 captures the impact of the shock, and the units are in percentage.