Global Banking and Firm Financing: A Double Adverse Selection Channel of International Transmission

Board of Governors of the Federal Reserve System International Finance Discussion Papers Number 1325 August 2021 Global Banking and Firm Financing: A Double Adverse Selection Channel of International Transmission Leslie Sheng Shen Please cite this paper as: Shen, Leslie Sheng (2021). “Global Banking and Firm Financing: A Double Adverse Selection Channel of International Transmission,” International Finance Discussion Papers 1325. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/IFDP.2021.1325. NOTE: International Finance Discussion Papers (IFDPs) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the International Finance Discussion Papers Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers. Recent IFDPs are available on the Web at www.federalreserve.gov/pubs/ifdp/. This paper can be downloaded without charge from the Social Science Research Network electronic library at www.ssrn.com.

Global Banking and Firm Financing: A Double Adverse Selection Channel of International Transmission∗ Leslie Sheng Shen† Federal Reserve Board of Governors April 2021 Abstract Thispaperproposesa“doubleadverseselectionchannel” ofinternationaltransmission. It shows, theoretically and empirically, that financial systems with both global and local banks exhibit double adverse selection in credit allocation across firms. Global (local) banks have a comparative advantage in extracting information on global (local) risk, and this double information asymmetry creates a segmented credit market where each bank lendstotheworstfirmsintermsoftheunobservedriskfactor. Givenabankfunding(e.g., monetary policy) shock, double adverse selection affects firm financing at the extensive and price margins, generating spillover and amplification effects across countries. ∗Anearlierdraftofthepaperwascirculatedunderthetitle“Globalvs. LocalBanking: ADoubleAdverseSelection Problem.” IamgratefultoPierre-OlivierGourinchas, UlrikeMalmendier, YuriyGorodnichenkoandAnnetteVissing- Jorgensenforinvaluableadviceandguidance. IalsowouldliketothankCarlosAvenancio-Leon,ThorstenBeck,Satoshi Fukuda,LindaGoldberg,GalinaHale,RebeccaHellerstein,SebnemKalemli-Özcan,AmirKermani,RossLevine,Robert Marquez,BenMoll,DamianPuy,JonathanRagan-Kelley,NickSander,NeeltjevanHoren,GaneshViswanath-Natraj, Michael Weber, James Wilcox, and seminar participants at the Federal Reserve Bank of New York, Federal Reserve Board, George Washington University, London School of Economics, University of Maryland, Santa Clara University, UCBerkeleyEconomicsandHaas,UniversityofBritishColumbia,UniversityofPittsburg,theECBForumonCentral Banking,CEBRAInternationalFinanceandMacroeconomicsMeeting,NorgesBankWorkshoponFrontierResearchin Banking, EmergingScholarsinBankingandFinanceConference, andtheannualmeetingoftheAmericanEconomics Associationforhelpfulcomments. MichaelNguyen-mason,YingjieWu,LisaZhang,andQinZhangprovidedexcellent researchassistance. ThisworkissupportedbyresearchfundingfromtheClausenCenter. Theviewsinthispaperare solelytheresponsibilityoftheauthorandshouldnotbeinterpretedasreflectingtheviewsoftheBoardofGovernors oftheFederalReserveSystemoranyotherpersonassociatedwiththeFederalReserveSystem. †Shen: FederalReserveBoard,20thandCStreetsNW,Washington,D.C.20551(email: leslie.shen@frb.gov).

1 Introduction How economic and financial shocks transmit across countries is a fundamental question in international finance. A growing body of literature has pointed out that global banks—banks thatlendtoforeignentitiesthroughcross-borderloans—playakeyroleintransmittingshocks abroad.1 Indeed,thepasttwodecadeshaveseenanexplosionincross-borderlendingbyglobal banks: global banking loans have more than tripled since the mid-1990s, reaching almost $15 trillionandaccountingforaround20%oftotaldomesticprivatecreditforanaveragedeveloped ormajoremergingmarketeconomy(AppendixFigureB.1). Existingliteraturemostlystudies the role of global banks in transmitting shocks by tracing changes in the credit they supply in response to monetary policy and liquidity shocks.2 In contrast, this paper delves into the economic mechanism that explains how global banking credit is allocated across firms in the firstplace, andshowsthattheunderlyingmechanismgeneratesanewchannelofinternational transmission. I propose an information view of credit allocation in financial systems with both global and local banks. I show, theoretically and empirically, that bank specialization in global versus local information constitutes a key mechanism driving credit allocation across firms in such globalized financial systems. Global banks specialize in information on global risk factors, and local banks specialize in information on local risk factors. This micro-foundation reveals a problem of double adverse selection in credit allocation: each bank type lends to the worst firms in terms of the risk characteristic it does not specialize in. This double adverse selection gives rise to a new channel of international transmission: when one bank is hit with a funding shock (e.g., monetary policy shock), firm financing is affected at both the extensive and intensive (interest rate) margin, generating spillover and amplification effects through adverse interest rates—a “double adverse selection channel” of international transmission. The information view of credit allocation builds on a long-standing literature in banking highlighting that the special role of banks derives from their ability to collect and process information,whichiskeyindeterminingbank-firmrelationships(CampbellandKracaw1980, Diamond1984,Ramakrishnanetal.1984,andBoydandPrescott1986). Ishow,however,that the traditional theory is not sufficient to explain bank-firm sorting in a globalized financial 1 See, among others, Cetorelli and Goldberg (2012a), Schnabl (2012), De Haas and Lelyveld (2014), Rey (2016),BräuningandIvashina(2020),Ongenaetal.(2017),Miranda-AgrippinoandRey(2018),Moraisetal. (2018),Takáts and Temesvary (2020), and Baskaya et al. (2021). 2Existingliteratureshowsthat,followingmonetarypolicyshocks,globalbanksadjustcross-borderflowsto othercountriesthroughbothexternalcapitalmarketsandinternalcapitalmarkets,increasingtheinternational propagationofdomesticliquidityshocks. Moraisetal.(2018)findthatUSandEuropeanglobalbanksincreases thesupplyofcredittoMexicanfirmsfollowingasofteningofmonetarypolicyathome. CetorelliandGoldberg (2012a)showthatglobalbanksactivelyuseinternalcapitalmarketstoreallocatefundsbetweentheheadoffice and their foreign offices in response to monetary policy shocks. 2

system with both global and local banks. The traditional theory posits that banks and firms sort based on hard and soft information: large banks are more likely to lend to firms with more readily available hard information, which tend to be large and established firms, while small banks are more likely to establish relationships with firms with more soft information, which tend to be small and young firms. Mapping this theory to the context of bank-firm sortinginglobalizedfinancialsystems,3 onewouldconjecturethatglobalbanksaremorelikely to lend to firms with more hard information, since global banks tend to be larger than local banks. However, using a detailed bank-firm loan-level dataset spanning across 24 countries, I find that both global and local banks lend to firms across the entire asset size and age distribution. This finding points to a puzzle in the mechanism driving the allocation of global banking credit across firms: why do firms of similar size and age borrow from different types of banks?4 In light of this puzzle, I raise a new perspective. I conjecture that global and local banks differ in their specialization in global and local information: global banks have a comparative advantageinextractinginformationonglobalriskfactors,andlocalbankshaveacomparative advantage in extracting information on local risk factors. Each bank type’s comparative informationaladvantageplaysakeyroleindeterminingbank-firmsortinginfinancialsystems withbothbanktypes. Thisideaismotivatedbytheobservationthatglobalbanksareuniquely positioned to extract information on global factors through global market making activities and research efforts within the banking organizations.5 At the same time, local banks are more conveniently positioned to extract information on local factors through local lending relationships (Petersen and Rajan 1994, Berger et al. 2005). Toanalyzeandtesttheconjecture,Ifirstdevelopamodeltoformalizethenewperspective and derive empirical predictions. I specifically focus on the prediction for bank-firm sorting in globalized financial systems and the implication for international transmission of bank funding shocks. I then test the model predictions using detailed cross-country loan-level data and empirical strategies that tightly maps to the model set-up. The model features an economy comprised of global and local banks, and firms that have returns dependent on global and local risk factors. The key ingredient of the model is each banktype’scomparativeinformationaladvantage: globalbankshavethetechnologytoextract 3 Globalized financial systems here refer to markets comprised of global and local banks as well as firms with access to both banks, which tend to be firms above a certain size threshold. 4 Another mechanism we may conjecture driving the sorting may be bank specialization in loans of particular currency denominations. I provide evidence in Section 2 showing that, in fact, global and local banks lend in both local and non-local currencies. 5 For instance, global banks heavily recruit PhD economists to work in their macro research departments. See past and current job listings from global banks such as Citi, JP Morgan, and Goldman Sachs on the American Economic Association’s Job Openings for Economists site: https://www.aeaweb.org/joe/listings. 3

informationonglobalfactorsbutnotlocalfactors, andviceversaforlocalbanks. Thisdouble information asymmetry is common knowledge and thereby incorporated in both the banks’ loan contract pricing and the firms’ selection of lenders. Each bank prices loans based on the component of firm return it observes, as well as its expectation of the component of return it does not observe for the subset of firms that selects the respective bank. Each bank type holds Nash-type conjectures about the other bank type’s loan pricing and plays best response strategies. The equilibrium features a fixed-point solution where each bank offers the best rate to the firms that select into that bank. The model generates a sharp prediction about credit allocation across firms: firms with higher expected return based on global factors relative to local factors are more likely to borrow from global banks, and vice versa for firms with returns more dependent on local factors. This allocation reveals a problem of double adverse selection: both global and local banks are adversely selected against by the firms, as firms select into borrowing from the bank which observes the more favorable component of their returns. In other words, each bank lends to the worst firms in terms of the unobserved characteristic. I further demonstrate that the double adverse selection problem generates a double adverse selection channel of international transmission through which bank funding shocks are transmitted to foreign (and domestic) firm financing. When one of the banks is hit with a funding (e.g., monetary policy) shock, the double adverse selection affects credit allocation at both the extensive and intensive margins, generating spillover and amplification effects through adverse interest rates. Specifically, suppose global banks face a decrease in funding cost due to expansionary monetary policy in the home country of the global banks. At the extensive margin, the model predictsthatfirmswithrelativelybalancedglobalandlocalriskexposurecomponentsaremore likely to switch into contracting with global banks. The result is driven by double adverse selection: since the firms with relatively balanced global and local risk exposure receive the most adverse interest rates relative to the first-best outcome, they are more likely to switch lenders given any changes in the credit market. These marginal firms that switch away from local banks into global banks are less risky than the infra-marginal firms that continue to borrow from either the local banks or the global banks. At the intensive margin, the model predicts that i) the interest rates of the infra-marginal firms that remain with the local banks are expected to increase (i.e., a spillover effect), and ii) theinterestratesoftheinfra-marginalfirmsthatremainwiththeglobalbanksareexpectedto decreasebymorethanthedirecteffectcausedbythefundingcostchange(i.e.,anamplification effect). The spillover effect on the infra-marginal firms that continue to borrow from local banks is solely driven by an exacerbation of the adverse selection problem. Since the marginal 4

firmsthatswitchawayfromlocalbanksarelessriskythantheinfra-marginalfirms,localbanks are left with a riskier pool of firms, which induces the banks to increase interest rates, despite no changes to their funding cost. On the other hand, the impact of the funding cost shock is positivelyamplifiedforinfra-marginalfirmsthatcontinuetoborrowfromglobalbanksbecause themarginalfirmsthatswitchintoglobalbanksarelessriskythantheseinfra-marginalfirms, which alleviates the adverse selection problem for the global banks. Next, I formally test the two model predictions on bank-firm sorting and international transmission of bank funding shock, using data on global syndicated corporate loans from Dealscan hand-matched with international firm-level databases including Amadeus, Orbis, Compustat, and Compustat Global. I categorize the lead bank on each loan into global banks andlocalbanks. Theresultingsampleincludes115,166loans, borrowedby12,979firmsacross 24 countries, over the period 2004-2017. This cross-country bank-firm loan-level dataset is uniquely appropriate for this study because it captures a significant portion of cross-border lending that other loan datasets such as credit registry data would not capture. To test the model prediction on bank-firm sorting, I construct measures for each firm’s global and local risk exposure that tightly maps to the model set-up. The empirical analysis reveals a stark pattern of bank-firm sorting: as predicted by the model, global banks lend more to firms with higher exposure to global risk relative to local risk, and vice versa for local banks. To make this result more concrete, consider two firms: Oil States International, an American multinational corporation that provides services to oil and gas companies, and Zale Corporation, an American jewelry retailer that has a large presence in malls around the US. While both firms are public firms, headquartered in Texas, and of similar size (with total assets around $1.3 billion in 2017), Oil States International’s return is more dependent on global risk factors, since, as a multinational firm in the petroleum industry, it is highly exposedtoglobaldemandandsupplyshocks. Ontheotherhand, ZaleCorporation’sreturnis more exposed on local risk factors, since its main sources of sales revenue are local customers. The model predicts that on average, Oil States International is more likely to borrow from global banks, while Zale Corporation is more likely to borrow from local banks. The data confirms this prediction: banks that lend to Oil States International are mostly global banks, including Bank of Nova Scotia, Credit Suisse, and Royal Bank of Canada, while mostly local banks such as Bank of Boston, First Republic Bank Dallas, and Rhode Island Hospital Trust National Bank lend to Zale Corporation. Totestthemodelpredictionsofhowfundingshockstobanksaffectcreditallocation,Itake the Euro area as an empirical laboratory and analyze how US and Euro area monetary policy shocks affect credit allocation across firms in the Euro area, through US and Euro area banks. To identify exogenous shocks to US and Euro area monetary policy, I use high-frequency 5

data on Federal Funds futures and Euribor futures. I find that an expansionary shock to US monetary policy induces firms in the Euro area with relatively balanced global and local risk componentstoswitchtheirborrowingfromEuroareabankstoUSbanks, conditionalonEuro area monetary policy. The analogue applies to an expansionary shock to Euro area monetary policy. Furthermore, I find that, conditional on Euro area (US) monetary policy and given expansionary US (Euro area) monetary policy, the interest rates of the infra-marginal firms that continue to borrow from Euro area (US) banks increase, reflecting a spillover effect. At the same time, the interest rate spreads of the infra-marginal firms that continue to borrow from US (Euro area) banks decrease, reflecting an amplification effect. The results are consistent with the model prediction on the effects of bank funding shocks on credit allocation at both the extensive and intensive margins, revealing a double adverse selection channel of international transmission. This channel not only sheds new light on how global banks propagate the transmission of shocks but also clarifies an existing view on this issue, namely the “international risk-taking channel” ofmonetarypolicytransmission.6 Thenewchannelrevealsthattheempiricalresults which the existing literature (e.g., Morais et al. 2018) points to as evidence for risk-taking behavior by global banks could be confounded with a force generated by the adverse selection problem, namely, substitution between global banking credit and local banking credit. Related Literature This paper contributes to three broad strands of literature. First, it contributes to the literature on the role of financial intermediaries in the transmission and amplification of shocks in the international context. While a long strand of work has studies the credit channel of monetary policy transmission (see e.g., Bernanke and Blinder 1992, Bernanke and Gertler 1995, and Kashyap and Stein 2000), there have been less work on the creditchannelintheinternationaldimensionuntiltheglobalfinancialcrisis. Sincethen,global banks emerged as a key channel for international transmission of liquidity conditions and monetary policy, sparking both theoretical and empirical research. On the theoretical front, several recent papers have introduced models with global banks for studying international transmission, including Dedola et al. (2013), De Blas and Russ (2013), Niepmann (2015), BrunoandShin(2015b),andAokietal.(2016). Whilethesemodelssolelyfocusonemergence and implications of one type of bank,7 this paper points out that the competitive interaction 6 The international risk-taking channel of monetary policy transmission is based on the view that low monetary policy rates and QE in developed economies could induce banks to lend to riskier firms abroad (Bruno and Shin 2015a, Coimbra and Rey 2017, and Morais et al. 2018). 7 In the framework in Bruno and Shin (2015b), there are both global and local banks. But local banks simply act as a conduit that intermediates funds from global banks to firms, which essentially make only one type of bank active in the economy. 6

between global and local banks plays an important role for international transmission. On the empirical front, a growing literature uses bank-level and loan-level data to trace out the channels through which global banking affects domestic bank lending, including Cetorelli and Goldberg (2012b), Popov and Udell (2012), Schnabl (2012), De Haas and Lelyveld (2014), Ivashina et al. (2015), and Baskaya et al. (2021). This paper contributes to this line of work bypointingoutanewchannelofinternationaltransmissionthroughglobalbanks—thedouble adverse selection channel of international transmission. Moreover, this channel is new to the literature on international transmission of monetary policy. RecentpapersbyRey(2016)andMiranda-AgrippinoandRey(2018)provideevidence of large spillovers of US monetary policy on credit creation around the world, suggesting global banks as the main source for transmission. Existing work has pointed to currency mismatches on global banks’ balance sheets (Ongena et al. 2017, Bräuning and Ivashina 2018, Bräuning and Ivashina 2020) and internal capital markets within global banks (Cetorelli and Goldberg 2012a) as channels of international monetary policy transmission. In addition, low international monetary policy rates and expansive quantitative easing in large developed economies over the past decade have prompted debates on the extent of a bank risk-taking channel of monetary policy transmission, as explained in Borio and Zhu (2012), Bruno and Shin (2015a), and Coimbra and Rey (2017). Morais et al. (2018), using bank-firm loan data, show that low monetary policy rates and QE in developed economies led global banks to increase credit supply to firms in Mexico, especially firms with higher-than-average exante loan rates. They consider this to be evidence of bank risk-taking. Contrary to their explanation, I show that the force driving increased credit supply to riskier firms could be substitution between global banking credit and local banking credit, raising adverse selection as a new channel of international transmission of monetary policy. Second, thispapercontributestotheliteratureinbanking. ThenewperspectiveIpropose builds on the traditional information view of banking from classic papers by Campbell and Kracaw (1980), Diamond (1984), Ramakrishnan et al. (1984), and Boyd and Prescott (1986). Through this lens, a subsequent strand of literature argues that different banks specialize in hard versus soft information, and lend to different types of firms as a result (Petersen and Rajan1994, Stein2002, Bergeretal.2005, andLibertiandPetersen2018). Iprovideevidence showing that hard versus soft information is insufficient for explaining bank-firm sorting in globalized banking systems, and propose an alternative dimension of bank specialization.8 In the context of cross-border banking specifically, this paper is related to the strand of banking literature that studies the effects of foreign bank entry on credit access. The frameworkdevelopedinthispaperbuildsontheworkbyDell’AricciaandMarquez(2004),Sengupta 8 Section 2 describes the traditional theory and the relevant empirical tests in detail. 7

(2007), Detragiache et al. (2008), and Gormley (2014), which emphasize the importance of (imperfect) information in shaping competition and credit allocation in economies with both local banks and foreign banks. The focus of that line of studies is foreign bank entry into low-income countries, where overall information asymmetries may be large. Local banks are considered to have an informational advantage over the foreign banks, which, as a result, are able to target only the largest or the least informationally opaque firms. In contrast, the focus of this paper is cross-border lending by global banks in developed economies, where the majority of global banking activity occurs. What sets this paper apart is the new perspective on how banks’ comparative advantage in different types of information, or global and local information specifically, can affect credit allocation.9 While the existing models predict that the smaller, more informationally opaque firms are more likely to borrow from local banks10, the framework in this paper predicts that some large and informationally transparent firms are still likely to borrow from local banks, as long as their returns are more dependent on local risk factors.11 Third, this paper adds to the work on the role of information frictions in international capital and credit markets. Papers have pointed out information friction as an important mechanism in driving international capital flows (Martin and Taddei 2013) and determining exchange rates (Bacchetta and Van Wincoop 2006). Albuquerque et al. (2009) specifically highlight the role of global information in international equity markets, showing that it helps explain U.S. investors’ trading behavior and performance. To the best of my knowledge, this is the first paper that studies the role of global information in international banking. Therestofthispaperisstructuredasfollows. Section2reviewsthetraditionaltheoryand presents a new puzzle on bank-firm sorting in globalized credit markets. Section 3 presents a model of global and local banking. Section 4 applies the framework to analyze bank-firm sortingandinternationaltransmissionofbankfundingshocks. Section5outlinestheempirical predictions and describes the data used for empirical testing. Section 6 presents the empirical 9 The key ingredient incorporated in my model to formalize the idea of banks’ differing specialization in global versus local information, double asymmetric information, and the ensuing result of double adverse selection,isnewtothelineofresearchincontracttheoryonadverseselectionincreditmarkets,startingwith the classic papers such as Stiglitz and Weiss (1981) and De Meza and Webb (1987). 10 Papers including Berger et al. (2001), Clarke et al. (2005), Mian (2006), and Gormley (2010) provide empirical evidence in support of this prediction, though the empirical settings studied in these papers are all low-income economies. 11 Detragiache et al. (2008), Beck and Peria (2010) and Gormley (2014) also explore the impact of foreign banking on overall credit access, relating it to debates on the benefits and costs of financial openness. They argue that foreign banking entry undermines overall access to credit since it worsens the credit pool left to localbanks,givesrisetoadverseselection,andtherebylowersoverallfinancialdevelopment. Whilemymodel also points to the possibility of a decline in aggregate credit due to adverse selection, I show that access to globalbankingcreditactuallyleadstoamoreefficientcreditallocationinthefinancialsystem. Thisisinline with papers which argue that the benefits of financial openness outweigh the costs, such as Levine (1996), Claessens et al. (2001), Edison et al. (2002), Claessens (2006), and Beck et al. (2007). 8

analysis on bank-firm sorting. Section 7 presents the empirical analysis on the effects of monetary policy shocks on credit allocation across foreign and domestic firms. Section 8 concludes. Proofs are relegated to APPENDIX A. 2 Traditional Theory and New Perspective Inthissection,Ireviewthetraditionaltheoryonbank-firmsortingandtestwhetheritpredicts the patterns of bank-firm sorting in globalized credit markets—markets comprised of global and local banks as well as firms with access to both banks. Classicbankingtheoryarguesthatbanksexistbecauseoftheiruniqueabilitytocollectand process information. Based on this view, a long strand of literature in banking and corporate finance has used the distinction between hard and soft information to explain how banks and firms establish relationships (see, e.g., Petersen and Rajan 1994, Stein 2002, Petersen and Rajan 2002, and Liberti and Petersen 2018). Hard information is information that is quantifiable, independent of its collection process, and easy to transmit in impersonal ways. Soft information is information that is not easily quantifiable, dependent on its collection process, and requires context-specific knowledge to fully understand. Theories based on this view conjecture that large banks are more likely to lend to firms with more readily available hard information, while small banks are more likely to establish relationships with firms with more soft information. As a first step to understand patterns of bank-firm sorting in globalized credit markets, I test whether the sorting patterns between firms and global versus local banks bear out the predictions of the traditional banking theory. Since global banks tend to be larger, I test whether global banks are more likely to lend to firms with more hard information, and local banks are more likely to lend to firms with more soft information, using a bank-firm loan-level dataset that spans across 24 countries and covers the period 2004-2017.12 For measures of hard and soft information, I follow the empirical literature (e.g., Berger et al. 2005 and Mian 2006), which often uses firm asset size and firm age to proxy for hard information. I sort firms into quartiles based on the distribution of firm asset size and firm age in each year in each country, and then calculate the proportion of loans given by global banks and local banks in each quartile. Figure 1 plots the distribution of lending from global and local banks over the entire sample. The plot shows that both global banks and local banks lend to firms of all sizes and ages, revealing that the traditional theory does not predict the pattern of bank-firm sorting in financial systems with both global and local banks. I further test whether the differences between global and local banks illustrated in Figure 12 See Section 5.2 of the paper for a detailed discussion of the data and data-cleaning procedure. 9

Figure 1: Bank-Firm Sorting, by Firm Size and Age Quartile 1 8. 6. 4. 2. 0 Firm Asset Size 1 2 3 4 1 8. 6. 4. 2. 0 Firm Age 1 2 3 4 Global Bank Local Bank Notes. The plot shows sorting patterns between firms and global versus local banks, with firms sorted into quartiles by asset size and age. The data sample consists of syndicated loans between global and local banks and firms across 24 countries from 2004-2017. Source: Dealscan, Amadeus, Orbis, Compustat, Compustat Global, and author’s calculation. 1 are indeed insignificant in a statistical sense. For each given variable measuring hard information, I test whether the value-weighted mean of that variable for global banks is different from that for local banks. Table 1 presents these means and their differences. The results confirm the takeaways from the graphical analysis: the differences in value-weighted means are statistically insignificant between global and local banks for firm asset size and firm age. Another conjecture about the mechanism driving the sorting between firms and global versus local banks may be bank specialization in loans of particular currency denominations. This is particularly motivated by recent papers by Maggiori et al. (2018) and Gopinath and Stein (2018) that highlight the prevalence of Dollar loans, and to a lesser extent Euro loans, in international financial markets. Given these considerations, I test whether global banks specialize in lending in non-local currencies, while local banks specialize in lending in local currency. As shown in Appendix Figure B.2, in fact, global and local banks make loans in both local and non-local currencies. This empirical observation holds even when the US or both the US and Euro area countries are excluded from the sample. The empirical evidence shows that the traditional banking theory of bank specialization 10

Table 1: Bank-Firm Sorting, by Firm Size and Age Quartile: Statistical Test (1) (2) Size Age (1) Mean: Global Bank 3.196*** 2.759*** (0.0299) (0.0208) (2) Mean: Local Bank 3.099*** 2.726*** (0.0674) (0.0367) (3) Difference 0.0969 0.0330 (0.0716) (0.0426) Observations 115,166 114,323 Notes. The dependent variable in each regression (Y) is one of the hard information variables, firm size (column 1) or firm age (column 2), coded 1-4 based on the quartile number to which each respective firm belongs. Note the firms are sorted every year by country. Row 1 and row 2 show the means for each variable forglobalbanksandlocalbanks,respectively,byrunningavalue-weightedregressionofYonaconstant. For differences in means of the two types of banks, the whole data is used in the regression and a dummy for globalbanksisadded(row3). Standarderrorsreportedinparenthesesareclusteredatthebanklevel. Source: Dealscan, Amadeus, Orbis, Compustat, Compustat Global, and author’s calculation. in hard or soft information, as well as the view of bank specialization in particular currency denominations, are insufficient to explain observed sorting patterns between firms and global versus local banks. This points to a puzzle in the mechanism driving bank-firm sorting in globalized credit markets. In light of the puzzle, I propose a new perspective. I argue that global and local banks’ differing specialization in global and local information constitutes a keymechanismforbank-firmsortingandcreditallocationinfinancialsystemswithbothtypes of banks. Global banks have a comparative advantage in extracting information on global risk, and local banks have a comparative advantage in extracting information on local risk. This new perspective builds on the classic information view of banking. Furthermore, it incorporates global banks’ unique position to acquire “global” information through global market-making activities andresearch effortsthey invest infor analyzingglobal economic and market trends. Next, I proceed to formalize the new perspective by developing a model with global and local banks in which each bank type’s comparative informational advantage serves as the key ingredient. 3 A Model of Global and Local Banking In this section I develop a model to study bank-firm sorting and credit allocation in an economy with two types of banks—global banks and local banks—and firms heterogenous in their exposure to global and local risks. Each type of bank can perfectly observe only one 11

component of firms’ risk exposure, giving rise to a double information asymmetry. I show that bank-firm sorting and credit allocation in equilibrium reveal a problem of double adverse selection. 3.1 Set-up Consider an economy with two periods (t = 0,1), a single good, and two classes of agents: firms and banks. All agents are risk neutral and cannot end with a negative amount of cash due to limited liability. Firms. There is a continuum of heterogenous firms that have access to a fixed-size project requiring an investment of one. Each firm i’s production technology is characterized by the following production function: z = zG+zL+u (1) i i i i where zG denotes firm i’s component of return due to global risk, zL denotes firm i’s compoi i nent of return due to local risk, and u denotes firm i’s idiosyncratic risk. Each component is i independently and uniformly distributed, with zG ∼ U(0,1), zL ∼ U(0,1), and u ∼ U(0,1). i i i More specifically, zG can be considered to encompass two components, zG = βGzG, where i i i βG denotes firm i’s exposure to global risk and zG denotes global risk. Similarly, zL can be i i considered to encompass two components, zL = βLzL, where βL denotes firm i’s exposure to i i i local risk and zL denotes local risk.13 Firms have full information on their returns due to global and local risk at the time of investment (period 0), while idiosyncratic risks are not realized until after investment (period 1). Firms have no private wealth; to implement the project, they need to raise one unit of external funds from a bank j through a loan contract. Banks. There are two types of banks, global banks (G) and local banks (L), denoted as j ∈ {G,L}. Theycanenterthefinancialmarketandcompeteonprojectsbyofferingstandard debt contracts. There is perfect competition within each bank type in the financial market. The key feature that differentiates global banks from local banks is their information acquisition technology on global and local information: global banks have the technology to evaluate firms’ return due to global risk (zG) but are not able to extract information on firms’ i return due to local risk (zL), while local banks are able to evaluate firms’ return due to local i risk but are not able to extract information on firms’ return due to global risk. This gives rise to an environment with double information asymmetry. The nature of the double information 13 These considerations become more applicable when mapping the model to empirics, which I describe more in detail in Section 6. 12

asymmetry problem and the distributions of the firms’ return due to global risk and local risk are common knowledge across banks and firms. Given the information structure, the loan rate offered by the two types of banks can be made contingent on the component of firm return observable to each respective bank type. Each type-contingent interest rate applies uniformly for all firms of the given observable component regardless of their unobserved return component. More specifically, global banks can offer type-contingent gross interest rate RG(zG) for firms with return component zG, and i i that rate applies for all firms with a given zG regardless of zL. Similarly, local banks can i i offer type-contingent interest rate RL(zL) for firms with return component zL, and that rate i i applies for all firms with a given zL regardless of zG. i i It follows that the interest rates offered by each type of bank can be generated by interest rate functions that map the observable return components to type-contingent interest rates fromtherespectivebanktype: globalbanksoffercontractsbasedontheinterestratefunction RG : zG (cid:55)→ RG(zG), and local banks offer contracts based on the interest rate function i i RL : zL (cid:55)→ RL(zL). For both types of banks, each bank’s objective is to maximize expected i i profitacrossfirmsofeachobservabletype: globalbanksmaximizeexpectedprofitacrossfirms of each given zG, and local banks maximize expected profit across firms of each given zL. i i GlobalbanksandlocalbanksfacegrossfundingraterG andrL, respectively, forthefunds they intermediate.14 Bank-Firm Sorting. This environment in which each type of bank can perfectly observe only one component of the firms’ return, while firms have full information on both return components, gives rise to a sorting process between banks and firms. The timing of the model is presented in Figure 2. Let E denote the expectation of firm i conditional on its information set. Between global i banks and local banks, each firm i selects the contract offered by bank j ∈ {G,L} that yields the higher expected utility E (cid:2) max(z −Rj(zj),0) (cid:3) .15 Firm selection results in a partition i i i of the set of all firms into two subsets, as each firm i with return component (zG,zL) selects i i to borrow from either a global bank or a local bank given the interest rate functions of the two bank types. One subset, denoted as SG, chooses to sign a lending contract with a global bank, and the other subset, denoted as SL, chooses to sign a lending contract with a local 14 Since the funding market is not of central importance to this paper, it is not explicitly modeled for analyticalconvenience. ThefundingratesrG andrL couldreflectfundingconditionsintheinterbankmarket, the deposit market, or other risk premiums. While funding is fully elastic here, the model predictions do not change if rG and rL are considered to be decreasing in loan amounts. 15 Note that the expectation here is taken with respect to idiosyncratic shocks only. 13

Figure 2: Model Timeline bank: (cid:26) (cid:27) SG = (zG,zL) : E (cid:2) max(z −RG(zG),0) (cid:3) ≥ E (cid:2) max(z −RL(zL),0) (cid:3) ; (2a) i i i i i i i i (cid:26) (cid:27) SL = (zG,zL) : E (cid:2) max(z −RL(zL),0) (cid:3) > E (cid:2) max(z −RG(zG),0) (cid:3) . (2b) i i i i i i i i The following assumptions about firm selection hold throughout the paper. Assumption1. SupposeRG(zG) > zG+zL+1orRL(zL) > zG+zL+1. Then(zG,zL) ∈ SG i i i i i i i i if RG(zG) ≤ RL(zL); and (zG,zL) ∈ SL otherwise. i i i i Assumption1statesthatintheregionoftheparameterspacewhenthefirm’sexpectedutility is zero when it borrows from either a global bank or a local bank, it chooses the bank that offers the lower interest rate. This assumption ensures that there is no ambiguity in firm selection across all regions of the parameter space. Remark 1. Based on Equations (2a) and (2b) and Assumption 1, each firm i selects into borrowing from a global bank if and only if RG(zG) ≤ RL(zL), and each firm i selects into i i borrowing from a local bank if and only if RG(zG) > RL(zL). In sum, each firm chooses the i i bank that offers the lowest rate. The selection of firms directly affects global and local banks’ expected profits. Let E G denote the expectation of a global bank conditional on its information set and E denote L 14

the expectation of a local bank conditional on its information set. The expected profits for a global bank (G) from lending to firms of a given zG and a local bank (L) from lending to i firms of a given zL are given by i (cid:90) (cid:18) (cid:19) G: E [π (zG)] = min zG+zL+u ,RG(zG) dF (zL,u )−rG, G G i i i i i G1 i i G1 (3a) (cid:26) (cid:27) where G 1 (z i G) = (z i L,u i ) (cid:12) (cid:12) z i L: (z i G,z i L) ∈ SG, 0 ≤ u i ≤ 1 ; (cid:90) (cid:18) (cid:19) L: E [π (zL)] = min zG+zL+u ,RL(zL) dF (zG,u )−rL, L L i i i i i L1 i i L1 (3b) (cid:26) (cid:27) where L 1 (z i L) = (z i G,u i ) (cid:12) (cid:12) z i G: (z i G,z i L) ∈ SL, 0 ≤ u i ≤ 1 . The first term on the right hand side of Equations (3a) and (3b) is the expected gross return across loan contracts to firms of a given zG and zL for a global bank and a local bank, i i respectively. In the global bank’s expected profit function, G (zG) summarizes the set of 1 i firms which select global banks given zG. This includes firms with idiosyncratic risk u from i i any part of the u distribution, and zL such that they are in the subset of firms that choose i i the global bank’s contract. Similarly in the local bank’s expected profit function, L (zL) 1 i summarizes the set of firms which select local banks given zL. This includes firms with i idiosyncratic risk u from any part of the u distribution, and zG such that they are in the i i i subset of firms that choose the local bank’s contract. The integrand in both equations shows the relationship between bank profit and firm profit in a standard debt contract: for each firm, when its realized return is less than the contractual interest rate, it defaults and gives up any realized project returns to the lending bank; otherwise, the firm is able to repay the loan at the contractual rate and keeps the difference between the project return and rate as profit. F (.) and F (.) denote the cumulative distribution function of the relevant variable G1 L1 conditional on G and L , respectively. The last terms in Equations (3a) and (3b) are the 1 1 funding costs for the global bank and local bank, respectively. 3.2 Strategies and Equilibrium Definition As shown in Equations (3a) and (3b), each type of bank’s choice of the interest rate function affects the expected profit of the other type of bank since it influences the subset of firms that selects loan contracts from one versus the other. I consider the competitive interplay between a global bank and a local bank as a non-cooperative game. In the game, the global bank’s strategy set UG consists of the set of possible interest rate 15

functions RG, and the local bank’s strategy set UL consists of the set of possible interest rate functions RL. The payoff function for the global bank is the expected profit function E [π (RG,RL)] across all firms, and that for the local bank is the expected profit func- G G tion E [π (RG,RL)].16 A given strategy RG is a best response to the strategy RL if L L E [π (RG,RL)] ≥ E [π (RG(cid:48),RL)] ∀ RG(cid:48) ∈ UG, and vice versa for RL. G G G G In a competitive equilibrium in this credit market, both global and local banks play best responses to each other’s strategies. Each operating bank earns an expected profit of zero given perfect competition and free entry, and the selection of firms is consistent with the banks’ equilibrium strategies. Formally, the definition of the competitive equilibrium is as follows: Definition 1. For a given set of parameters rG, rL, and the distributions of zG, zL, and u , i i i a competitive equilibrium with free entry in the credit market is a strategy profile {RG,RL} and sets SG and SL such that: 1. (No Unilateral Deviation): E [π (RG,RL)] ≥ E [π (RG(cid:48),RL)] ∀ RG(cid:48) ∈ UG; G G G G E [π (RG,RL)] ≥ E [π (RG,RL(cid:48))] ∀ RL(cid:48) ∈ UL; L L L G 2. (Zero Profit Condition, Global Bank): (cid:18) (cid:19) (cid:82) min zG+zL+u ,RG(zG) dF (zL,u ) = rG; i i i i G1 i i G1 3. (Zero Profit Condition, Local Bank): (cid:18) (cid:19) (cid:82) min zG+zL+u ,RL(zL) dF (zG,u ) = rL; i i i i L1 i i L1 4. (Firm Selection): (cid:26) (cid:27) Sj = (zG,zL):E (cid:8) max[z −Rj(zj),0] (cid:9) ≥E (cid:8) max[z −Rk(zk),0],j (cid:54)=k ∈{G,L} . i i i i i i i i Part 1 of Definition 1 requires that no unilateral deviation in strategy by any bank is profitable for that bank. Parts 2 and 3 impose zero profit among global banks and local banks, respectively. Part 4 defines the set of firms that select the loan contract with either of the two types of banks in an incentive-compatible fashion. All banks that enter the market hold correct expectations about both banks’ pricing choices and the pool of firms that will accept the contract. As a consequence, the allocations of credit across firms are consistent with the banks’ equilibrium strategies. Before turning to characterizing the equilibrium in the credit market of two bank types under double information asymmetry, I describe two useful benchmarks. 16 Banks also strictly prefer making a loan with zero expected profit to not making a loan. 16

First Best. In an environment where both types of banks observe full information on each firm’s return due to global and local risk, the only margin that differentiates the loan rate charged by global banks versus local banks is the funding cost faced by each bank type. As a result,onlythebanktypewithlowerfundingcost(r)existsinthecreditmarketinequilibrium, and its interest rate function is strictly decreasing in (zG+zL). Panel (a) of Figure 3 shows i i an illustration of the first-best equilibrium in an economy with full information. The diagonal line zL +zG +1/2 = r denotes a threshold.17 The firms in the region below this threshold FB FB are not able to receive loans, as their expected profits are too low for the bank to break even in expectation. Closed Economy. In an environment where there exist only local banks that observe information on each firm’s return due to local risk, the interest rate function RL(zL) is i strictly decreasing in zL and uniform across the entire distribution of zG. Panel (b) of Figure i i 3 shows an illustration of the equilibrium in this economy. Firms with zL below zL = rL−1 i CE (firms in Regions a and c) are not able to receive loans. Relative to the first-best allocation without information asymmetries, the equilibrium in a closed economy overfunds firms whose return due to local risk is high relative to return due to global risk (firms in Region b) and underfunds firms whose return due to local risk is low relative to return due to global risk (firms in Region c). Figure 3: Benchmark Equilibrium: First-Best and Closed Economy (a) First-Best (b) Closed Economy 1 1 First-best zL =r−1/2 FB zL zL i i b No Loans zG =r−1/2 FB zL =rL−1 CE a No Loans c 1 1 zG zG i i Notes. Panel(a)illustratesthefirst-bestequilibriuminaneconomywithfullinformation. Panel(b)illustrates the equilibrium credit allocation in a closed economy in which there are only local banks. 17 Note E[u ]=1/2. i 17

3.3 Equilibrium Characterization In the following I characterize the equilibrium in a credit market of two bank types under double information asymmetry. I start by establishing the properties of the bank interest rate functions in equilibrium. Subject to the zero profit conditions from Parts 2 and 3 of Definition 1, Equation (3a) determines the global banks’ type-contingent interest rate function RG given firm selection as specified in Equation (2a), and Equation (3b) determines the local banks’ type-contingent interestratefunctionRL givenfirmselectionasspecifiedinEquation(2b). Sincefirmselection depends on interest rates from both types of banks in equilibrium, Equations (3a) and (3b) given Equations (2a) and (2b) simultaneously determine the type-contingent interest rate functions RG and RL in equilibrium. Let E [zL | (zG,zL) ∈ SG,zG] denote the global banks’ expectation of the average zL G i i i i i for the set of firms with (zG,zL) in SG conditional on zG, and E [zG | (zG,zL) ∈ SL,zL] i i i L i i i i denote the local banks’ expectation of the average zG for the set of firms with (zG,zL) in SL, i i i conditional on zL. Proposition 1 characterizes RG and RL. i Proposition 1. (Type-Contingent Interest Rate Functions) 1. RG is strictly decreasing in zG for zG ∈ [zG,1], where zG ≡ rG −E [zL | (zG,zL) ∈ i i G i i SG,zG]−1/2. i 2. RL is strictly decreasing in zL for zL ∈ [zL,1]„ where zL ≡ rL −E [zG | (zG,zL) ∈ i i L i i SL,zL]−1/2. i Part 1 of Proposition 1 establishes that the global banks’ interest rate function is strictly monotone for zG ∈ [zG,1]. The lower bound zG pins down a cut-off point on zG below which i i the expected profits of the pertinent firms are too low for the global banks to break even in expectation. In other words, zG defines the lowest zG firm to which the global banks i lend. The lower bound zG is increasing in global bank’s funding cost (rG), decreasing in the averagezL ofthesetoffirmsthatareexpectedtoselecttheglobalbank,anddecreasinginthe i expectedidiosyncraticshocksforfirms. Theexplanationforlocalbanks’interestratefunction RL established in Part 2 of Proposition 1 is analogous. Panel (a) of Figure 4 illustrates the interest rate functions in a graph with zL on the x-axis. Since global banks cannot observe i zL, RG is uniform across zL. RL is strictly decreasing in zL, as established in Proposition 1. i i i Using strict monotonicity, I next establish that the competitive interplay between global and local banks generates a unique form of horizontal segmentation in equilibrium, in which thereexistsasetofmarginalfirmsthatareindifferentbetweentakingloansfromglobalbanks and local banks. Formally, 18

Proposition 2. (Threshold Functions) Let RG = {RG(zG) | zG ∈ [zG,1]} and RL = i i {RL(zL) | zL ∈ [zL,1]}. In the region RG ∩ RL, there exist threshold functions z¯L(zG) i i i and z¯G(zL) such that: i 1. RG(zG) = RL(z¯L(zG)). i i RL(zL) = RG(z¯G(zL)). i i 2. SG = {(zG,zL) : zL ≤ z¯L(zG)}, and SL = {(zG,zL) : zL > z¯L(zG)}. i i i i i i i i SL = {(zG,zL) : zG < z¯G(zL)}, and SG = {(zG,zL) : zG ≥ z¯G(zL)}. i i i i i i i i Part 1 of Proposition 2 establishes that, for every firm with zG (resp. zL), there exists a i i threshold on zL (resp. zG), denoted as z¯L(zG) (resp. z¯G(zL)), at which both the global bank i i i i and local bank offer the same interest rate. Panel (b) of Figure 4 illustrates the threshold: for a given zG, there exists a threshold z¯L(zG) at which the interest rate functions of the two i i banks intersect, RG(zG) = RL(z¯L(zG)). i i Part2ofProposition2followsfromthemonotonicpropertyofthetype-contingentinterest rate. Given RG(zG) and RL(zL) are strictly decreasing in zG and zL, respectively, firms i i i i (zG,zL) with zL < z¯L(zG) face a lower rate from global banks and therefore select global i i i i banks (i.e, the firms are in SG). Firms with zL > z¯L(zG) face a lower rate from local banks i i and thereby select local banks (i.e, they are in SL). This idea is shown in Panel (b) of Figure 4. An analogous explanation applies to firms with zG < z¯G(zL) and zG > z¯G(zL). i i i i Parts 1 and 2 of Proposition 2 establish the existence of thresholds that segment the credit market into two parts, with global banks as the lender in one, and local banks as the lender in the other. In equilibrium, the threshold values z¯L(zG) and z¯G(zL) are determined i i by the interaction between the interest rate schedules of the global and local banks, where z¯L(zG) = (RL)−1(RG(zG)) and z¯G(zL) = (RG)−1(RL(zL)). i i i i The following corollary characterizes the threshold functions, describing how they change given changes in zG, zL, and the interest rate functions. Let z˜G be a cut-off that pins down i i an upper bound on zG, above which firms with zL from any part of the zL distribution are i i i expected to select the global bank (i.e., z¯L(zG) = 1 for all zG ≥ z˜G), and the analogue applies i i to z˜L. Corollary 1. (Threshold Functions Characterization) Let z˜G = min{zG : z¯L(zG) = 1} and i i z˜L = min{zL : z¯G(zL) = 1}. i i 1. z¯L(zG) is increasing in zG for zG ∈[zG,min(z˜G,1)]. i i i z¯G(zL) is increasing in zL for zL ∈[zL,min(z˜L,1)]. i i i 19

Figure 4: Illustration of Interest Rate Functions and Threshold Functions (a) (b) RL (z i L ) RL (z i L ) RG (z i G ) RG (z i G ) SG SL R R i i 0 1 0 z¯ L (z i G ) 1 zL zL i i (c) (d) RL (z i L ) RL (z i L ) RG (z i G ) RG (z i G ) RG (z i G′) RL′( z i L ) SG SL SG SL R R i i 0 z¯ L (z i G )z¯ L (z i G′) 1 0 z¯ L (z i G ) z¯ L′( z i G ) 1 zL zL i i Notes. Panel (a) illustrates Proposition 1, showing the monotonically decreasing property of the interest rate functions, given information asymmetry. Panel (b) illustrates Part 1 and 2 of Proposition 2, showing, for a given zG, there exists a threshold z¯L(zG) at which RG(zG)=RL(z¯L(zG). Firms below the threshold borrow i i i i from global banks; firms above which borrow from local banks. Panel (c) illustrates Part 3 of Proposition 2, showinganincreaseinzG lowersRG(zG)andincreasesz¯L(zG),holdingallelseconstant. Panel(d)illustrates i i i Part 4 of Proposition 2, showing an increase in RL(zL) increases z¯L(zG), holding all else constant. i i 20

2. z¯G(zL)isdecreasinginRL(zL)andz¯L(z¯G(zL))isincreasinginRL(zL),forzG ∈[zG,min(z˜G,1)] i i i i i and zL ∈[zL,min(z˜L,1)]. i z¯L(zG)isdecreasinginRG(zG)andz¯G(z¯L(zG))isincreasinginRG(zG),forzG ∈[zG,min(z˜G,1)] i i i i i and zL ∈[zL,min(z˜L,1)]. i The intuition for Part 1 of Corollary 1 is straightforward. Suppose there is an increase in zG from zG to zG(cid:48), or in other words, the global component of firm i’s return strengthens. i i i Global banks’ expected profit increases, and perfect competition drives down RG(zG). At the i margin, this attracts firms with higher zL to contract with global banks. Thus, the threshold i on zL increases, z¯L(zG(cid:48)) > z¯L(zG). This relationship is illustrated in Panel (c) of Figure 4. i i i The intuition for Part 2 of Corollary 1 (shown in Panel (d) of Figure 4) is as follows. Suppose the local banks’ interest rate function changes such that RL(zL) increases for some i zL ∈ [zL,min(z˜L,1)]. A higher interest rate induces a set of marginal firms to switch from i contracting with local banks to global banks, holding constant zG and RG(zG). In particular, i i the local component (zL) of the switching firms is greater than that of the firms in global i banks’originalportfolio, whichimpliesanincreaseofthethresholdz¯L(zG). Atthesametime, i the global component (zG) of the switching firms is higher than that of the firms that remain i with local banks, which implies a decrease of the threshold z¯G(z¯L(zG)). i Based on the results from Proposition 1 and 2 and Corollary 1, I next characterize the competitive interaction between the two interest rate functions offered by the two types of banks. Proposition 3. (Interaction of Rate Functions in Equilibrium) Given zG, for any increase in i RL(zL) such that z¯L(zG) increases, RG(zG) declines. Given zL, for any increase in RG(zG) i i i i i such that z¯G(zL) increases, RL(zL) declines. i i Proposition 3 points out that each bank’s type contingent interest rate function is determined by two inputs: the observed risk component of each firm’s return and the threshold valueoftheunobservedriskcomponent. ForagivenzG,ifthereisachangeinthelocalbanks’ i interest rate function RL such that the threshold z¯L(zG) increases, a set of marginal firms i with zL greater than all the zL’s in global banks’ original portfolio switches into borrowing i i from global banks. As a result, the global banks offer a lower RG(zG) for the firms with the i given zG. The interaction between the interest rates functions of global and local banks point i to a stable equilibrium in which the two banks interact as strategic substitutes. Propositions 1–3 lead to a full characterization of the equilibrium solution on RG and RL. The solutions for the equilibrium interest rates RG(zG) and RL(zL), and thresholds i i z¯L = z¯L(zG) and z¯G = z¯G(zL), for zG ∈ [zG,1] and zL ∈ [zL,1] are described in detail in i i i i i i APPENDIX A, I.A. 21

4 Model Analysis and Implications In this section, I analyze the model by studying bank-firm sorting in equilibrium, and how creditallocationacrossfirmsrespondstochangesinbanks’fundingcostattheextensive(firm selection) and intensive (interest rate) margins. I show that the the model delivers two sharp empirical predictions on bank-firm sorting and international transmission of funding shocks on credit allocation across firms. 4.1 Equilibrium Bank-Firm Sorting To build intuition, I focus on bank-firm sorting in a symmetric equilibrium where global and local banks face the same funding cost, rG = rL = r. This can be motivated by the idea that both types of banks have access to funds from a global interbank market that provides an elastic supply of funds at the risk-free interest rate r. This case allows me to focus solely on the implications of the double information asymmetry on bank-firm sorting. Appendix I.B discusses the equilibrium bank-firm sorting in the general case when there is variation between the funding costs of global and local banks (rG (cid:54)= rL). Given the assumption rG = rL = r, the expected profit functions of the global and local banksbecomecompletelysymmetric. Withperfectcompetitionandfreeentry,theequilibrium thresholds also become symmetric. Lemma 1. (Thresholds: Symmetric Case) If rG = rL, then z¯L(zG) = zG and z¯G(zL) = zL. i i i i Given Lemma 1, sorting between firms and global versus local banks immediately follows. Corollary 2. (Bank-Firm Sorting: Symmetric Case) Let rG = rL. A firm selects a global bank if and only if zG ≥ zL. A firm selects a local bank if and only if zL > zG. i i i i Panel (a) of Figure 5 provides a simple illustration of bank-firm sorting for the symmetric case. Global and local banks compete for loans borrowed by firms with zG ∈ [zG,1] and i zL ∈ [zL,1]. In equilibrium, the thresholds form a 45 degree line that segments the credit i market. Firms in Region L, which have zL > zG, select into local banks, and firms in Region i i G, which have zG ≥ zL, select into global banks. i i Corollary 2 reveals that the information asymmetry problem faced by global and local banks creates a segmented credit market affected by double adverse selection. Both types of banks are adversely selected against, as firms select into borrowing from the bank which observesthemorefavorablecomponentoftheirriskexposure. Specifically,firmswithaweaker local component (zL) relative to their global component (zG) select into a global bank — the i i bank that cannot observe the weaker component. 22

Furthermore, firms are borrowing at higher interest rates relative to the first-best outcomes. This is because banks, given the information asymmetry problem, can only assign interest rates contingent on the component of firms’ risk exposure that they observe, but not on the unobserved component, for which their rates must be uniform, as shown by the isointerest rate curves in Panel (b) of Figure 5. Knowing the firm selection process, they assign interest rates based on the expected risk of the firms which will approach them. This gives risetoheterogeneityamongfirmsinthedegreetowhichtheyarechargedhigherinterestrates relative to the first-best outcomes. The firms that are riskier in their unobserved exposure component face more favorable interest rates, and firms with relatively balanced global and localriskexposure(i.e., closertothethresholds)facemoreadverseinterestrates. Specifically, consider firms a and b in Panel (a) of Figure 5. In this economy, both firms select into borrowing from a global bank in equilibrium, and are offered the same interest rate RG(zG) since i their zG component is the same. However, the zL component of firm a is much stronger than i i that of b, which means that firm a faces a worse outcome relative to the first-best outcome. 4.2 Bank Funding Cost Shock Next, I proceed to study how the equilibrium credit allocation responds to changes in banks’ funding cost (e.g., a change in monetary policy of the home country of one of the banks) at both the extensive (firm selection) and intensive (interest rate) margins. In addition, I apply the model to clarify the forces underlying the international risk-taking channel of monetary policy. The following corollary summarizes the effects of a shock to banks’ funding cost on the thresholds and the equilibrium interest rates. Corollary 3. (Funding Shock) Holding all else constant, 1. z¯L(zG) is decreasing in rG and increasing in rL; z¯G(zL) is decreasing in rL and increasi i ing in rG. 2. RG(zG) is increasing in rG and decreasing in rL; RL(zL) is increasing in rL and dei i creasing in rG. To expand on its intuition and implications, I describe the results from Corollary 3 in the context of a decrease in global banks’ funding cost, e.g., a decrease in funding rate due to expansionary monetary policy in the home country of the global banks. The effects of a lower funding cost, rG, are also illustrated in Figure 6, which is based on simulation results with parameter values rG = 1.015, rG(cid:48) = 1.005, and rL = 1.040, where rG(cid:48) denotes the new gross funding rate for global banks. 23

Figure 5: Bank-Firm Sorting and Interest Rates Under Symmetric Equilibrium (a) (b) 1 1 L L . . B B z i L d c . G z i L G a . b zL zL A C A C zG 1 zG 1 zG zG i i RG(zG) RL(zL) i i Increasing Notes. Panel (a) shows the equilibrium bank-firm sorting when rG = rL. Panel (b) shows iso-interest rate curvesbyglobalbanksandlocalbanks. Forbothplots,RegionAdepictstheregionwherenoloansaregiven. Region B depicts the region where only local bank loans are given and no global banks would give loans. Region C depicts the region where only global bank loans are given and no local banks would give loans. RegionLdepictstheregionwherebothglobalandlocalbankcompeteforloans,andloansaregivenbylocal banks in equilibrium. Region G depicts the region where both global and local bank compete for loans, and loans are given by global banks in equilibrium. 24

Extensive Margin Effects. A decrease in global banks’ funding costs lowers the equilibrium interest rates offered by global banks for all firms. Based on Part 4 of Proposition 2, z¯L(zG) increases, and z¯G(zL) decreases, which implies that a set of marginal firms switch i i from local banks to global banks. The changes in the thresholds are illustrated in Panel (a) of Figure 6. It is interesting to point out that the marginal firms that switch into global banks arelessriskythantheinfra-marginalfirmsthatcontinuetoborrowfromeitherthelocalbanks or the global banks. Moreover, the funding cost change affects zG and zL, the cut-offs on zG i i i and zL below which global and local banks, respectively, would not make loans. A set of risky i firms that initially were not able to get loans from either bank can now get loans from global banks (firms in Region G(cid:48)), while a set of firms that initially were getting loans from local 2 banks are no longer able to borrow from either class of bank (firms in Region G(cid:48)). 3 Thisresultshowsthatashocktobankfundingcostaffectscreditallocationattheextensive margin. Specifically, the model predicts that firms near the thresholds, which are firms with relatively balanced global and local risk exposure components, are more likely to switch into contracting with global banks. The result is driven by adverse selection: since the firms with relatively balanced global and local risk exposure are more adversely selected against, they are more likely to switch lenders given any changes in the credit market. Intensive Margin Effects. Changes in bank funding cost also affect credit allocation at the intensive margin. Given a decline in rG, for each value of zL, the zG components of i i the marginal firms that switch away from local banks are higher than those of all the inframarginal firms that remain with the local banks. Since the local banks are left with a riskier pool of firms, they charge higher interest rates, despite no changes to their funding cost. This pointstoaspillovereffect, onethatissolelydrivenbyanexacerbationoftheadverseselection problem. Simulation results show that, given a 100 basis point decrease in rG (specifically a decreasefromrG = 1.015torG(cid:48) = 1.005), theinterestratesthatlocalbanksoffertotheinframarginal firms that continue to borrow from them increase by 126 basis points on average, as shown in the red region in Panel (b) of Figure 6. From the global banks’ perspective, the zL components of the marginal firms that switch i into them are higher than those of all the infra-marginal firms that were getting loans from them in the initial equilibrium, conditional on zG. Since the pool of firms that borrows from i global banks is less risky given the funding cost shock, they lower RG(zG). In other words, i the interest rates of the infra-marginal firms that remain with the global banks are expected to decrease by an amount more than that caused by the decrease in global banks’ funding cost, reflectinganamplificationeffect. Theimpactofthefundingshockispositivelyamplified for those infra-marginal firms because firm switching alleviates the initial adverse selection 25

Figure 6: Effects of a Positive rG Shock (rG lowers) (a) Equilibrium Characterization (b) Rate Change: Infra-marginal Firms 1 z¯G(zi L) L G1′ zL G i z¯L(zi G) zL′ zL G3′ G2′ pre post 0 zG′zG 1 zG i (c) Rate Change: Marginal Firms (d) Rate Change: Marginal Firms (zoomed in) Notes. Simulations based on parameter values rG =1.015, rG(cid:48) =1.005, and rL =1.040. Panel (a) Illustrates the equilibrium characterization before and after a decrease in rG. Panel (b) shows ∆R = Rpost−Rpre for i i i the infra-marginal firms. Panel (c) shows ∆R =(Rpost−Rpre)/(Rpre−1) for the marginal firms. Panel (d) i i i i shows a zoomed-in version of part (c) of this figure. 26

problem for the global banks. Simulation results show that a decrease of 100 basis points in rG translates to a decrease of 180 basis points for an average infra-marginal firm that borrows from global banks, as shown in the blue region in Panel (b) of Figure 6. Panels (c) and (d) of Figure 6 illustrate the change in interest rate for the marginal firms thatswitchbanksgiventhefundingcostshock(firmsinRegionG(cid:48) inpanel(a)oftheFigure). 1 The effects are heterogenous across the firms: while interest rates decrease for the switching firms that are closer to initial threshold; rates increase for firms closer to new threshold. Nevertheless, those firms would have been worse off if there were frictions to switching that left them with the local banks. 4.3 DiscussionofDoubleAdverseSelectionChannelofInternationalTransmission Altogether, the analysis of the effects of a funding cost shock on credit allocation reveals a double adverse selection channel of international transmission.18 It results from the key ingredient in the model: competitive interactions between banks with differing specialization in global versus local information. One of factors that can affect banks’ funding cost is monetary policy rate changes. When this happens, the model points to a novel adverse selection channel of international monetary policy transmission through bank lending, one that is distinct from the channels discussed in the existing literature, including currency mismatches on global banks’ balance sheets (?, Ongena et al. 2017, Bräuning and Ivashina 2018) and internal capital markets within global banks (Cetorelli and Goldberg 2012a). One channel of international monetary policy transmission that has received much attention in recent years is the risk-taking channel. Papers, including Bruno and Shin (2015a) and CoimbraandRey(2017),arguethatlowinternationalmonetarypolicyratesandQEcouldinduceglobalbankstoreachforyieldandtakeonexcessrisk. Inparticular,Moraisetal.(2018), usingloan-leveldata, showthatlowmonetarypolicyratesandQEindevelopedeconomiesled global banks in Mexico to increase credit supply to firms charged higher-than-average ex-ante interest rates (riskier firms). They consider this result as evidence for risk-taking behavior by global banks. Tobetterunderstandtheforcesunderlyingtheirresult, Iimplementtheempiricalexercise in Morais et al. (2018) in my model using numerical simulation and examine whether bank risk-taking is indeed the main driving force. Following their procedure, I first categorize 18 APPENDIX B provides additional discussion on how double adverse selection in globalized financial markets sheds new light on the effects of banks’ funding shocks on banks portfolio riskiness and the benefits and costs of financial integration. 27

each firm in the model into a high-risk group and a low-risk group based on whether the firm’s ex-ante rate is above or below the average interest rate in the credit market in the initial equilibrium. I then examine, given a decline in global banks’ funding cost due to expansionary monetary policy in their home country, whether it is the high-risk firms that receive more loans from the global banks. The specific parameter values I use for the simulation are rG = 1.015, rG(cid:48) = 1.005, and rL = 1.040, where the change in rG reflects the decline in monetary policy rate in developed economiesinthepost-globalfinancialcrisisperiodandrL reflectstheaveragemonetarypolicy rate in Mexico over the period. Panel (c) of Figure 6 shows a line pinpointing the firm with the average ex-ante interest rate in that parameter space. As shown, the set of marginal firms that switch into global banks in response to the funding cost change are firms in the high-risk group. Therefore, this model recovers the result that Morais et al. (2018) find in the paper, predicting that an expansionary monetary policy in the home country of the global banks leads to a higher supply of credit to high-risk firms in the local economy. However, in contrast to their explanation, in my model the driving force for the result is substitution between global banking credit and local banking credit. 5 Empirical Predictions and Data 5.1 Mapping Theory to Empirics The model delivers two sharp empirical predictions on bank-firm sorting and international transmission of funding shocks on credit allocation across firms: Prediction 1: Conditional on funding cost differences between global and local banks, global banks lend more to firms with higher return due to global risk relative to local risk, and local banks lend more to firms with higher return due to local risk relative to global risk. Prediction 2a: A shock to the funding cost of one type of bank induces the segment of firms with relatively balanced global and local risk components (i.e., the marginal firms near the thresholds z¯L(zG) and z¯G(zL)) to switch to borrowing from the other type of bank. i i Prediction 2b: Given a decrease in global banks’ funding cost, the interest rates of the infra-marginalfirmsthatremainwiththelocalbanksareexpectedtoincrease(spillovereffect). The interest rates of the infra-marginal firms that remain with the global banks are expected to decrease by more than the direct effect due to the decrease in funding cost (amplification effect). The effects on interest rates of the marginal firms that switch banks are ambiguous. I proceed to test these predictions in the subsequent sections. 28

5.2 Data and Summary Statistics. The main data source for the empirical analysis is syndicated corporate loans from Loan Pricing Corporation’s Dealscan database.19 Syndicated loans are extended by a group of banks to a borrower under a single loan contract. Within each group of lenders, the “lead arranger” isthebankthatestablishesarelationshipwiththeborrowingfirm, negotiatesterms of the contract, and guarantees a loan amount for a price range. It then turns to “participant” lenders that fund part of the loan.20 Ivashina and Scharfstein (2010) report that syndicated loan exposures represent about a quarter of total commercial and industrial loan exposures on US banks’ balance sheets, and about a third for large US and foreign banks. De Haas and Van Horen (2013) note that syndicated loans are a key source of cross-border funding for firms from both advanced and emerging market countries. For the purpose of this study, the ideal dataset is one that encompasses the universe of loans to firms that genuinely have access to both global and local banking credit, which are likely to be firms above a certain threshold in size. The global syndicated loans are viewed as a proxy of that universe of loans. Despite potential selection issues, syndicated loans are uniquely appropriate for this study because they capture a significant portion of cross-border lending, which would not be captured by other loan datasets such as credit registry data. In the Dealscan data, there is detailed information on each loan contract, including terms of the loans at origination (interest rate, whether or not the loan is secured, the maturity of the loan), the type of loan (e.g., line of credit versus term loan), the purpose of the loan, the size of the loan, and the contract activation and ending dates. The dataset also contains information on the name of the borrowers and lenders as well as the country of syndication. Using the names of the borrowers, I hand-match the Dealscan data with international firmlevel databases including Orbis, Amedeus, Compustat, and Compustat Global to extract firm balance sheet data.21 I further implement a series of data-cleaning procedures to correct for basic reporting mistakes, including dropping firm-year observations that have missing information on total assets and operating revenues, dropping firms with negative total assets or employment in any year, and dropping firm-year observations with missing information regardingtheirindustryofactivity. Finally,Ialsoexcludefirmsinfinancialindustries,identified 19AllofthecommerciallylicenseddatausedinthispaperwereobtainedunderthepurviewoftheUniversity of California at Berkeley licenses, while the author was a student of the university. 20 See Sufi (2007) and Ivashina (2009) for more background description of syndicated loans. 21 TheAmadeusandOrbisdatasetsaremainlyusedtoextractinformationonEuropeanandothernon-US firms, including private firms. Compustat is used to extract information on US firms. A well-known problem in the Orbis and Amadeus dataset is that key variables, such as employment and materials, are missing once the data are downloaded. I follow the data collection process described in Kalemli-Ozcan et al. (2015) to maximize the coverage of firms and variables for the sample. Specifically I merge data across historical disks instead of downloading historical data all at once from the WRDS website. 29

by SIC codes 60 through 64 from the sample. For the purpose of this empirical analysis, one of the key variables needed is one that identifies whether the lender of each loan is a global bank or a local bank. To this end, I categorize the lead lender(s) of each loan as global or local. The focus is on the lead bank(s) of each loan contract because they are the entities that are responsible for due diligence prior to loan syndication, while the participant banks rely on the information collected by the lead banks (Ivashina and Scharfstein 2010).22 The bank categorization is based on the following criteria: 1. Local banks: a lender is categorized as a local bank if the corresponding loan is not a cross-border loan, i.e., the borrower of the loan operates in the country where the lender resides. This includes local subsidiaries of foreign banks.23 2. Global banks: • Method 1: a lender is categorized as a global bank if it is considered a globally systemicallyimportantbank(G-SIB),orifthecorrespondingloanisacross-border loan. • Method 2: a lender is categorized as a global bank if the corresponding loan is a cross-border loan. The resulting sample encompasses 115,166 loans, borrowed by 12,979 firms across 24 countries, in the period 2004-2017. Table 2 presents the summary statistics on the loan counts and firm counts for each country in the sample, with the loan counts decomposed into the share given by global banks and that given by local banks, based on Method 1 of the categorization criteria for global banks.24 The majority of the countries in the sample are developed economies, where most global banking activities take place. For most of the countries, theloansaresplitrelativelyevenlybetweenglobalbankingcreditandlocalbanking credit. Table 3 presents the summary statistics on a set of firm balance sheet variables. All the variables in the table are in billions of dollars, except for age and employment. Value added, 22 For loans that involve multiple lead banks of which some are global banks and some are local banks, I consideraloanisgivenbyglobalbankif≥50%ofthelendersareglobalbanks. Thesecasesmakeuparound 20% of the loans. Based on the model predictions, I conjecture that firms with relatively balanced global and local risk components are more likely to get loans that involve both global and local lead banks. I find empirical evidence that supports this conjecture. 23 E.g.,forfirmsinGermany,JPMorganHoldingDeutschlandisalocalbank,whileJPMorganChaseUSA isaglobalbank. Localsubsidiariesareconsideredseparatelegalentitiesfromtheirparentbank,incorporated in host countries and supervised by the host regulator. 24 Table B.1 in APPENDIX C presents summary statistics on the same variables as Table 2 but with the banks categorized based on Method 2 of the categorization criteria for global banks. 30

Table 2: Summary Statistics: Loan and Firm Count by Country (Method 1) Country Loan GB LB Firm Country Loan GB LB Firm Australia 4507 0.70 0.30 701 Japan 21341 0.45 0.55 2865 Austria 387 0.53 0.47 61 Mexico 601 0.70 0.30 137 Belgium 704 0.69 0.31 123 Netherlands 2028 0.54 0.46 406 Canada 6760 0.64 0.36 903 New Zealand 1023 0.70 0.30 127 Czech Republic 197 0.68 0.32 77 Norway 1017 0.66 0.34 253 Denmark 327 0.56 0.44 84 Poland 318 0.54 0.46 87 Finland 587 0.65 0.35 113 Portugal 254 0.65 0.35 64 France 5876 0.67 0.33 996 Spain 4380 0.68 0.32 839 Germany 5987 0.68 0.32 942 Sweden 875 0.66 0.34 190 Greece 309 0.66 0.34 47 Switzerland 790 0.69 0.31 175 Ireland 404 0.70 0.30 107 United Kingdom 6810 0.69 0.31 1528 Italy 2378 0.67 0.33 688 United States 46732 0.70 0.30 1466 Notes. Sample constructed from Dealscan, Amadeus, Orbis, Compustat, Compustat Global, and author’s calculation. Sample period covers the year 2004-2017. wage bill, total assets, and exporter revenue are deflated with gross output price indices with a base year of 2017. I first calculate the means and standard deviations of each variable across firms in each given year and country without weighting across firms. Entries in the table denote the means and standard deviations averaged across all years and countries. The summary statistics exhibit significant variation in each variable in the sample, which shows that the sample contains firms from a wide distribution of asset size and age. For all variables except exporter revenue, there does not seem to be a significant difference between the firms that borrow from global banks and the ones that borrow from local banks. On the other hand, it seems that firms that borrow from global banks export significantly more than firms that borrow from local banks. 6 Empirical Analysis I: Double Adverse Selection in Credit Allocation In this section, I test whether the bank-firm sorting patterns predicted by the model are consistent with the observed patterns in the data (model Prediction 1). To that end, I follow an empirical strategy that tightly maps to the model set-up. Methodology. In order to test whether global banks lend more to firms with higher return due to global risk (zG) relative to local risk (zL), and vice versa for local banks, I need to i i 31

Table 3: Summary Statistics: Firm Characteristics by Bank Type Global Bank Local Bank Mean SD Mean SD Value Added 512.55 1256.45 468.55 895.09 Age 25.29 24.67 25.23 24.98 Employees 1657.34 6073.34 1719.58 5326.32 Wage Bill 209.73 1030.76 163.35 786.41 Working Capital 110.58 1089.56 123.34 1173.32 Fixed Asset 918.03 2465.80 732.76 2987.84 Total Assets 1344.5 4658.56 1134.53 4034.32 Exporter Revenue 587.00 1789.34 113.31 456.68 Notes. Value added is constructed as the difference between operating revenue and materials with negative valuesdropped. Ageofthefirmiscalculatedasthedifferencebetweentheyearofthebalancesheetinformation and the year of firm incorporation plus one. Except for age and employment, all entries in the table are in billions of dollars. Value added, wage bill, total assets, and exporter revenue are deflated with gross output price indices with a base year of 2017. I first calculate the means and standard deviations without weighting across firms for each year in each country. Entries in the table denote the means and standard deviations averaged across all years and countries. Data from Amadeus, Orbis, Compustat, and Compustat Global. Sample period covers the year 2004-2017. construct measures for zG and zL for each firm in the sample. Recall from the model that the i i production function for each firm is z = zG+zL+u . I take that as a simplified version of a i i i i typical Cobb-Douglas production function Y = z KγL1−γ, where there is one unit of K and i i i i i L . The parameter z , in turn, can be interpreted as a firm revenue productivity measure that i i capturestotalexposuretobothproductivityanddemandrisk, andzG (zL)canbeinterpreted i i as total exposure to global (local) productivity and demand risk. I start by estimating a time-varying revenue productivity measure z for each firm in each it year based on the method of Solow growth accounting.25 Specifically, I compute the z based it on the following equation: logz = log(Y /L )−γ log(K /L ) (4) it it it t it it where Y denotes nominal value added divided by the 2-digit industry-level output price it deflatorforeachcountry,wherevalueaddedisconstructedasthedifferencebetweenoperating revenue and material costs with negative values dropped, L denotes the wage bill divided it 25 Gorodnichenko (2012) shows that this can be used as a robust non-parametric method to estimate productivity. Healsopointsoutthatanumberofexistingparametricmethodsforestimatingproductivityare misspecified or poorly identified. In particular, inversion/control-function estimators (e.g., Olley and Pakes 1996, Levinsohn and Petrin 2003) can lead to inconsistent estimates because they ignore variation in factor prices. GMM/IVestimatorsusinglagsofendogenousvariablesasinstruments(e.g.,BlundellandBond1998) can be poorly identified because of economic restrictions on the comovement of inputs and output. 32

by the same output price deflator, K denotes fixed assets divided by the aggregate price of it investment goods, and the factor share γ uses country-specific and industry-specific shares t extracted from the National Accounts of each country. Appendix Figure B.3 plots the estimates of the productivity measure, log z , averaged it acrossfirmsandtimebycountry. Asexpected,averageproductivityishigherfortherelatively more developed economies such as the US and high-income European economies. NextIdecomposethefirm-specificproductivitymeasure,z ,whichcapturestotalexposure i to productivity and demand risk, into two components: exposure to global risk (zG) and i exposuretolocalrisk(zL). Firms’totalexposuretoglobalriskcanbeconsideredtoencompass i twocomponents,zG = βGzG,whereβG denotesfirmi’sexposuretoglobalriskandzG denotes i i i global risk. The same applies to firms’ total exposure to local risk: zL = βLzL, where βL i i i denotes firm i’s exposure to local risk and zL denotes local risk. I implement a principal component analysis to extract estimates for zG and zL, following Stock and Watson (2002). Specifically, I estimate the following equation: z =βGzG+βLzL +u (5) ict ic t ic ct ict where z is the productivity measure for firm i in country c in year t, zG is the global factor, ict t zL is the local factor in country c, and u is a firm-specific component. ct ict The factors can be estimated consistently with a two-step procedure. In the first step, the common global factor is obtained from the principal components of the z series across the ict 24 countries in the sample. The first principal component explains 58% of the total variance, which I take as the global factor, zG. Appendix Figure B.4 plots the global factor.26 As t shown, it declines around 2007-2008, the period of the global financial crisis, and gradually recovers thereafter. In the second step, I orthogonalize the global component by regressing the productivity measures z on the global factor and taking the residuals. I then extract local (country) ict factors by computing the principal components based on the residualized z series for each ict country. The first principal component from output for each country is taken as the local factor, zL. Finally, I estimate the firm-specific global and local exposure measures using ct OLS regressions. βG and βL are extracted from the loadings on the global and local factor, i i respectively. Results. Using the estimated measures for zG and zL, I proceed to test the first model i i prediction on bank-firm sorting. Similar to the procedure I used to test the traditional theory 26 TomapcloselytothemodelsetupwherezG andzL onlytakepositivevalues,thefactorvalueshavebeen i i adjusted upward by their minimum so that all the values are positive. 33

Figure 7: Bank-Firm Sorting, by zG/zL Quartile (Method 1) i i 1 8. 6. 4. 2. 0 zG/zL i i 1 2 3 4 Global Bank Local Bank Notes. The plot shows sorting patterns between firms and global versus local banks, with firms sorted into quartiles by their exposure to global versus local risk (zG/zL), uses variables that are constructed based on i i Method1ofthebankcategorizationcriteriaforglobalbanks. Datasampleconsistsofsyndicatedloansbetween firmsglobalandlocalbanksandfirmsacross24countriesfrom2004-2017. Source: Dealscan,Amadeus,Orbis, Compustat, Compustat Global, and author’s calculation. on bank-firm sorting in Section 2 but now using the new measures, I sort firms into quartiles based on the distribution of firm exposure to global versus local risk (zG/zL) in each year by i i country, and calculate the proportion of loans given by global banks and local banks in each quartile. Figure 7 plots the resulting distribution of lending from global and local banks over the entire sample. The plot shows a stark pattern of bank-firm sorting: global banks lend more to firms with higher return given global risk (zG) relative to local risk (zL), and local i i banks lend more to firms with higher return due to local risk relative to global risk.27 As before, I further test whether the differences between global and local banks illustrated in Figure 7 and B.5 are statistically significant. For the measure on firm exposure to global versus local risk (zG/zL), I test whether the value-weighted mean of that variable for global i i banks is different from that for local banks. Table 4 presents these means and their differences. The results confirm the graphical analysis: the differences in value-weighted means are statistically significant between global and local banks for the measure of firm exposure to global versus local risk (zG/zL), supporting the model prediction on bank-firm sorting. i i 27 Figure B.5 is a parallel version of Figure 7, with the banks categorized based on Method 2 of the bank categorization criteria for global banks. 34

Table 4: Bank-Firm Sorting, by zG/zL Quartile: Statistical Test i i Method 1 Method 2 z /z z /z iG iL iG iL Mean: Global Bank 2.905*** 3.382*** (0.046) (0.040) Mean: Local Bank 2.107*** 2.507*** (0.113) (0.097) Difference 0.798*** 0.875*** ( 0.122) (0.105) Observations 98,345 98,345 Notes. The dependent variable in each regression (Y) is the measure of firm exposure to global versus local risk,(zG/zL),coded1-4basedonthequartilenumbertowhicheachrespectivefirmbelong. Notethefirmsare i i sortedbasedontheexposuremeasureeveryyearbycountry. Row1androw2showthemeansforeachvariable forglobalbanksandlocalbanks,respectively,byrunningavalue-weightedregressionofYonaconstant. For differencesinmeansofthetwotypesofbanks,thewholedataisusedintheregressionandadummyforglobal banks is added (row 3). Standard errors reported in parentheses are clustered at the bank-level. Results in column 1 and column 2 are based on the banks categorized using Method 1 and Method 2, respectively, of thebankcategorizationcriteriaforglobalbanks. Source: Dealscan,Amadeus,Orbis,Compustat,Compustat Global, and author’s calculation. The results show that the new perspective I raise in this paper, bank specialization in global versus local information, plays an important role in determining bank-firm sorting in financial systems with both global and local banks. But does the traditional theory of bank specialization in hard versus soft information still play a role? I investigate this question by studying how the measures that capture global information and the measures that capture hard information jointly predict the likelihood of getting loans from global banks. I run a set ofregressionswiththedependentvariablebeingadummyvariablethattakesthevalue1ifthe loan is given by a global bank and 0 otherwise. The independent variables in the regressions are firm exposure to global risk relative to local risk (zG/zL), firm asset size, and/or firm i i age, each coded by the quartile number to which each observation of the respective variable belongs. TheresultsarepresentedinTable5. Resultsincolumn1showthatbetweenfirmsin twoconsecutivequartilesbasedonthemeasureofexposuretoglobalriskrelativetolocalrisk, the firms in the higher quartile group are 33% more likely to get loans from a global bank. Columns 2 and 3 present results from regressions that include firm asset size and firm age, respectively. Theresultsshowthat,controllingforfirmexposuretoglobalriskrelativetolocal risk, firms that are larger and more established are significantly more likely to get loans from global banks, which is consistent with the predictions from the traditional banking theory. The results in column 4 show that each of the three measures still have predictive power on the likelihood of getting loans from global banks, even when the other two measures are 35

also included as regressors. Overall, these results suggest that the bank-firm sorting patterns predictedbythetraditionalbankingtheorycanberecoveredoncebankspecializationinglobal and local information are taken into account. Table 5: Bank-Firm Sorting, Traditional Theory and New Perspective (1) (2) (3) (4) 1(GB) 1(GB) 1(GB) 1(GB) zG/zL 0.329*** 0.221*** 0.261*** 0.198** i i (0.086) (0.074) (0.080) (0.081) Size 0.268*** 0.236*** (0.081) (0.073) Age 0.157** 0.138* (0.075) (0.078) Industry FE Yes Yes Yes Yes Country FE Yes Yes Yes Yes Observations 98,345 98,345 98,345 98,345 Notes. Resultsfromregressionswiththedependentvariablesbeingadummyvariablethattakesthevalue1if theloanisgivenbyaglobalbankand0otherwise. Theindependentvariablesarefirmexposuretoglobalrisk relative to local risk (zG/zL), firm asset size, and/or firm age, each coded by the quartile number to which i i each observation of the respective variable belongs. Each regression controls for industry and country fixed effects. Standard errors reported in parentheses are clustered at the firm level. Source: Dealscan, Amadeus, Orbis, Compustat, Compustat Global, and author’s calculation. Finally, I explore the characteristics of the firms that borrow from global banks, and the characteristics of the loans are given by global banks. For the former, I study if exporters are more likely to have a higher value of zG/zL and thereby more likely to get loans from i i global banks. I run a firm-level panel regression with zG/zL as the dependent variable, and a i i dummy variable that takes the value 1 if the exporting revenue for the respective firm for a given year is nonzero and 0 otherwise as the main regressor, controlling for time and country fixed effects. The results, reported in column 1 of Table 6, show that exporting firms tend to have significantly higher zG/zL values, or higher exposure to global risk relative to local risk. i i Combined with the results from the sorting exercises, this empirical evidence suggests that exporters are more likely to get loans from global banks. In light of these evidence, I further investigate into the loan-level data to see whether loans of specific purposes such as trade finance are more likely to be funded by global banks. Irunaloan-levelregressionswiththemainregressorsbeingdummiesonspecificloanpurposes, includingprojectfinance,workingcapital,tradefinance,andothers28. Thedependentvariable 28 Others include IPO related finance, real estate, stock buyback, etc. They are grouped together in one 36

of the regression is a dummy variable that takes the value 1 if the loan is given by a global bank and 0 otherwise. The results (column 2 of Table 6) show that it is not the case that global banks mainly finance loans for the purpose of trade finance. A significant portion of the loans they finance are for general project finance and working capital. Table 6: Determinants of zG/zL and Global Banking Credit i i (1) (2) zG/zL 1(GB) i i Exporter 0.565*** (0.103) Project purpose Project finance 0.013*** (0.001) Working capital 0.020*** (0.001) Trade finance 0.004** (0.002) Year FE Yes Yes Industry FE Yes Yes Country FE Yes Yes Observations 129,309 98,345 Notes. Column1reportsresultsfromafirm-levelpanelregressionwithzG/zL asthedependentvariable,and i i a dummy variable that takes the value 1 if the exporting revenue for the respective firm for a given year is nonzero and 0 otherwise as the main regressor. Column 2 reports results from a loan-level regression with a dummyvariablethattakesthevalue1iftheloanisgivenbyaglobalbankand0otherwiseasthedependent variable, and dummy variables on loan purpose as the main regressors. Time, industry and country fixed effects are included in both regressions. Standard errors reported in parentheses are clustered at the firm level. Source: Dealscan, Amadeus, Orbis, Computstat, Compustat Global, and author’s calculation. 7 Empirical Analysis II: A Double Adverse Selection Channel of International Transmission In this section, I study how shocks to bank funding cost, specifically monetary policy shocks, affect credit allocation at the extensive and intensive margins, testing model Predictions 2. I take the Euro area as the empirical laboratory of this study, and analyze how US and Euro area monetary policy, through US and Euro area banks, respectively, affect credit allocation across firms in the Euro area. From the perspective of Euro area firms, US banks are global banks, and Euro area banks are local banks. Given this context, I raise two conjectures based variable. 37

on the model predictions and the results on bank-firm sorting from the last section: i)ConditionalonEuroareamonetarypolicy,anexpansionaryUSmonetarypolicyinduces firms in the Euro area with relatively balanced global and local risk components—firms in the second tercile of the zG/zL distribution—to switch their borrowing from Euro area banks to i i US banks. ii)ConditionalonEuroareamonetarypolicyandgivenexpansionaryUSmonetarypolicy, the interest rates of the infra-marginal firms that continue to borrow from Euro area banks— firms in the first tercile of the zG/zL distribution (firms with relatively low zG relative to i i i zL)—are expected to increase (spillover effect). The interest rates of the infra-marginal firms i that continue to borrow from US banks—firms in the third tercile of the zG/zL distribution i i (firms with relatively high zG relative to zL)—are expected to decrease by more than the i i direct effect due to expansionary US monetary policy (amplification effect). The effects on interestratesofthemarginalfirmsthatswitchbanks—firmsinthesecondtercileofthezG/zL i i distribution—are ambiguous. To test the conjectures, I perform regressions of the following form, using data on loans borrowed by Euro area firms in the loan-level data and the firm-specific zG/zL measure: i i 3 3 3 ∆Y = (cid:88) βq(∆USR x Tq )+ (cid:88) δq(∆EUR x Tq )+ (cid:88) γqTq +ν +σ +(cid:15) (6) it t it−1 t it−1 it−1 i t it q=1 q=1 q=2 where i indexes firm, t indexes the date on which a specific loan is issued, ∆(.) denotes the differenceinthereferredvariablebetweenthedateonwhichthecurrentloanisissuedandthe date on which the last loan was issued, Y denotes the applicable dependent variable which I explain below, USR denotes US monetary policy shocks, EUR denotes Euro area monetary policy shocks, q indexes each of the three terciles of the zG/zL distribution, Tq are dummy i i it−1 variables that take the value 1 when firm i’s zG/zL measure at the time of the last loan i i issuance belongs to tercile q and 0 otherwise, ν are firm dummies, and σ are year dummies. i t The standard errors are clustered by time, to take into consideration potential correlations across firms in borrowing behavior or borrowing term changes since the monetary policy shocks are aggregate. For measures of US and Euro area monetary policy shocks, I use intraday data on the Federal Funds 30-day futures contracts and the three-month Euribor futures contracts, respectively, from Gorodnichenko and Weber (2016) and CQG Data Factory.29 The Federal 29 The US monetary policy shock measure based on intraday data on the Federal Funds futures contracts has been used in a number of papers, including Kuttner (2001), Cochrane and Piazzesi (2002), Rigobon and Sack (2004), Gertler and Karadi (2015), Gorodnichenko and Weber (2016), Nakamura and Steinsson (2018), 38

Funds futures data is based on trading on the Chicago Board of Trade (CBOT) Globex electronic trading platform. It reflects the market expectation of the average effective Federal Funds rate during that month. The Euribor futures rates is based on trading on ICE Futures Europe and reflects the market expectation of the Euribor rate for three-month Euro deposits.30 Therefore, both series provide a market-based measure of the anticipated path of the monetary policy rates for the respective region. In order to identify exogenous shocks to US and Euro area monetary policy, the monetary policy shocks are calculated as changes in the futures rates within a time window around the Federal Open Market Committee (FOMC) or European Central Bank (ECB) monetary policy announcements.31 The identifying assumption is that changes in the interest rate futures within the specified windows around the announcements only reflect market responses to the monetary policy news, not changes in other domestic or foreign economic conditions. For measures of US monetary policy shocks, I consider a window of 60 minutes around the announcements that starts 15 minutes (∆−) prior to the event, following Gorodnichenko and Weber (2016) and Nakamura and Steinsson (2018). As for ECB monetary policy, its key target rate decision since 2001 has been announced at13:45CETthroughapressrelease, followedbyapressconferenceat14:30pmCET.Atthe press conference, the ECB President and Vice-President discuss the future path of monetary policy and announce any additional non-conventional measures.32 To give a sense of how the ECB policy rate announcement and the press conference affect the market expectation of the Euribor rate, I illustrate the three-month Euribor futures rate in high frequency on two specific announcement dates in Figure B.6. The upper panel plots the Euribor futures rate from 08:00 to 18:00 CET for April 6, 2006. At 13:45 CET, the ECB announced through a press release that it is keeping the target rate unchanged. Since this decision was expected by the market, the futures rate did not exhibit significant change around the press release time. But it decreased sharply during the press conference window. This is because, contrary to market expectation of an interest rate hike later in the year, Jean-Claude Trichet told the and Wong (2018). The Euro area monetary policy measure based on the three-month Euribor futures has been used in papers including Bernoth and Hagen (2004), Rosa and Verga (2008), and Ranaldo and Rossi (2010). They show that the three-month Euribor futures rate is an unbiased predictor of Euro area policy rate changes. 30 Tobemorespecific,thethree-monthEuriborfutureisacommitmenttoengageinathree-monthloanor depositofafacevalueof1,000,000Euros. Futurespricesarequotedonadailybasis. Therearefourdelivery dates during a year, namely the third Wednesday of March, June, September and December. 31 I obtain the dates of the FOMC meetings from the Federal Reserve Board website at http://www.federalreserve.gov/monetarypolicy/fomccalendars.htm, and those of the ECB meetings from the ECBwebsiteathttps://www.ecb.europa.eu/press/govcdec/mopo. Ialsoverifytheexacttimesofthemonetary policy announcements using the first news article about them on Bloomberg. 32 SeeRosaandVerga(2008)foradescriptionoftheinstitutionalfeaturesuniquetoECBmonetarypolicy announcements. 39

press that “the current suggestions regarding the high probability of an increase of rates in our next meeting do not correspond to the present sentiment of the Governing Council.” The decline in Euribor futures rate during the press conference time window thus reflect market’s revision of its expectations. The bottom panel of Figure B.6 plots the Euribor futures rate for November 3, 2011, when the ECB unexpectedly cut interest rates by 25bps for the first time in two years. The sharp decline in the Euribor futures rate around the time of the press release reflect the change in market expectation. Given the unique institution features of ECB monetary policy announcements, I apply a window of 120 minutes that starts 10 minutes (∆−) prior to the press release and ends 10 minutes (∆+) after the press conference to construct measures of ECB monetary policy shocks. Furthermore, I consider two measures of monetary policy shocks for each region: a current period shock based on current month futures (mp1), and a long-term path shock based on three-month-ahead futures (mp4). The long-term path shock is aimed at capturing any persistent effects of current period shocks on long-term investment, which can occur when the current period shocks change expectations about the future path of monetary policy rates. The shock measures take the general form: mp = (fx −fx ) (7) t t+∆+ t−∆− where t is the time when the FOMC or ECB issues an announcement, f is the Federal t+∆+ FundsfuturesortheEuriborfutures∆+ minutesaftert,f istheFederalFundsfuturesor t−∆− the Euribor futures ∆− minutes before t, and x denotes either 1 for current month futures or 4 for three-month-ahead futures. For the US current monetary policy shock measure (mp1), Equation (7) is adjusted by the term D , where D is the number of days in the month. This D−t is because the Federal Funds futures settle on the average effective overnight Federal Funds rate. Iaggregateuptheidentifiedshockstoobtainmonthlymeasuresofmonetarypolicyshocks, following Cochrane and Piazzesi (2002). I use the monetary policy measures from the month prior to the loan dates (t) when estimating Equation (6), to ensure time consistency. Extensive Margin To analyze how monetary policy shocks affect credit allocation across firmsintheEuroareaattheextensivemargin,IestimateEquation(6)withthethedependent variable being the change in a dummy variable that takes the value 1 if the loan is given by a US bank and 0 if the loan is given by a Euro area bank between two consecutive loans for each given firm i (denoted as ∆USB ). The main coefficients of interest are βq and δq. I it conjecture β2 to be negative, and δ2 to be positive, since, based on the model prediction, contractionary US monetary policy would induce firms in the second tercile of the zG/zL i i 40

distribution to switch away from US banks, and contractionary ECB monetary policy would induce firms in the second tercile of the zG/zL distribution to switch into US banks. All the i i specifications include firm fixed effects to account for potential demand-driven explanations for changes in the trends of firms’ borrowing behavior, as well as time fixed effects to control for common shocks. Table 7 reports the regression results. Columns 1 and 3 show the average effects of the US and Euro area monetary policy shocks, based on measures of mp1 and mp4, respectively, on the firms’ switching behavior. Results in Column 1 show that, on average, a 25-basis-point shock to the current US monetary policy rates decreases the probability of firm switching from a Euro area bank into a US bank by 3.4 percentage points, while a 25-basis-point shock to the Euro area monetary policy rates increases the probability of a firm switching from a Euro area bank into a US bank by 4.1 percentage points. The effects are larger and more significant when considering shocks to the path of monetary policy rates. Results in Column 3 show that, on average, a 25-basis-point shock to the path of US monetary policy rates decreases the probability of firm switching into a US bank by 5.2 percentage points, while such shock to the path of Euro area monetary policy rates increases the probability of a firm switching into a US bank by 5.3 percentage points.The coefficients are statistically significant at the 5% level. The findings point to evidence of firm switching in the Euro area in response to monetary policy shocks on average. In particular, firms respond slightly more to domestic monetary policy shocks. Turning to the coefficients of interest, columns 2 and 4 in Table 7 show the estimations of how these effects vary for firms in different terciles of the zG/zL distribution (Equation i i (6)). Across both specifications, the effects of US and Euro area monetary policy shocks on the probability of firm switching are around two times larger in the second tercile of the zG/zL distribution than the other terciles, and highly significant. The point estimates of i i β2 imply that a 25-basis-point shock to the current and long-term US monetary policy rate decreases the probability of firm switching into a US bank by 6.0 and 7.6 percentage points, respectively,forfirmsinthesecondtercileofthezG/zL distribution. Forthosefirms,thepoint i i estimates of δ2 imply that a 25-basis-point shock to the Euro area monetary policy increases the probability of firm switching into a US bank by 6.6-8.5 percentage points. The effects are again larger when considering shocks to the path of monetary policy rates, suggesting that firm investments respond more to changing expectations about the future path of monetary policy rates. The results for the other two terciles are mostly statistically insignificant. Overall, the results suggest that most of the firm switching effects are concentrated in the second tercile of the zG/zL distribution, where firms have relatively balanced exposure to i i global risk relative to local risk. This evidence supports the model prediction on the effects 41

of bank funding shocks on credit allocation across firms at the extensive margin. Intensive Margin Next, I turn to analyzing how monetary policy shocks affect credit allocation across firms in the Euro area at the intensive (interest rate) margin. I implement Equation(6)withthedependentvariablebeingthechangeintheinterestratespreadbetween two consecutive loans for each given firm i (denoted as ∆R ).33 The spread describes the it amount the borrower pays in basis points over the LIBOR. The main coefficients of interest are again βq and δq. The model predicts that, conditional on Euro area (US) monetary policy and given contractionary US (Euro area) monetary policy, the interest rates of the infra-marginal firms that continue to borrow from Euro area (US) banks decrease, reflecting a (positive) spillover effect. Thus, β1 (which summarizes the group of firms that are more likely to be borrowing from Euro area banks) and δ3 (which summarizes the group of firms that are more likely to be borrowing from US banks) are conjectured to be negative. The model also predicts that, under the above scenario, the interest rate spreads of the infra-marginal firms that continue to borrow from US (Euro area) banks increase, reflecting a (negative) amplification effect. Thus, β3 and δ1 are conjectured to be positive. Since these predictions are based on the assumption that there is stronger pass-through from US monetary policy to the interest rates offered by US banks, and similarly Euro area monetary policy to Euro area banks, I first perform a series of regressions to validate these assumptions. Columns 1, 2, 4 and 5 in Table 8 report the results from regressions of changes in firm interest rate spreads (∆R ) on changes in US and Euro area monetary policy shocks it (∆USR and ∆EUR, respectively), a dummy variable that takes the value 1 for US or Euro area banks and 0 otherwise (1(USB) or 1(EUB)), and interactions of these two variables: eitheraninteractionbetweentheUSmonetarypolicyshockandtheUSbankdummyvariable (USR∗1(USB)), or one between the Euro area monetary policy shock and the Euro area bank dummy variable (∆EUR∗1(EUB)). The results confirm the assumption. Columns 1 and 4 show that a 25-basis-point shock to the current and long-term US monetary policy rate disproportionately increases the interest rate spread charged by US banks, by around 25 and 33 basis points, respectively, on average relative to other banks. Results in columns 2 and 5 show that a 25-basis-point shock to the current and long-term Euro area monetary policy rate disproportionately increases the interest rate spread charged by Euro area banks, by 34 and 37 basis points, respectively, on average relative to other banks. Turning to the coefficients of interest, columns 3 and 6 in Table 8 report the results of how theeffectsofmonetarypolicyshocksoninterestratespreadsvaryforfirmsindifferentterciles 33 To make the interest rate spreads as comparable as possible, the type of loan facilities (e.g., revolving line, bank term loan, and institutional term loan) between two consecutive loans are matched. 42

ofthezG/zL distribution(Equation(6)). Aspredicted, thecoefficientsβ1 andδ3 arenegative i i acrossallspecifications. Specifically,a25-basis-pointshocktothecurrentUSmonetarypolicy rate decreases the interest rate spread for the infra-marginal firms that continue to borrow from Euro area banks by 22 basis points, while such a shock to the Euro area monetary policy rate decreases the interest rate spread for the infra-marginal firms that continue to borrow from US banks by 25 basis points. The effects are larger and more significant when considering shocks to the path of monetary policy rates (column 6). A 25-basis-point shock to the long-term US (Euro area) monetary policy rate decreases the interest rate spread for the infra-marginal firms that continue to borrow from Euro area (US) banks by 27 (32) basis points. These results point to a (positive) spillover effect. Furthermore, the coefficients β3 and δ1 are positive across all specifications, as predicted, andhighlystatisticallysignificant. Specifically, a25-basis-pointshocktothecurrentUSmonetary policy rate increases the interest rate spread for the infra-marginal firms that continue to borrow from US banks by 25 basis points. The effect increases to 32 basis points given a 25-basis-point shock to the path of US monetary policy rate. Similarly, a 25-basis-point shock tothecurrentandlong-termEuroareamonetarypolicyrateincreasestheinterestratespread for the infra-marginal firms that continue to borrow from Euro area banks by 34 and 40 basis points, respectively. These results point to a (negative) amplification effect. Furthermore, the effects on interest rates of the firms in the second tercile of the zG/zL distribution, which, i i based on the results from Table 7, is mostly comprised of marginal firms that may switch banks, are ambiguous, as predicted. Overall, the results in Table 8 support the model prediction on the effects of bank funding shocks on credit allocation across firms at the intensive margin. Combined with the results on the extensive margin effects, they point to evidence of a novel adverse selection channel of monetary policy transmission. 43

Table 7: Monetary Policy Shocks and Credit Allocation: Extensive Margin (1) (2) (3) (4) mp1 mp1 mp4 mp4 ∆USR -0.134* -0.209** (0.071) (0.083) ∆EUR 0.164** 0.211** (0.074) (0.089) ∆USR∗T1 -0.049 -0.054 (0.119) (0.128) ∆USR∗T2 -0.241** -0.302** (0.120) (0.131) ∆USR∗T3 -0.117 -0.163 (0.118) (0.127) ∆EUR∗T1 0.057 0.062 (0.118) (0.137) ∆EUR∗T2 0.264** 0.339*** (0.118) (0.135) ∆EUR∗T3 0.173 0.220* (0.116) (0.127) Firm FE Yes Yes Yes Yes Time FE Yes Yes Yes Yes Observations 11,454 11,454 11,454 11,454 R-squared 0.067 0.068 0.067 0.068 Notes. Regressions with the dependent variable being the change in a dummy variable that takes the value 1 iftheloanisgivenbyaUSbankand0iftheloanisgivenbyaEuroareabankbetweentwoconsecutiveloans for each given firm i (denoted as ∆USB). USR denotes US monetary policy shocks, and EUR denotes Euro area monetary policy shocks. Tq is a dummy variable that takes the value 1 when the firm’s zG/zL measure i i at the time of the last loan issuance belongs to tercile q and 0 otherwise. For the specifications in columns 1 and 2, the monetary policy measures used are current period shocks constructed from current month futures (mp1). For the specifications in columns 3 and 4, the monetary policy measures used are long-term path shocks constructed from three-month-ahead futures (mp4). Year and firm fixed effects are included in all specifications. Standard errors reported in parentheses are clustered by time. Source: Dealscan, Amadeus, Orbis,Computstat,CompustatGlobal,andauthor’scalculation. Significanceatthe1percent,5percent,and 10 percent levels is indicated by ***, **, and *, respectively. 44

Table 8: Monetary Policy Shocks and Credit Allocation: Intensive Margin (1) (2) (3) (4) (5) (6) mp1 mp1 mp1 mp4 mp4 mp4 ∆USR∗1(USB) 98.543** 132.458*** (43.765) (47.986) ∆EUR∗1(EUB) 136.633*** 147.375*** (42.543) (49.864) ∆USR∗T1 -89.354* -108.564* (48.542) (54.875) ∆USR∗T2 62.796 78.342 (52.769) (60.875) ∆USR∗T3 98.427** 126.653** (46.293) (58.975) ∆EUR∗T1 136.864** 158.539*** (56.249) (57.986) ∆EUR∗T2 76.563 83.457 (52.087) (59.357) ∆EUR∗T3 -101.876* -127.978** (54.681) (54.975) Firm FE Yes Yes Yes Yes Yes Yes Time FE Yes Yes Yes Yes Yes Yes Observations 3,367 3,367 3,367 3,367 3,367 3,367 R-squared 0.051 0.051 0.052 0.051 0.051 0.052 Notes. Regressions with the dependent variable being the change in the interest rate spread between two consecutive loans for each given firm i (denoted as ∆R). USR denotes US monetary policy shocks, and EUR denotes Euro area monetary policy shocks. 1(USB) and 1(EUB) are dummy variables that takes the value 1 for US and Euro area banks, respectively, and 0 otherwise. Tq is a dummy variable that takes the value 1 when the firm’s zG/zL measure at the time of the last loan issuance belongs to tercile q and 0 otherwise. For the specifications in columns 1-3, the i i monetary policy measures used are current period shocks constructed from current month futures (mp1). For the specifications in columns 4-6, the monetarypolicymeasuresusedarelong-termpathshocksconstructedfromthree-month-aheadfutures(mp4). Thespecificationsincolumn1and4 includeUSR,1(USB),andEURasregressors. Thespecificationsincolumn2and5includeUSR,1(EUB),andEURasregressors. Yearandfirm fixed effects are included in all specifications. Standard errors reported in parentheses are clustered by time. Source: Dealscan, Amadeus, Orbis, Computstat,CompustatGlobal,andauthor’scalculation. Significanceatthe1percent,5percent,and10percentlevelsisindicatedby***,**,and *, respectively. 45

8 Conclusion The rise of global banking has transformed financial systems and corporate financing across the world over the past two decades. This paper provides a new theory on the mechanism driving credit allocation in globalized financial systems, and tests it using cross-country loanlevel data. I show that bank specialization in global versus local information—information on global versus local risk factors—plays a key role in determining bank-firm sorting and credit allocation in financial systems with both global and local banks. I point out that that the traditional theory of bank specialization in hard or soft information is insufficient to explain observed sorting patterns between firms and banks in globalized financial systems, revealing a puzzle in the mechanism driving global banking credit. In light of the puzzle, I develop a model of banking in which there are global and local banks, and firms that have return dependent on exposure to global and local risk. Each bank faces a problem of asymmetric information: global banks have the technology to extract information on global risk factors but not local risk factors, and vice versa for local banks. The model shows that this double information asymmetry creates a segmented credit market affected by double adverse selection: banks are adversely selected against by firm selection, as firms select into borrowing from the bank which observes the more favorable component of their risk exposure. I further apply the model to analyze the macroeconomic implications of the adverse selectionproblem, studyingtheimpactoncreditallocationoffundingshockstobanks. Themodel demonstrates that, given a monetary policy shock, adverse selection affects credit allocation at both the extensive and intensive margins. It induces firms with relatively balanced global and local risk components to switch banks, and generates spillover and amplification effects through adverse interest rates. I test the model using a cross-country bank-firm loan-level dataset matched with firm balance sheet data. I find bank-firm sorting patterns, and evidence of firm switching behavior and interest rate changes given US and Euro area monetary policy shocks, that support the model predictions. The results point to a new double adverse selection channel of international monetary policy transmission. Overall, the evidence substantiates that bank specialization in global versus local information is a key mechanism driving credit allocation in globalized banking systems. This mechanism has potentially important policy implications. Relative to the traditional view that firms and banks sort based on hard versus soft information, this new mechanism suggests that global banks’ balance sheet may be more loaded on global risk than previously thought, since firms with returns more dependent on global risk are more likely to select into borrowing from them. This, in turn, calls for considerations from policy-makers for bank 46

regulations on exposure limits and macroprudential policies. References Albuquerque, R., G. H. Bauer, and M. Schneider (2009). Global private information in international equity markets. Journal of Financial Economics 94(1), 18–46. Aoki, K., G. Benigno, and N. Kiyotaki (2016). Monetary and financial policies in emerging markets. Working Paper. Bacchetta, P. and E. Van Wincoop (2006). Can information heterogeneity explain the exchange rate determination puzzle? American Economic Review 96(3), 552–576. Baskaya, Y., J. Giovanni, S. Kalemli-Ozcan, and M. Ulu (2021). International spillovers and local credit cycles. Forthcoming. Beck, T., A. Demirguc-Kunt, and M. S. M. Peria (2007). Reaching out: Access to and use of banking services across countries. Journal of Financial Economics 85, 234–266. Beck, T. and M. S. M. Peria (2010). Foreign bank participation and outreach: Evidence from Mexico. Journal of Financial Intermediation 19(1), 52–73. Berger, A., N. Miller, M. Petersen, R. Rajan, and J. Stein (2005). Does function follow organizationalform? Evidence from the lending practicesof large and small banks. Journal of Financial economics 76(2), 237–269. Berger, A. N., L. F. Klapper, and G. F. Udell (2001). The ability of banks to lend to informationally opaque small businesses. Journal of Banking and Finance 25, 2127–2167. Bernanke, B. S. and A. Blinder (1992). The federal funds rate and the channels of monetary transmission. The American Economic Review 82(4), 901–921. Bernanke, B. S. and M. Gertler (1995). Inside the black box: the credit channel of monetary policy transmission. Journal of Economic perspectives 9(4), 27–48. Bernoth, K. and J. v. Hagen (2004). The euribor futures market: Efficiency and the impact of ecb policy announcements. International finance 7(1), 1–24. Blundell, R. and S. Bond (1998). Initial conditions and moment restrictions in dynamic panel data models. Journal of econometrics 87(1), 115–143. Borio, C. and H. Zhu (2012). Capital regulation, risk-taking and monetary policy: A missing link in the transmission mechanism? Journal of Financial stability 8(4), 236–251. Boyd,J.H.andE.C.Prescott(1986). Financialintermediary-coalitions. Journal of Economic theory 38(2), 211–232. Bräuning, F. and V. Ivashina (2018). US monetary policy and emerging market credit cycles. Working Paper. 47

Bräuning, F. and V. Ivashina (2020). Monetary policy and global banking. The Journal of Finance 75(6), 3055–3095. Bruno, V. and H. S. Shin (2015a). Capital flows and the risk-taking channel of monetary policy. Journal of Monetary Economics 71, 119–132. Bruno, V. and H. S. Shin (2015b). Cross-border banking and global liquidity. Review of Economic Studies 82(2), 535–564. Campbell, T. S. and W. A. Kracaw (1980). Information production, market signalling and the theory of financial intermediation. The Journal of Finance 35(4), 863–882. Cetorelli, N. and L. Goldberg (2012a). Banking globalization and monetary transmission. Journal of Finance 67(5), 1811–1843. Cetorelli, N. and L. S. Goldberg (2012b). Liquidity management of US global banks: Internal capital markets in the great recession. Journal of International Economics 88(2), 299–311. Claessens, S. (2006). Competitive implications of cross-border banking. In G. Caprio, D. D. Evanoff, and G. G. Kaufman (Eds.), Cross-Border Banking: Regulatory Challenges, pp. 151–182. Hackensack, NJ: World Scientific Publishing. Claessens, S., A. Demirgüç-Kunt, and H. Huizinga (2001). How does foreign entry affect domestic banking markets? Journal of Banking and Finance 25(5), 891–911. Clarke, G. R., R. J. Cull, M. S. M. Peria, and S. M. Sanchez (2005). Bank lending to small businesses in Latin America: Does bank origin matter? Journal of Money, Credit and Banking 37(1), 83–118. Cochrane, J. H. and M. Piazzesi (2002). The fed and interest rates—a high-frequency identification. American Economic Review 92(2), 90–95. Coimbra, N. and H. Rey (2017). Financial cycles with heterogeneous intermediaries. Working Paper. De Blas, B. and K. N. Russ (2013). All banks great, small and global: Loan pricing and foreign competition. International Review of Economics & Finance 26, 4–24. De Haas, R. and I. V. Lelyveld (2014). Multinational banks and the global financial crisis: Weathering the perfect storm? Journal of Money, Credit and Banking 46(1), 333–364. De Haas, R. and N. Van Horen (2013). Running for the exit? International bank lending during a financial crisis. The Review of Financial Studies 26(1), 244–285. De Meza, D. and D. C. Webb (1987). Too much investment: A problem of asymmetric information. Quarterly Journal of Economics 102(2), 281–292. Dedola, L., P. Karadi, and G. Lombardo (2013). Global implications of national unconventional policies. Journal of Monetary Economics 60, 66–85. 48

Dell’Ariccia, G. and R. Marquez (2004). Information and bank credit allocation. Journal of Financial Economics 72(1), 185–214. Detragiache, E., T. Tressel, and P. Gupta (2008). Foreign banks in poor countries: theory and evidence. Journal of Finance 63(5), 2123–2160. Diamond, D.(1984). Financialintermediationanddelegatedmonitoring. Review of Economic Studies 51(3), 393–414. Edison, H. J., R. Levine, L. Ricci, and T. Sløk (2002). International financial integration and economic growth,. Journal of International Money and Finance 21, 749–776. Gertler, M. and P. Karadi (2015). Monetary policy surprises, credit costs, and economic activity. American Economic Journal: Macroeconomics 7(1), 44–76. Gopinath, G. and J. C. Stein (2018). Banking, trade and the making of a dominant currency. Working Paper. Gormley, T.A. (2010). The impact offoreignbank entryin emergingmarkets: Evidence from India. Journal of Financial Intermediation 19(1), 26–51. Gormley, T. A. (2014). Costly information, entry and credit access. Journal of Economic Theory 154(633-667). Gorodnichenko, Y.(2012). Usingfirmoptimizationtoevaluateandestimateproductivityand returns to scale. Working Paper. Gorodnichenko, Y. and M. Weber (2016). Are sticky prices costly? Evidence from the stock market. American Economic Review 106(1), 165–99. Ivashina, V. (2009). Asymmetric information effects on loan spreads. Journal of financial Economics 92(2), 300–319. Ivashina, V. and D. Scharfstein (2010). Bank lending during the financial crisis of 2008. Journal of Financial Economics 97(3), 319–338. Ivashina,V.,D.S.Scharfstein,andJ.C.Stein(2015). Dollarfundingandthelendingbehavior of global banks. Quarterly Journal of Economics 130(3), 1241–1281. Kalemli-Ozcan, S., B. Sorensen, C. Villegas-Sanchez, V. Volosovych, and S. Yesiltas (2015). HowtoconstructnationallyrepresentativefirmleveldatafromtheORBISglobaldatabase. Working Paper. Kashyap, A. K. and J. C. Stein (2000). What do a million observations on banks say about the transmission of monetary policy? American Economic Review 90(3), 407–428. Kuttner, K. N. (2001). Monetary policy surprises and interest rates: Evidence from the fed funds futures market. Journal of monetary economics 47(3), 523–544. 49

Levine, R. (1996). Foreign banks, financial development and economic growth. In C. E. Barfield (Ed.), International Financial Markets: Harmonization versus Competition, pp. 224–254. Washington, DC: AEI Press. Levinsohn, J. and A. Petrin (2003). Estimating production functions using inputs to control for unobservables. Review of Economic Studies 70(2), 317–342. Liberti, J. M. and M. A. Petersen (2018). Information: Hard and soft. Working Paper. Maggiori, M., B. Neiman, and J. Schreger (2018). International currencies and capital allocation. Working Paper. Martin, A. and F. Taddei (2013). International capital flows and credit market imperfections: A tale of two frictions. Journal of International Economics 89(2), 441–452. Mian, A. (2006). Distance constraints: The limits of foreign lending in poor economies. Journal of Finance 61(3), 1465–1505. Miranda-Agrippino, S. and H. Rey (2018). US monetary policy and the global financial cycle. Working Paper. Morais, B., J.-L. Peydró, J. Roldan, and C. Ruiz (2018). The international bank lending channel of monetary policy rates and quantitative easing: credit supply, reach-for-yield and real effects. Journal of Finance, Forthcoming. Nakamura, E. and J. Steinsson (2018). High frequency identification of monetary nonneutrality: The information effect. Quarterly Journal of Economics 133(3), 1283–1330. Niepmann, F. (2015). Banking across borders. Journal of International Economics 96(2), 244–265. Olley,S.andA.Pakes(1996). Dynamicsofproductivityinthetelecommunicationsequipment. Econometrica 64(6), 1263–1297. Ongena, S., I. Schindele, and D. Vonnák. (2017). In lands of foreign currency credit, bank lending channels run through? Working Paper. Petersen, M. and R. G. Rajan (1994). The benefits of lending relationships: Evidence from small business data. Journal of Finance 49(1), 3–37. Petersen, M. A. and R. G. Rajan (2002). Does distance still matter? The information revolution in small business lending. The journal of Finance 57(6), 2533–2570. Popov, A.andG.F.Udell(2012). Cross-borderbanking, creditaccessandthefinancialcrisis. Journal of international economics 87(1), 147–161. Ramakrishnan, R. TS, and A. V. Thakor (1984). Information reliability and a theory of financial intermediation. Review of Economic Studies 51(3), 415–432. 50

Ranaldo, A. and E. Rossi (2010). The reaction of asset markets to swiss national bank communication. Journal of International Money and Finance 29(3), 486–503. Rey, H. (2016). International channels of transmission of monetary policy and the Mundellian Trilemma. IMF Economic Review 64(1), 6–35. Rigobon, R. and B. Sack (2004). The impact of monetary policy on asset prices. Journal of Monetary Economics 51(8), 1553–1575. Rosa, C. and G. Verga (2008). The impact of central bank announcements on asset prices in real time. International Journal of Central Banking 4(2), 175–217. Schnabl, P. (2012). The international transmission of bank liquidity shocks: Evidence from an emerging market. Journal of Finance 67(3), 897–932. Sengupta, R. (2007). Foreign entry and bank competition. Journal of Financial Economics 84(2), 502–528. Stein, J. (2002). Information production and capital allocation: Decentralized versus hierarchical firms. Journal of Finance 57(5), 1891–1921. Stiglitz, J. E. and A. Weiss (1981). Credit rationing in markets with imperfect information. American Economics Review 71(3), 393–410. Stock, J. H. and M. W. Watson (2002). Macroeconomic forecasting using diffusion indexes. Journal of Business & Economic Statistics 20(2), 147–162. Sufi, A. (2007). Information asymmetry and financing arrangements: Evidence from syndicated loans. The Journal of Finance 62(2), 629–668. Takáts, E. and J. Temesvary (2020). The currency dimension of the bank lending channel in international monetary transmission. Journal of International Economics, 103309. Wong, A. (2018). Transmission of monetary policy to consumption and population aging. Working Paper. 51

APPENDIX A MODEL SOLUTION AND PROOFS I.A Equilibrium Solution Propositions 1–3 lead to a full characterization of the equilibrium solution on RG and RL. Basedonthesecharacterizations,IsolvefortheequilibriuminterestratesRG(zG)andRL(zL), i i and thresholds z¯L = z¯L(zG) and z¯G = z¯G(zL), for zG ∈ [zG,1] and zL ∈ [zL,1] as follows. i i i i i i First, letRG(zG,z¯L)andRL(zL,z¯G)betheimplicitfunctionswhichgivetherateatwhich i i i i each banks’ expected profit (Equation (3a) and (3b)) would be zero for a given observed component combined with a given threshold on the unobserved component34: RG(zG,z¯L) = RG(zG) s.t. E [π (zG,z¯L,RG(zG))] = 0; i i i G G i i i RL(zL,z¯G) = RL(zL) s.t. E [π (zL,z¯G,RL(zL))] = 0. i i i L L i i i Based on Proposition 2, for each given zG, the corresponding threshold z¯L is the zL for i i i the firm (zG,zL) for which RL(zL)=RG(zG). By symmetry, z¯G(z¯L) = zG. Therefore, the i i i i i i equilibrium rate RG(zG) and threshold z¯L are the solutions to the system of equations: i i RG(zG,z¯L) = RL(z¯L,zG). i i i i (A.1) RG(zG) = RG(zG,z¯L). i i i Similarly, for each given zL, the equilibrium rate RL(zL) and threshold z¯G are the solutions i i i to the system of equations: RL(zL,z¯G) = RG(z¯G,zL). i i i i (A.2) RL(zL) = RL(zL,z¯G). i i i Furthermore, I apply Proposition 2 to solve for zG and zL, the cut-offs below which the expected profits of the firms are too low for the global bank and local bank to break even in expectation, regardless of the rate charged. At these cut-off points, the maximum expected profits of the banks are zero, all firms default given the equilibrium interest rates. The next lemma establishes that the cut-offs zG and zL are thresholds to each other. Lemma 2. zG = z¯G(zL), and zL = z¯L(zG). Given Lemma 2, zG and zG are the solutions to the system of equations: zL 1 (cid:90) (cid:90) (zG+zL+u ) dF (u ,zL) = rG; i i i zL i i 0 0 zG 1 (cid:90) (cid:90) (zG+zL+u ) dF (u ,zG) = rL. i i i zG i i 0 0 where F (.) and F (.) denote the cumulative distribution function of the relevant variable zL zG 34 The implicit equations are fully written out in the appendix as Equations (A.5a) and (A.5b). 52

conditional on zL ≤ zL and zG ≤ zG, respectively. The solutions to this system is: i i 1 1 zG = (4rG−2rL−1) and zL = (4rL−2rG−1). (A.3) 3 3 The bounds zG and zL define the cut-offs on zG and zL, respectively, below which global i i banks and local banks would not make loans. They are increasing in the banks’ own funding cost and decreasing in the funding cost faced by the other bank type. In other words, facing higher funding cost induces the respective banks to be more restrictive on the riskiest firm to which they lend, while higher funding cost faced by the other bank type induces them to lend to riskier firms. Interestingly, each banks’ own funding cost has a stronger effect on the respective lower bound than the other banks’ funding cost. Appendix Figure A.1 illustrates the cut-offs zG and zL in a space that summarizes all the firms in the economy. Given the cut-offs, firms in Region A are not offered loans. Firms in Region B can only receive loans from local banks, and firms in Region C can only receive loans from global banks. Figure A.1: Firm Space 1 zL B i zL A C zG 1 zG i Notes. The plot summarizes all the firms in the economy. The bounds zG and zL define the cut-offs below which global banks and local banks, respectively, would not make loans. Firms in Region A are not offered loans. Firms in Region B can only receive loans from local banks. Firms in Region C can only receive loans from global banks. I.B Model Equilibrium Asymmetric Equilibrium. I solve the model numerically to study firm-bank sorting in the general case when there is variation between the funding costs of global and local banks (rG (cid:54)= rL). Panel (a) of Figure A.2 provides an illustration of the equilibrium when rG < rL, where 53

rG = 1.00andrL = 1.01. Comparedtothesymmetriccase, globalbanksareabletocapturea greater share of the loan market given their funding advantage. In particular, they are able to attract all the firms with zG > z˜G, and they provide loans to firms with lower zG components i i than before, since the cut-off zG is increasing in rG (Equation (A.3)). Panel (b) of Figure A.2 illustrates the equilibrium when rG > rL, where rL = 1.00 and rG = 1.01. The results are analogous. Figure A.2: Firm-Bank Sorting under Asymmetric Equilibrium (a) (b) 1 1 L L z˜L B B z i L G z i L G zL zL A C A C zG z˜G 1 zG 1 zG zG i i rG=rL rG<rL rG=rL rG>rL Notes. Panel (a) illustrates the firm-bank sorting when rG <rL, where rG =1.00 and rL =1.01. Panel (b) illustrates the firm-bank sorting when rG > rL, where rL = 1.00 and rG = 1.01. For both plots, Region A depictstheregionwherenoloansaregiven. RegionB depictstheregionwhereonlylocalbankloansaregiven and no global banks would give loans. Region C depicts the region where only global bank loans are given andnolocalbankswouldgiveloans. RegionLdepictstheregionwherebothglobalandlocalbankscompete for loans, and loans are given by local banks in equilibrium. Region G depicts the region where both global and local banks compete for loans, and loans are given by global banks in equilibrium. 54

I.C Proofs Proof of Proposition 1. Based on Equations (2a) and (3a), RG(zG) is given implicitly by i the global bank’s expected profit function: (cid:20)(cid:90) (cid:18)(cid:90) (cid:90) (cid:19) (cid:21) E [π (zG)] = (zG+zL+u ) dF(u )+ RG(zG) dF(u ) dF(zL) −rG = 0 G G i i i i i i i i Gc Ga G b (cid:26) (cid:27) where G a = u i (cid:12) (cid:12) 0 ≤ u i < min(max(0,RG(z i G)−z i G−z i L),1) (cid:26) (cid:27) G b = u i (cid:12) (cid:12) min(max(0,RG(z i G)−z i G−z i L),1) ≤ u i ≤ 1 (cid:26) (cid:27) G c = z i L (cid:12) (cid:12) z i L: (z i G,z i L) ∈ S G (A.4) Equation (A.4) can be decomposed into two regions over zG: i 1. No loans: zG such that zG+E [zL | (zG,zL) ∈ SG]+1/2 < rG. i i G i i i 2. Loans: zG such that zG+E [zL | (zG,zL) ∈ SG]+1/2 ≥ rG. i i G i i i Equilibrium rates RG(zG) are defined in region 2. i Analyzing ∂EG[πG(z i G)] : An increase in zG lowers the probability of default and increases ∂zG i i the bank’s expected return. Thus ∂EG[πG(z i G)] > 0 ∀zG. ∂zG i i Given that, I first prove that RG is weakly decreasing in zG. Assume otherwise: there i exists zG > zG such that RG(zG) > RG(zG). Given perfect competition with free entry, j i j i E[π (zG)] = 0 for RG(zG). Because ∂EG[πG(z i G)] > 0, another global bank could offer at most G i i ∂zG i the same RG(zG) for zG and at least break even. Therefore, it could offer RG(zG) ≤ RG(zG), i j j i which is a contradiction. RL is similarly weakly decreasing in zL. i Analyzing ∂EG[z i L|(z i G,z i L)∈SG] : An increase in the rate RG(zG) may cause some marginal ∂RG(zG) i i values of zL to switch from selecting the global to the local bank. Since both RG(zG) and i i RL(zL) are non-increasing, those that do will be those with the lowest RL(zL) and therefore i i the highest zL, lowering the expected value of zL over firms which select the global bank. i i Therefore, ∂EG[z i L|(z i G,z i L)∈SG] ≤ 0. ∂RG(zG) i Analyzing ∂EG[πG(z i G)] : An increase in RG(zG) drives the expected return to the global ∂RG(zG) i i bank through two effects: 1. It increases the return in all outcomes where previously there was no default. 2. It decreases the expected value of zL for firms which will select the global bank, which i decreases the expected return in case of default. 55

Absent other constraints, at any point, ∂EG[πG(z i G)] could be dominated by either term and ∂RG(zG) i be positive, negative, or zero. Now I prove that RG is strictly decreasing in zG (where loans are made, in region 2). i Assume otherwise: there exists zG > zG such that RG(zG) ≥ RG(zG). Consider again j i j i the perfect competition and free entry among global banks. E [π (zG)] = 0 for RG(zG). G G i i Because ∂EG[πG(z i G)] > 0, if RG(zG) = RG(zG) there would be excess profit: E [π (zG)] > 0. ∂zG j i G G j i Regardless of the sign of ∂EG[πG(z i G)] , another bank could charge a lower rate RG(zG) without ∂RG(zG) j i losing money in expectation: • If ∂EG[πG(z i G)] ≤ 0, decreasing the rate would leave profit unchanged or increased and ∂RG(zG) i clearly be possible. • If ∂EG[πG(z i G)] > 0, a competing global bank could trade the excess profit to offer a lower ∂RG(zG) i rate and capture the market while still at least breaking even. Therefore RG(zG) < RG(zG), which is a contradiction. j i The proof that RL is strictly decreasing in zL is entirely analogous. i Further analysis. Consider the two effects which drive ∂EG[πG(z i G)] . The first is trivially ∂RG(zG) i continuous. The second is continuous because RL being strictly decreasing means that differential changes in RG(zG) cannot have discontinuous effects on selection SG. i ConsideralsotheimplicitfunctionofRG(zG)wherethethebankprofitiszero: E [π (zG)] = i G G i 0. By the implicit function theorem, dRG(z i G) = − ∂EG[πG] / ∂EG[πG(z i G)] . We know that dzG ∂zG ∂RG(zG) i i i dRG(z i G) < 0 (RG is strictly decreasing) and ∂EG[πG] > 0. Therefore, ∂EG[πG(z i G)] > 0, and the dzG ∂zG ∂RG(zG) i i i positive profit effect of increasing RG(zG) dominates the negative selection effect. i Finally, considering the regions over zG, the boundary between the two regions occurs i when zG+E [zL | (zG,zL) ∈ SG]+1/2 = rG. Since ∂EG[z i L| (z i G,z i L)∈SG] < 0 and dRG(z i G) < 0, i G i i i ∂RG(zG) dzG i i E [zL | (zG,zL) ∈ SG] is increasing in zG. Therefore there is a unique zG = rG −E [zL | G i i i i i G i (zG,zL) ∈ SG]−1/2. Equilibrium rates RG(zG) are defined for all zG ≤ zG ≤ 1. i i i i i All analyses apply to the analogous terms for local banks. Proof of Proposition 2. 1) In an equilibrium market configuration that supports both types of banks, there must exist a set of marginal firms that are indifferent between the contracts by global banks and local banks, which occur when RG(zG) = RL(zL). Let f(zG,zL) = RG(zG)−RL(zL) = 0. i i i i i i By Proposition 1, ∂f(z i G,z i L) = − ∂RL(z i L) > 0 for zL ∈ [zL,1]. By the implicit function ∂zL ∂zL i i i i 56

theorem, for each zG ∈ [zG,1], there exists a threshold function z¯L: zG (cid:55)→ z¯L, such that i i i i RG(zG) = RL(z¯L). i i The proof on the existence of a threshold function z¯G: zL (cid:55)→ z¯G such that RL(zL) = i i i RG(z¯G) is analogous. i 2) Consider a marginal firm that faces RG(zG) = RL(zL). As zL decreases, RL(zL) i i i i increases by Proposition 1, while RG(zG) remains constant. Since now RL(zL) > RG(zG), i i i those firms would select a global bank. Therefore, firms with zL < z¯L ∈ SG. Conversely, i i as zL increases, RL(zL) decreases by Proposition 1, while RG(zG) remains constant. Since i i i RL(zL) < RG(zG), those firms would select a local bank. Therefore, SG = {(zG,zL) : zL ≤ i i i i i z¯L(zG)}, and SL = {(zG,zL) : zL > z¯L(zG)} i i i i i The proof that SL = {(zG,zL) : zG < z¯G(zL)} and SG = {(zG,zL) : zG ≥ z¯G(zL)} is i i i i i i i i analogous. Proof of Proposition 3. The equilibrium interest rate functions are solution to the bank expected profits equations subject to zero profits conditions and firm selection: (cid:20)(cid:90) (cid:18)(cid:90) (cid:90) (cid:19) (cid:21) E [π (zG)] = (zG+zL+u ) dF(u )+ RG(zG) dF(u ) dF(zL) −rG = 0, G G i i i i i i i i Gc Ga G b (cid:26) (cid:27) where G a = u i (cid:12) (cid:12) 0 ≤ u i < min(max(0,RG(z i G)−z i G−z i L),1) , (cid:26) (cid:27) G b = u i (cid:12) (cid:12) min(max(0,RG(z i G)−z i G−z i L),1) ≤ u i ≤ 1 , (cid:26) (cid:27) G c = z i L (cid:12) (cid:12) 0 < z i L ≤ z¯L(z i G)) ; (A.5a) (cid:20)(cid:90) (cid:18)(cid:90) (cid:90) (cid:19) (cid:21) E [π (zL)] = (zG+zL+u ) dF(u )+ RL(zL) dF(u ) dF(zG) −rL = 0, L L i i i i i i i i Lc La L b (cid:26) (cid:27) where L a = u i (cid:12) (cid:12) 0 ≤ u i < min(max(0,RL(z i L)−z i G−z i L),1) , (cid:26) (cid:27) L b = u i (cid:12) (cid:12) min(max(0,RL(z i L)−z i G−z i L),1) ≤ u i ≤ 1 , (cid:26) (cid:27) L c = z i G (cid:12) (cid:12) 0 < z i G ≤ z¯G(z i L)) . (A.5b) Analyzing ∂EG[πG(z i G)] : An increase in RL(z¯L(zG)) shifts marginal firms from the local to ∂RL(z¯L(zG)) i i global bank at (zG,z¯L(zG)). This increases the threshold value z¯L(zG) at zG. As a result, i i i i the expected profit of the global bank increases, all else held constant, so ∂EG[πG(z i G)] > 0. ∂RL(z¯L(zG)) i 57

The analysis that ∂EG[πG(z i G)] > 0 is outlined in the proof for Proposition 1. ∂RG(zG) i By the implicit function theorem, dRG(z i G) = − ∂EG[πG(z i G)] / ∂EG[πG(z i G)] < 0. dRL(z¯L(zG)) ∂RL(z¯L(zG)) ∂RG(zG) i i i Proof of Lemma 2. At zG, the equilibrium rate RG(zG) is such that all firms which approach global banks default: RG(zG) = zG + z¯L(zG) + 1. Similarly at zL, RL(zL) = z¯G(zL)+zL+1. It is clear that at least one entry zj must be the threshold for the other zk: z¯j(zk) = zj. Without loss of generality, let j = G and k = L: z¯G(zL) = zG. Assume otherwise, z¯L(zG) > zL. Given z¯G(zL) = zG, RL(zL) = zG+zL+1. It follows RG(zG) = zG+z¯L(zG)+ 1 > zG+zL+1 = RL(zL). This implies RL(z¯L(zG)) > RL(zL), which contradicts the strict monotonicity of RL. At the same time, z¯L(zG) < zL is a contradiction, since local banks make no loans to firms with zL < zL by definition. Therefore, zL = z¯L(zG). i The proof that zG = z¯G(zL) is analogous. Proof of Lemma 1. Let rG = rL. The expected profit equations for global banks and local banks subject to the break even conditions and firm selection, given by Equations (A.5a) and (A.5b), respectively, are symmetric. The result that z¯L(zG) = zG and z¯G(zL) = zL follows. i i i i Proof of Corollary 2. Let rG = rL. Assume firm i selects into borrowing from a global bank. Based on firm selection criteria from Equations (2a) and 2b and Assumption 1, RG(zG) ≤ RL(zL), which implies zG ≥ zL by Proposition 1 and Lemma 1. Now assume i i i i zG ≥ zL. Based on Equations (A.5a) and (A.5b), R (zG) ≤ R (zL), which implies firm i i i Gi i Li i selects into borrowing from a global bank. The proof that a firm selects a local bank if and only if zL > zG is analogous. i i 58

APPENDIX B Additional Discussion of Model Implications This section supplements the model discussion in Section 4.3. I apply the theoretical framework to discuss how the double adverse selection in globalized financial markets sheds new light on the effects of banks’ funding shocks on banks portfolio riskiness and the benefits and costs of financial integration. Overall Riskiness in Bank Portfolios. The prior exercise suggests that credit substitution driven by adverse selection is an important effect of monetary policy transmission. Furthermore, it could potentially confound with bank risk-taking behavior. I investigate this issuefurtherbyanalyzinghowafundingshockaffectstheoverallriskinessofbanks’portfolios, and decomposing the overall effect into the changes due to credit substitution and those due to bank risk-taking. Let the riskiness of the portfolio held by a bank j be expressed in terms of the firms’ average output R = (cid:80)n (zG +zL)/n, where j denotes either a global bank or local bank, j i=1 i i and i denotes the firm in the respective bank portfolio. Higher average output R implies j lower risk. IcomputeR beforeandafteradeclineinrG usingnumericalsimulation, andexaminethe j change in R of each bank’s portfolios given the change. Specifically, I run the simulation for j two sets of parameter values for the initial equilibrium. In scenario 1, rG < rL in the initial equilibrium: rG = 1.015, rL = 1.050, andrG(cid:48) = 1.005ex-post. Inscenario2, rG = rL = 1.015 in the initial equilibrium, and rG(cid:48) = 1.005. Table A.1 presents the results. The local bank’s portfoliosbecomeunambiguouslyriskierafterthefundingcostchangeduetonegativespillover effects. On the other hand, the overall riskiness of the global bank’s portfolio may increase or decrease depending on the relationship between rG and rL in the initial equilibrium. In scenario 1, the overall riskiness of the global bank’s portfolio increases given the decline in funding cost. This is due to the risk profiles of both the marginal firms that switch into the global bank and the newly added firms that were too risky to receive loans before (Region G(cid:48) and Region G(cid:48) in Panel (a) of Figure A.3, respectively). The average risk of the firms 1 2 that newly enter the credit market and borrow from the global bank is unambiguously higher than that of the infra-marginal firms that were getting loans from the global bank, driving up the overall riskiness of the global bank’s portfolio. This change can be attributed to bank risk-taking. The marginal firms that switch into borrowing from the global bank, despite having higher zL components conditional on zG, have lower zG components on average—and, i i i asaresult, highercombinedaveragerisk—thanthoseoftheinfra-marginalfirms. Thisfurther drives up the overall riskiness of the global bank’s portfolio, and the driving force is credit substitution. 59

Inscenario2, theoverallriskinessoftheglobalbank’sportfoliolowers. Whiletheriskiness of the firms that newly enter the credit market is still unambiguously higher than that of the infra-marginal firms (Region G(cid:48) in Panel (b) of Figure A.3), the average riskiness of the 2 switching firms is lower. The average riskiness of both the zL and zG components of the i i switching firms are lower than the infra-marginal firms that were initially getting loans from global banks. The risk profile of the marginal firms dominate the risk adjustments in global bank’s portfolio given the change in rG. In other words, the effects due to credit substitution dominate the effects due to bank risk-taking in this scenario. Figure A.3: Effects of a Decline in Funding Cost rG (a) Scenario 1, rG <rL (b) Scenario 2, rG =rL 1 1 L L G1′ G1′ z i L G z i L G zL′ zL′ zL zL G3′ G2′ pre G3′ G2′ pre post post 0 0 zG′zG 1 zG′zG 1 zG zG i i Notes. Panel (a) Illustrates the equilibrium before and after a decline in rG based on simulations with parameter values rG = 1.015, rG(cid:48) = 1.005, and rL = 1.050. Panel (a) Illustrates the equilibrium before and after a decline in rG based on simulations with parameter values rG =1.015, rG(cid:48) =1.005, and rL =1.015. Closed Economy vs. Financial Integration. An interesting counterfactual to consider is how this financially integrated economy compares with the benchmark closed economy, in terms of bank-firm sorting, aggregate credit, and efficiency. In a closed economy where there are only local banks, firms with zL < rL−1 are considered too risky to get loans (illustrated i in Panel (b) of Figure 3). With financial integration, most of those firms, specifically firms with zG > zG, would be able to get loans from global banks (firms in Region n in Panel (a) i of Figure A.4). Furthermore, a set of firms with stronger global components (zG) relative to i their local components (zL) would switch into borrowing from global banks (firms in Region i G in Panel (a) of Figure A.4), since they would receive lower interest rates from global banks, as shown in Panel (b) of the figure. Those firms would all benefit from financial integration. However, the switching of firms leaves local banks with a riskier pool of firms, inducing 60

Figure A.4: Effects of Financial Integration (a) Bank-Firm Sorting (b) Rate Change 1 L zL i zL G e zL =rL−1 CE n zG 1 zG i (c) Efficiency: Closed Economy (d) Efficiency: Financial Integration 1 First-best 1 First-best d d zL =r−1/2 zL =r−1/2 FB FB z i L z i L b zL c b zG =r−1/2 a zL CE =rL−1 a FB c b zG FB =r−1/2 1 zG z i G z i G 1 Notes. Panel (a) illustrates the equilibrium characterization after financial integration. Relative to a closed economy, upon financial integration, firms in Region e are no longer able to get loans, and firms in Region n are able to get loans. Panel (b) shows the interest rate change as measured by ∆R = RFI−RCE upon i i i financialintegration. TheplotisbasedonsimulationsusingparametervaluesrG =1.05andrL =1.05. Panel (c) and (d) compares the firm space in a closed economy and a financially integrated economy, respectively, to that in the benchmark full-information economy. According to the first-best outcome, firms in Regions a and b would not get loans, and firms in Regions c and d would get loans. 61

Table A.1: Banks’ Overall Risk Before and After a Decline in rG Pre Post Switching New Scenario 1 G 1.163 1.157 1.155 0.509 L 0.943 0.917 – – Scenario 2 G 1.087 1.155 1.516 0.508 L 1.155 1.085 – – Notes. The table shows the riskiness of the portfolios held by a global bank (G) and a local bank (L) before (“pre”)andafter(“post”)adeclineinrG. Theposteffectisfurtherdecomposedbyshowingtheriskinessofthe “switching” firms and “new” firms that select into global banks after the change. Riskiness of bank portfolios is measured as R = (cid:80)n (zG+zL)/n, where j denotes either a global bank or local bank, i denotes all the j i=1 i i firms in the respective bank portfolio. The higher the R measure, the lower the risk. In scenario 1, rG <rL j in the initial equilibrium: rG =1.015, rL =1.050, and rG(cid:48) =1.005 ex-post. In scenario 2, rG =rL =1.015 in the initial equilibrium, and rG(cid:48) =1.005. an increase in interest rate for the infra-marginal firms that remain with local banks (firms in Region L in Panel (a) of Figure A.4), as shown in Panel (b) of the Figure. This means that financial integration can give rise to an adverse selection problem. Moreover, this adverse selection problem would force a set of firms to exit the credit market (firms in Region e in Panel (a) of Figure A.4). This result suggests that financial integration can induce a decline in aggregate credit due to adverse selection, which is in line with the arguments raised in Detragiache et al. (2008) and Gormley (2014). Despite the potential decline in aggregate credit, it is important to point out that credit allocationinafullyintegratedfinancialsystemismoreefficientrelativetoaclosedeconomy. I defineefficiencyintermsofhowcloselycreditallocationcorrespondstothatinthebenchmark full-information economy. As shown in Panels (c) and (d) in Figure A.4, in a full information economy, firms in Regions a and b would not get loans, and firms in regions c and d would get loans. In both a closed economy and a financially integrated economy, firms in Region b are overfunded, while firms in Region c are underfunded. Nevertheless, for all reasonable parameters values, Regions b and c in a financially integrated economy are smaller than the corresponding regions in a closed economy. Quantitatively, let efficiency be defined as the share of total credit in the economy relative to the benchmark full-information economy (Efficiency = 1−(b+c)/(a+b+c+d) based on the illustrations Panels (c) and (d) in Figure A.4). Given parameter values rG = 1.05 and rL = 1.05, the closed economy is 85% efficient, while a financially integrated economy is 95% efficient. 62

APPENDIX C ADDITIONAL FIGURES AND TABLES Figure B.1: Global Banking Credit to Private Sector (a) Total Global Banking Credit, All Countries noillirT $ 51 01 5 0 (b) Share in Total Private Credit, by Country 1995q1 2000q1 2005q1 2010q1 2015q1 2020q1 tnecreP 06 04 02 0 Argentin A a ustralia Austri B a elgium Cana C da zec C h h R ile epub D lic enmark Finland Fran G ce ermany Greec H e ungary Ireland Italy M N e e xi t c h o e N rl e a w nd Z s ealand Norway Polan P d ortugal Spai S n w S ed w e it n zerla U n n d T ite u d rk K ey i U ng n d it o e m d States Notes. Panel (a) plots a time-series of total cross-border credit to the non-bank private sector across all BIS reporting countries. Source: BIS Locational Banking Statistics. Panel (b) plots the share of cross-border credit in total private credit, averaged over 2005-2016, for a cross-section of developed and major emerging market economies. Source: BIS Locational Banking Statistics and IMF International Financial Statistics. 63

Figure B.2: Loan Currency Denominations by Global and Local Banks 1 8. 6. 4. 2. 0 US Excluded GB LB 1 8. 6. 4. 2. 0 US and Eurozone Excluded GB LB Local Currency Non-Local Currency Notes. Theplotshowstheshareofloansinlocalcurrencyversusnon-localcurrencygivenbyglobalandlocal banks. The left panel is based on loans from all countries in the sample except the US. The right panel is based on loans from all countries in the sample except the US and Euro area countries. Source: Dealscan, Amadeus, Orbis, Compustat, Compustat Global, and author’s calculation. 64

Figure B.3: Estimates of Average Productivity Measure log z by Country it Australia Austria Belgium Canada Czech Republic Denmark Finland France Germany Greece Ireland Italy Japan Mexico Netherlands New Zealand Norway Poland Portugal Spain Sweden Switzerland United Kingdom United States 0 1 2 3 log(z) i Notes. Estimates of the productivity measure log z averaged across firms and years by country, calculated it based on Equation (4). Source: Dealscan, Amadeus, Orbis, Compustat, Compustat Global, and author’s calculation. 65

Figure B.4: Estimates of Global Factor zG Gz 4 3 2 1 0 2004 2006 2008 2010 2012 2014 2016 Notes. AplotoftheglobalfactorzG,extractedfromthefirstprincipalcomponentofthez series. Thefactor ict values have been adjusted upward by their minimum so that all the values are positive. Source: Dealscan, Amadeus, Orbis, Compustat, Compustat Global, and author’s calculation. 66

Figure B.5: Bank-Firm Sorting, by zG/zL Quartile (Method 2) i i 1 8. 6. 4. 2. 0 zG/zL i i 1 2 3 4 Global Bank Local Bank Notes. The plot shows sorting patterns between firms and global versus local banks, with firms sorted into quartiles by their exposure to global versus local risk (zG/zL), uses variables that are constructed based on i i Method2ofthebankcategorizationcriteriaforglobalbanks. Datasampleconsistsofsyndicatedloansbetween firmsglobalandlocalbanksandfirmsacross24countriesfrom2004-2017. Source: Dealscan,Amadeus,Orbis, Compustat, Compustat Global, and author’s calculation. 67

Figure B.6: Three-Month Euribor Rates around ECB Announcements Target Announcement PC Start PC End dleiY deilpmI 80.360.340.320.3 3 89.2 April 6, 2006 8 9 10 11 12 13 14 15 16 17 18 Target Announcement PC End PC Start dleiY deilpmI 3.1 52.1 2.1 51.1 1.1 Nov 3, 2011 8 9 10 11 12 13 14 15 16 17 18 Time Stamp Notes. The figure plots the three-month Euribor rates on April 6, 2006 (upper panel) and November 3, 2011 (bottompanel)between08:00and18:00. Verticallinesrepresentthetargetpolicyrateannouncement(13:45), the start of the press conference (14:30), and the end of the press conference (15:30). All times are in CET. Source: CQG Data Factory. 68

Table B.1: Summary Statistics: Loan and Firm Count by Country (Method 2) Country Loan GB LB Firm Country Loan GB LB Firm Australia 4507 0.70 0.30 701 Japan 21341 0.45 0.55 2865 Austria 387 0.53 0.47 61 Mexico 601 0.70 0.30 137 Belgium 704 0.69 0.31 123 Netherlands 2028 0.54 0.46 406 Canada 6760 0.64 0.36 903 New Zealand 1023 0.70 0.30 127 Czech Republic 197 0.68 0.32 77 Norway 1017 0.66 0.34 253 Denmark 327 0.56 0.44 84 Poland 318 0.54 0.46 87 Finland 587 0.65 0.35 113 Portugal 254 0.65 0.35 64 France 5876 0.67 0.33 996 Spain 4380 0.68 0.32 839 Germany 5987 0.68 0.32 942 Sweden 875 0.66 0.34 190 Greece 309 0.66 0.34 47 Switzerland 790 0.69 0.31 175 Ireland 404 0.70 0.30 107 United Kingdom 6810 0.69 0.31 1528 Italy 2378 0.67 0.33 688 United States 46732 0.70 0.30 1466 Notes. Sample constructed from Dealscan, Amadeus, Orbis, Compustat, Compustat Global, and author’s calculation. Sample period covers the year 2004-2017. 69