University of Hawai`i at Mānoa Department of Economics Working Paper Series

Transcription

1 Univerity of Hawai`i at Mānoa Department of Economic Working Paper Serie Saunder Hall 542, 2424 Maile Way, Honolulu, HI Phone: (808) Working Paper No Who i to Blame: Foreign Ownerhip or Foreign Funding? By Inea Love Roberto Rocha Erik Feyen Samuel Munzele Maimbo Raquel Letelier September 2014

2 Who i to Blame: Foreign Ownerhip or Foreign Funding? Inea Love, Roberto Rocha, Erik Feyen, Samuel Munzele Maimbo, and Raquel Letelier 1 JEL Claification: G01, G21, F36. Key word: Global financial crii, bank credit, foreign bank, funding model. 1 Inea Love i an Aociate Profeor at the Univerity of Hawaii (ilove@hawaii.edu), Roberto Rocha i a Senior Advior in the World Bank Financial Sytem Department (rrocha@worldbank.org), Erik Feyen i a Lead Financial Economit in the World Bank Financial Sytem Department (efeijen@worldbank.org), Samuel Munzele Maimbo i a Lead Financial Sector Specialit in the World Bank' Europe and Central Aia Department (maimbo@worldbank.org) and Raquel Letelier i a Financial Analyt in the World Bank Europe and Central Aia Department (rletelier@worldbank.org). We thank María Soledad Martínez Pería, Stijn Claeen, participant of the June 2013 World Bank ECA region retreat, participant of the Univerity of Hawaii eminar and participant of the 2014 FIRS conference for ueful comment and uggetion.

3 Who i to Blame: Foreign Ownerhip or Foreign Funding? Abtract We invetigate whether the credit contraction that followed the global financial crii i due to high foreign ownerhip or high reliance on foreign finding. We apply panel vector autoregreion to quarterly data for 41 countrie and find that dometic credit growth i highly enitive to cro-border funding hock around the world. However, high foreign ownerhip per e doe not appear to increae the enitivity of credit to foreign funding hock. Rather, the enitivity i higher in countrie with high reliance on foreign funding and high loan-to-depoit ratio. Thee finding have important policy implication for many countrie involved in cro-border funding. JEL Claification: G01, G21, F36. Key word: Global financial crii, bank credit, foreign bank, funding model.

4 I. Introduction Although mot countrie acro the world experienced a evere contraction of credit during the recent global financial crii, the Eatern Europe and Central Aia (ECA) region wa hit harder than mot of the developing countrie. One poible reaon for the evere decline lie in cloe economic and financial link with Wetern Europe, including a high degree of foreign ownerhip (i.e. foreign equity) and greater reliance on foreign funding (i.e. foreign liabilitie) in the banking ector. The credit crunch wa generally driven by foreign bank, a they initiated effort to repair their balance heet and retrench to home market. The concerted policy repone, including the Vienna Initiative, ucceeded in mitigating a evere credit contraction in the participating countrie (de Haa et al, 2012). However, thee event triggered a renewed debate about the benefit and cot aociated with the preence of foreign bank, cro-border finance, and financial integration. We ue the pecial cae of ECA countrie to provide new evidence on the relative role that foreign ownerhip and foreign funding play in the tranmiion of global financial crie. Previou reearch ha etablihed that foreign bank have much to contribute to financial ector development (ee Claeen and Van Horen, 2014, for a comprehenive review of the literature). 2 However, a the global financial crii ha made clear, their preence can alo have large detabilizing impact (Mihkin, 2007). Recent reearch how that foreign bank drove the 2 Specifically, foreign bank can accelerate financial development by lowering the cot of financial intermediation, increaing the quality of and acce to financial ervice (Giannettia and Ongena, 2012), increaing competition in the hot country, bringing in up-to-date technology to the market, introducing new, more diverified product and ervice, and preing regulator to reform and modernize the regulation and uperviion of financial ytem. Allen et al (2012) alo dicu cot and benefit of cro-border banking in detail. 3

5 credit boom in ECA before the crii and exacerbated the credit contraction after the crii. 3 Even the US branche of foreign bank were ignificantly affected (Cetorelli and Goldberg, 2012). Concerned about their potentially detabilizing impact, many countrie are reconidering the merit of limit on foreign bank ownerhip. 4 Depite the widepread agreement that foreign bank have amplified the tranmiion of the global hock to hot countrie, the evidence appear more nuanced and there i ignificant heterogeneity among the foreign bank. For example, foreign bank that focued on depoitbaed funding contracted lending le than dometic bank during the crii (Claeen and Van Horen, 2013, Ongena et al 2012). Alo, foreign bank which were more integrated into a network of dometic co-lender reduced lending le during the crii (De Haa and Van Horren, 2013). Kapan and Minoiu (2013) how that bank with tronger balance heet were better able to maintain lending during the crii, while thoe more dependent on market funding which i typically horter term and more fickle reduced the upply of credit more than other bank. Cerutti and Claeen (2013) point that market-baed meaure of vulnerabilitie of banking ytem to hock were more important than accounting baed meaure in explaining bank deleveraging. In addition, Choi, Gutierrez, and Martinez Peria (2013) find that bank with better capitalized parent were more table in their lending. Thee finding ugget that the type of bank funding model, along with the regulatory tandard and uperviory arrangement, maybe more important for tability than foreign ownerhip per e. Foreign bank can finance their operation predominantly with local depoit- 3 Cetorelli and Goldberg (2011), de Haa and Van Lelyveld (2014), de Haa et al (2012), Cull and Martinez Peria (2012), Popov and Udell (2012), Feyen and González del Mazo (2013), Impavido et al (2013). 4 For example, Indoneia and Namibia have impoed new retriction on foreign bank in

6 baed funding, a they do in mot Latin American countrie, or with the foreign funding from their parent bank or cro-border wholeale funding ource, a in mot ECA countrie. Caual obervation ugget that the ource of funding played an important role during the recent crii. For example, ome Eatern European countrie with a high degree of foreign ownerhip but le reliance on foreign funding eem to have been le affected (e.g. Czech Republic, Slovakia), while other countrie with a lower degree of foreign ownerhip but trong reliance on croborder funding were highly affected (e.g. Kazakhtan and the Ukraine). It i important for policymaker to know whether the detabilizing impact of foreign bank depend on the origin of bank funding (i.e. liabilitie) or the origin of bank ownerhip (i.e. equity). While other author have argued that cro-border funding i one of the main channel of tranmiion of global crie, earlier tudie cited above lacked the direct meaure of reliance on foreign funding becaue they ued bank-level data, motly from Bankcope databae, which doe not include information on the origin of bank liabilitie. 5 Thu, the previou reearch ha not been able to directly pinpoint the foreign funding channel a the culprit of the foreign-bank induced credit crunch. In thi paper we provide direct evidence on the impact of foreign funding on credit growth, which ha immediate implication for the regulatory debate on cro border banking. We make everal contribution to the ongoing reearch effort to dientangle the role of foreign ownerhip and foreign funding in the tranmiion of the global financial crii. Firt, we ue precie meaure of a country reliance on foreign funding uing data on foreign liabilitie of 5 While ome previou author ue the ratio of depoit to aet to proxy for the level of foreign funding, thi variable i crude and doe not accurately meaure the expoure to foreign funding, and thu have only indirect relevance for the policy debate. 5

7 the banking ytem. Second, we compare how factor affecting credit growth were different in ECA v. other region. Third, we follow a different methodology Panel Vector Autoregreion (PVAR) allowing u to examine the dynamic relationhip between funding tructure and credit growth over time. There are everal advantage in uing PVAR methodology to tudy the importance of different factor in explaining credit growth. Firt, in VAR, all variable are treated a endogenou and interdependent o all the feedback effect are explicitly included in the model. Thu, VAR are deigned to explicitly addre the endogeneity problem, which i a eriou challenge in tudying the empirical relationhip between credit growth and funding ource. Second, unlike ingle country VAR that need long time erie for efficiency, PVAR can be ued with relatively hort time-erie. Thi i important for our tudy ince the quarterly data on foreign liabilitie are only available for about 10 year. Third, the PVAR methodology can ditinguih between the hort-term impact of each of the factor baed on the impule-repone function and the long-term cumulative impact of hock baed on variance decompoition. Fourth, a in any panel-baed method, PVAR allow u to control for country- and year- fixed effect. Country-fixed effect will capture time-invariant country characteritic that can explain credit growth, uch a intitution, the rule of law, the quality and depth of credit information and other relatively tatic feature of the buine environment. Year-fixed effect capture global hock affecting finance and growth, uch a the effect of the global financial crii that i common for all countrie in the ame time period. Fifth, the VAR model can ditinguih whether demand (e.g. GDP) or upply (e.g. foreign liabilitie and depoit) factor drove the decline of the private credit, which i important becaue all thee variable deteriorated imultaneouly 6

8 during the crii. Thu, PVAR i a methodology that i very well-uited to the quetion thi tudy aim to addre. Our reult can be ummarized a follow. We apply PVAR to a global panel that conit of quarterly data for 41 countrie in ECA, Aia, Latin America, and the Middle Eat and Africa for the period We document four main finding. Firt, we find that private credit growth i highly enitive to cro-border funding hock around the world. Second, we find that thi relationhip i ignificantly tronger in the average ECA country compared to the ret of our ample. Specifically, the repone i 72% higher in the initial period and about three time higher after one quarter. Third, we how that countrie with high loan-to-depoit ratio and high reliance on foreign funding exhibit a tronger repone of private credit to foreign funding hock. However, foreign ownerhip per e doe not lead to different repone in our ample of countrie. Fourth, we how more directly that high loan-to-depoit ratio and high foreign funding are able to explain the difference between ECA and the ret of the world, while high foreign ownerhip by itelf i inufficient to explain thee difference. Taken together, our finding ugget that funding model difference were at the heart of ECA protracted pot-crii credit growth contraction, not it high prevalence of foreign bank ownerhip. Thee finding have important policy implication. Rather than trying to curtail and cale back foreign bank preence, thi paper ugget regulator ought to focu on buine model of local affiliate of international bank and the regulation of cro-border funding. We dicu policy implication in more detail in the concluion of the paper. 7

9 II. Data We compile a panel databae with quarterly data covering 41 countrie in variou region with about 11 year of data available for mot countrie. Table 1 provide a decription of the variable and ource. We have two et of variable: the time erie variable are ued for the panel VAR and the cro-country variable are ued for ample plit. We decribe each of thee et of variable in detail below. II.1. Panel VAR Variable There four main variable in our PVAR model, which are taken from the IMF International Financial Statitic (IFS): private credit, foreign liabilitie, GDP, and depoit. For mot of the countrie in our ample thee variable are available quarterly tarting with Q2, 2001 and ending with Q4, The private credit variable meaure the quarter-on-quarter real growth rate of private credit. Private credit iolate credit iued to the private ector and therefore exclude credit iued to government, government agencie, and public enterprie. Private credit alo exclude credit iued by central bank. The foreign liabilitie variable meaure the quarter-on-quarter real growth rate of foreign liabilitie. Thee liabilitie include claim of non-reident on the dometic banking ytem which include depoit, ecuritie, loan, financial derivative and other liabilitie. GDP i the quarter-on-quarter, eaonality-adjuted growth rate of real GDP. The depoit variable meaure quarter-on-quarter real growth rate of 6 Some countrie have horter erie due to data availability, detail available on requet. 8

10 bank depoit, coniting of the dometic time, aving, and demand depoit of depoit money bank. All variable are denominated in local currency. For all variable we firt calculate quarter-on-quarter nominal growth rate. Then we contruct real growth rate by ubtracting the quarter-on-quarter percent change in the CPI index. 7 Thu, our main PVAR variable are expreed in growth rate, which enure their tationarity. Table 2A how that the mean real quarterly growth rate of foreign liabilitie were higher than thoe of private credit and depoit, and it tandard deviation wa coniderably higher, by a factor of three or larger. Thi reult reveal the major weakne of foreign funding it volatility in period of crii. The mean growth rate of foreign liabilitie and private credit were much higher in the cae of ECA countrie, relative to Non-ECA countrie, reflecting the greater reliance of the ECA region on foreign funding and the key role played by cro-border finance in ECA pre-crii credit boom. Interetingly, the tandard deviation of foreign liabilitie in the ECA and non-eca region are imilar, revealing that countrie in the two region are potentially expoed to the ame volatility in period of crii, but ECA countrie are much more expoed, given their greater reliance on foreign funding. Thi reult will be further explored in Section V. II.2. Sample Split Variable 7 We have alo ued GDP deflator to adjut GDP growth rate, but it had le coverage. All our reult hold uing the GDP deflator, however the GDP reult become lightly le ignificant perhap becaue of the added noie. In addition, we removed ignificant country-pecific eaonality in GDP growth rate. To do that we regreed real GDP growth rate on country-pecific quarter dummie and ubtracted the predicted value. 9

11 One of the objective of thi paper i to undertand what drive the difference in the repone of private credit to other PVAR variable between ECA and the ret of the world. In doing o, we ue three variable hypotheized to explain thee difference: the loan-to-depoit ratio (LDR), foreign ownerhip, and foreign funding. LDR i calculated a the ratio of private credit to depoit, both variable were defined above. LDR proxie the extent to which dometic bank credit i funded by dometic depoit. An LDR in exce of 100 percent implie that credit ha been financed with other ource of funding uch a unecured wholeale funding in the international or dometic interbank market. Thee type of funding are typically hort-term and highly enitive to market condition in contrat with (retail) depoit which are uually much more table. The foreign funding meaure i a proxy of the banking ytem reliance on foreign ource of funding relative to total funding. It i calculated a the ratio of foreign liabilitie to the um of foreign liabilitie and depoit, which roughly repreent total debt funding. Latly, foreign ownerhip meaure the fraction of banking ytem aet that are majority-owned by non-reident. Thee data are taken from the World Bank Bank Regulation and Superviion Survey. To gauge the impact of thee variable, we plit the ample baed on their pre-crii value. 8 Table 2B report the pre-crii country average for our three ample-plit variable broken for ECA and non-eca countrie. Table 2B how that LDR, foreign funding and foreign ownerhip are ignificantly higher in ECA countrie, relative to non-eca countrie. However, the tandard deviation alo ugget that there i ubtantial variation acro countrie within 8 We ue 2007:Q4 value for LDR and foreign funding. We ue the cloet available obervation for foreign ownerhip, which i 2008 for mot countrie. 10

12 each of the two region. Thi i an important reult that will alo be further explored in Section V. III. Trend and decriptive tatitic For our graphical analyi we break down the ECA region into the Commonwealth of Independent State (CIS) and the Central, Eatern and Southeatern Europe region (CESEE). Many countrie in both ample enjoyed a credit boom prior to the global financial crii, the CIS growing by 50-60% per year while the CESEE growing at 30%-40% between 2005 and mid-2008 (Figure 1, Panel A). Among other region, only the Gulf Cooperation Council (GCC) countrie grew at a imilar pace, while the non-gcc countrie in the Middle Eat and North Africa (MENA) region, and the Latin American and Aia region howed moderate rate of credit growth during that period. Depoit growth rate in thee economie were alo growing at high rate (Panel B), although lower than the rate of credit growth, a reflected in the growing loan-to-depoit ratio (Panel F). Loan-to-depoit ratio in the CIS went a high a 170% before the crii, and continued to increae even after the crii, reaching almot 180% at it peak in mid CESEE reached it maximum ratio in early 2009, with level of 140%, while GCC went a high a 125% around the ame period. Converely, other emerging region howed more teady level, and in the cae of non-gcc and Emerging Aian countrie, alway below 100%. Thee trend can be largely explained by difference in banking model acro region. While LAC, non-gcc and Aia financed their credit growth with local depoit (keeping their 11

13 loan-to-depoit ratio under 100%), CESEE, CIS and GCC countrie were heavily reliant on foreign borrowing. Figure 1, Panel D diplay the hare of foreign liabilitie to total liabilitie (i.e. it i the foreign funding meaure decribed above) and how the different path thee region followed. While foreign liabilitie in LAC, non-gcc and Aia were in the range of 10%- 20% of total liabilitie, foreign borrowing in CESEE, CIS and GCC increaed prior to the crii, exceeding on average 40% of total liabilitie in early The ECA region i generally known for the dominant preence of foreign ownerhip in comparion with emerging peer, but thi i only true for the CESEE countrie (Panel G). By 2008, 80% of the banking ytem in CESEE wa in foreign hand, a figure that barely urpaed 40% in LAC and 30% in other emerging region. The level of foreign ownerhip were lower in the large CIS countrie, epecially in countrie like Ruia and the Ukraine, where foreign bank accounted for le than 20% of total aet. After the onet of the crii in 2008, credit experienced a ignificant lowdown in all emerging region. Thi wa more acutely felt in thoe region where credit had been increaing at a rapid pace and GDP contracted heavily in 2009, uch a CESEE and CIS (Panel C). A a conequence of the credit crunch, thee region witneed an increaing deterioration in aet quality, a hown by the ratio of nonperforming loan to total loan (Panel E). 9 III.1. Correlation 9 High NPL in CIS are exacerbated by Kazakhtan, which experienced a houing market collape, leading to a deterioration in the quality of bank aet, who were large invetor in the real etate market. 12

14 Panel A of Table 3 preent the correlation between our PVAR variable for the full ample, a well a eparate correlation for ECA and the ret of the ample ( non-eca ). A expected, the correlation between depoit and private credit i high in all ample. We alo oberve that foreign liabilitie are correlated with private credit (0.40). However, thi correlation i almot twice a high in ECA compared to non-eca (0.50 v. 0.23). Thi ugget credit in the ECA region could be much more vulnerable to foreign funding hock than in the ret of the world, a hypothei we formally tet with PVAR in Section V. Panel B of Table 3 diplay the correlation for the ample plit variable. It illutrate the different funding model in ECA reflected in the high correlation between foreign funding and LDR in ECA, but the abence of uch correlation in the non-eca region. Indeed, many bank in ECA funded their operation by attracting foreign funding to upplement dometic depoit in order to utain high credit growth, reulting in elevated loan-to-depoit ratio. The correlation ugget thi practice wa not a widepread in non-eca countrie. While foreign funding and foreign ownerhip are poitively correlated, the correlation i not very high: 0.35 in the full ample and near zero in ECA (Table 3, panel B). The poitive correlation i expected, a foreign owned bank often obtain funding from parent bank and have better acce to international funding market. However, in ECA there are many cae of countrie with high level of foreign liabilitie but low level of foreign ownerhip and vice vera. For example, Kazakhtan and Ukraine had relatively low level of foreign ownerhip, but high level of foreign liabilitie, while Czech Republic and Slovakia had high level of foreign ownerhip (around 90 percent of aet) but limited reliance on foreign funding (around 10 percent of liabilitie). Thee correlation ugget that many dometically-owned bank in ECA, epecially thoe in CIS, relied to a ignificant extent on foreign funding. 13

15 III.2. T-tet: I ECA Different from the Ret of the World? In thi ection we tet whether the pattern dicued above are tatitically ignificant. Our main focu i on the difference between ECA and the ret of the world (i.e. non-eca). Therefore, we plit our ample into an ECA and non-eca ubample and run baic mean comparion t-tet. Specifically, we track the evolution of our variable before ( ), during ( ), and after ( ) the global financial crii hit. Table 4 preent the reult. Panel A document the ECA v. non-eca difference for the PVAR variable: private credit, foreign liabilitie, GDP, and depoit. We oberve that in the pre-crii period, ECA wa ignificantly different from the ret of the ample in term of the growth of foreign liabilitie, private credit, and depoit. The real rate of growth of foreign liabilitie and private credit for ECA wa more than double relative to non-eca: 10.06% v. 3.64% for foreign liabilitie and 7.87% v. 3.15% for private credit. Thee difference are tatitically ignificant and highlight the ditinctivene of the ECA cro-border, wholeale funding model which conit of rapid private credit growth utained by inflow of foreign funding. The reulting ECA credit boom i demontrated by higher average real depoit growth (5.42% v. 2.65%), but it did not reult in ignificant difference in real GDP growth (0.95% v. 0.51%). Next, we tudy the PVAR variable during the peak of the crii in During thi period, GDP growth fell ignificantly more in ECA compared to non-eca (-1.86% v. - 14

16 0.51%), while the private credit growth wa till ignificantly higher in ECA (2.15% v. 1.45%). Yet, depoit growth difference are not tatitically ignificant. The difference for foreign liabilitie wa not tatitically ignificant in thi period (even though the mean are till higher for ECA, there wa ubtantial variation within the region). Thi may point to the lag in adjuting bank operation, which quare well with our ordering of the foreign liabilitie a firt (i.e. the variable that repond to all other with a lag). In the immediate pot-crii year ( ), many non-eca countrie had tarted their recovery a evidenced by the return of poitive growth in foreign liabilitie, private credit, and depoit. In contrat, foreign liabilitie and credit growth were ignificantly lower in ECA, reflecting the onet of the EU overeign and banking crii, which had protracted effect on cro-border lending.. The continuation of the crii in the EU i alo reflected in the lower GDP and depoit growth rate in ECA, although the difference are not tatitically ignificant. Thee finding ugget that non-eca countrie have recovered fater from the impact of the crii. Given uch ditinctive difference in the impact of the global crii and the path of recovery in ECA v. non-eca economie, we eek to undertand what key pre-crii feature of financial ytem in ECA could be reponible for uch difference. A dicued, we hypotheize the difference could be attributed to diimilaritie in the funding model and/or the extent of foreign ownerhip. Therefore, to gauge funding model difference we perform t-tet on LDR, foreign ownerhip and foreign funding. Table 4, Panel B how the reult. The average LDR in ECA i ignificantly higher than in non-eca (126.4% v. 98.4%). Since ECA LDR exceed 100% by a wide margin, ECA had to fund it liquidity gap with other wholeale ource, which are typically regarded a le table than retail depoit. Thee wholeale ource were often of foreign nature in ECA a evidenced by the fact that foreign 15

17 funding wa twice a high ECA compared to non-eca (31.9% v. 15.8%). While thee foreign ource of funding can act a tabilizer when tre originate from the dometic financial ector (e.g. De Haa and Van Lelyveld, 2014, they can alo function like conduit of foreign financial tre. The ECA funding model contrat harply with the one of foreign bank affiliate in Latin America which typically ource their funding in dometic market motly retail depoit which impede exceive credit expanion and the tranmiion of foreign tre. At the ame time, we find that foreign ownerhip i alo very high in ECA. In the average non-eca country, 31.4% of banking aet are majority-owned by non-reident compared to 66.7% in the average ECA country. While ECA ha high average of all three ratio, not all countrie have equally high foreign bank ownerhip and high foreign funding dependence, a we dicued above. We ue thi variation to ae which of thee two factor i reponible for the difference in credit growth in our PVAR analyi below. IV. Methodology PVAR combine the advantage of the traditional Vector Autoregreion (VAR) with the advantage of panel-data model. Baic VAR allow for the imultaneou analyi of the evolution of a ytem of endogenou variable. Moreover, the PVAR allow u to eparately tudy demand and upply factor to better undertand the evolution of private credit. The prime benefit of VAR i that the dynamic impact of orthogonal hock can be evaluated i.e. the iolated impact of a hock of one variable on the ytem over time, keeping the hock of the other variable equal to zero. 16

18 The PVAR technique i particularly uitable for our purpoe becaue we eek to model the evolution of a ytem of our four variable of interet private credit, foreign liabilitie, GDP, and depoit in a et of countrie which ignificantly differ along variou dimenion uch a the development of their financial and economic ytem, financial regulation and uperviory effectivene, and exchange rate regime. Thee unobervable country-pecific difference are captured with country fixed effect in our model. Moreover, thee countrie are faced with common expoure uch a the global cycle, global rik appetite, key interet rate, extraordinary meaure of large central bank, and other global financial and economic hock. Controlling for thee common expoure i alo important becaue our ample include the period of the global financial crii during which financial tre rapidly pilled over to other part of the world ignificantly increaing the ize of common rik factor. Thee common time hock are captured with time fixed effect in our model. The tructure of our baeline PVAR model can be written in reduced form a follow: Ζ it = Γ 0 + Γ 1 Ζ it-1 + f i + d t +e it (1) where Ζ it i a vector of our four variable for quarter t and country i and modeled a a function of the firt-order lag of all variable in the ytem. The f i and d t term are country fixed effect and time fixed effect repectively. The country fixed effect preent an etimation challenge which arie in any model which include lag of the dependent variable: the fixed effect are correlated with the regreor and therefore the mean-differencing procedure commonly ued to eliminate fixed effect would create biaed coefficient. Our PVAR implementation follow Love and Zicchino 17

19 (2006). 10 Specifically, to remove fixed effect we ue forward mean-differencing, alo referred to a the `Helmert procedure' (ee Arellano and Bover, 1995). 11 Thi procedure remove only the forward mean, i.e. the mean of all the future obervation available for each firm-year. Thi tranformation preerve the orthogonality between tranformed variable and lagged regreor, which allow u to ue lagged regreor a intrument and etimate the coefficient by ytem Generalized Method of Moment (GMM). 12 The time-fixed effect are removed by timedifferencing all the variable prior to GMM etimation, which i equivalent to putting time dummie in the ytem. Finally, to minimize the influence of outlier, we winorize all variable by replacing the data below/above the 1 t /99 th percentile with the value of the 1 t /99 th percentile. 13 The PVAR allow u to model the impact of an iolated hock or innovation of a ingle variable on the whole ytem over time, while etting the innovation of all other variable equal to zero. Thi produce o-called impule-repone function, which take into conideration the etimated coefficient matrix (given by Γ 1 in equation 1), a well a the correlation of reidual acro equation (i.e. e e). However, the error acro the variable in the ytem are typically correlated which inhibit the attribution of the impact of an innovation of a ingle variable to that variable only. To iolate the impact of innovation, it i neceary to decompoe the reidual in uch a way that they become orthogonal. The common olution i to adopt a particular variable ordering which 10 More recently, Love and Turk Ari (2014) have ued the ame PVAR etimation methodology to evaluate the impact of macroeconomic hock on bank loan portfolio quality in Egypt. 11 The forward mean differencing i an alternative to firt differencing. It i preferable becaue it preerve more data, preerve the variance, and doe not induce firt order autocorrelation, which then need to be corrected for a in the cae of firt differencing. 12 In our cae the model i jut identified becaue the number of regreor equal the number of intrument, therefore ytem GMM i mathematically equivalent to equation-by-equation 2SLS. Alo, there are no overidentifying retriction becaue the number of intrument i equal to the number of variable in the model. 13 We chooe to winorize the data to avoid a ignificant reduction in our ample ize by dropping thee outlier. However, our reult are robut to excluding thee obervation from the ytem (reult available upon requet). 18

20 aume that variable that come earlier in the order affect all following variable contemporaneouly, while variable that come later affect previou variable only with a lag. While all the variable are treated a endogenou, the ordering only affect the timing of the repone a the variable that come later in the ordering have a delayed repone on the variable that come earlier in the ordering. In other word, the correlation between the reidual of two variable i allocated to the variable that come firt in the ordering. The ordering can therefore have implication for the hape of impule-repone function and for variance decompoition. Thi procedure i known a a Choleki decompoition of the variancecovariance matrix of reidual and i equivalent to tranforming the ytem into a recurive'' VAR for identification purpoe (ee Hamilton (1994) for the derivation and dicuion of impule-repone function). In our baeline model, we aume the following ordering: foreign liabilitie, GDP, depoit and private credit. We place foreign liabilitie firt ince it i to a large extent driven by external upply factor uch a global rik appetite, parent bank health, economic condition in a home country, and global funding market. Thi aumption implie that foreign liabilitie affect all other variable contemporaneouly. In contrat, the other variable can only affect foreign liabilitie with a 1-quarter lag. Thi i a reaonable aumption ince revering the flow of foreign liabilitie i likely to take ome time. We place private credit lat in the order becaue arguably private credit can react to all other factor quickly, i.e. in the ame quarter; however, the private credit only affect other variable with a 1-quarter lag. Indeed, typically there i a delay between loan origination and loan deployment, o an impact on other variable can only be expected with a lag. 19

21 Having etablihed foreign liabilitie and private credit a firt and lat in the order, repectively, in our baeline model we put GDP econd, followed by depoit. Our rationale i that GDP ha a more immediate impact on depoit a money demand repond more quickly to change in GDP, while change in depoit are likely to affect GDP only with a lag. However, the alternative ordering in which depoit enter econd and GDP third produce qualitatively imilar reult (available on requet). To analyze the impule-repone function we need an etimate of their confidence interval. Following Hamilton (1994) we calculate confidence interval with Monte Carlo imulation. 14 Finally, we alo employ variance decompoition to undertand the cumulative impact of the hock of a particular variable on the ytem. Variance decompoition how the percent of the total variation in one variable that i explained by the hock of another variable after a certain amount of time. They therefore provide an indication of the magnitude of the total effect one variable exert on another. We report the total effect accumulated over 10 quarter, but longer time horizon produced equivalent reult ince after 10 quarter the effect of a hock ha motly worked it way through the ytem (and mot impule repone have converged to zero in our etimation). 14 Specifically, we randomly generate a draw of coefficient Γ 0 and Γ 1 from Equation (1) uing the model etimation and the variance-covariance matrix and re-calculate the impule-repone. We repeat thi procedure 200 time and then generate the 5th and 95th percentile of thi ditribution which we ue a confidence interval. 20

22 V. Empirical Reult V.1. The Baeline Model for the Entire Sample Our baeline PVAR model the interaction between private credit, foreign liabilitie, GDP, and depoit uing Generalized Method of Moment (GMM) etimation for our entire ample of 41 countrie. The baeline model will provide a good idea of the baic interaction between the variable for an average country and help u interpret the finding when we tart plitting the ample. We report the coefficient etimate of our reduced form PVAR model for the full ample, and the ECA and non-eca ubample in Table 5. Since our main focu i to iolate how hock to one variable affect another variable, we don t focu on the coefficient etimate and intead turn our attention to the impule-repone function, which take into account the coefficient, a well a the variance-covariance matrix of the error. Figure 2 how 16 impule-repone function for our baeline model. Each row correpond to a particular variable and how 4 impule-repone function which diplay how thi variable repond to an iolated hock of each of the variable in the ytem (i.e. including the variable itelf). Each repone i traced for 6 period (i.e. 1.5 year) after which the hock ha motly worked it way through the ytem and any reidual impact i minimal. For example, the graph in row 1, column 4 report the repone of foreign liabilitie to a hock in private credit and the graph in row 4, column 1 report the repone of private credit to a hock in foreign liabilitie. Becaue private credit come later in the ordering, it impact on foreign liabilitie i delayed by one period (and hence there i no effect at time zero), while the impact of foreign 21

23 liabilitie on private credit i immediate and poitive at time zero. Each graph how the point etimate of the impule-repone function (the middle line) a well a the 5% and 95% confidence bound (the top and the bottom line) baed on Monte Carlo imulation. The repone i tatitically ignificant if the confidence interval doe not include the zero line. Given our reearch objective to explain the behavior of private credit, we focu on row 4 which diplay how private credit repond to variou hock. The ame point etimate of the impule repone plotted on the Figure 2 (i.e. the middle line) are alo reported in Table 6. We document a ignificantly poitive repone of private credit to a hock in foreign liabilitie. A one tandard deviation hock in foreign liabilitie reult in a 1.24% increae in private credit growth at time zero which i large given that average private credit growth in our entire ample i 3.2% and it tandard deviation i 5.12% (ee Table 2). Thi i a key reult which we will tudy further in order to undertand repone difference in ECA v. non-eca countrie. We alo find a poitive repone of private credit to GDP, which in our model capture the demand for credit. Thi repone i alo tatitically ignificant, but omewhat maller in magnitude: 0.59% at time zero. Finally, we oberve a poitive and ignificant repone of private credit to a depoit hock: 1.53% at time zero. Thu, we conclude that all our upply and demand factor are ignificant driver of private credit. 15 Other impule-repone function in Figure 2 how intereting and intuitive reult that are tatitically ignificant. For example, foreign liabilitie repond poitively to a private credit hock (row 1, column 4) uggeting that a udden increae in credit can be funded by foreign liabilitie which can typically be attracted on hort notice. At the ame time, depoit alo 15 Takát (2010) find that both demand and upply factor contributed to the fall in gredit during the crii, but the impact of upply factor wa tronger. 22

24 repond poitively to a private credit hock (row 3, column 4) but the magnitude i maller than for foreign liabilitie, arguably becaue it i more difficult to ignificantly raie depoit in the hort term. Together, thee finding imply that foreign ource of funding can be ueful to temporarily fill dometic funding gap. We alo find that a foreign liabilitie hock ha a poitive impact on depoit (row 3, column 1) which implie that bank purue depoit growth when foreign funding increae, poibly to avoid a relative over-reliance on foreign funding. Alternatively, an increae in depoit could be due to the perception that better performing bank are able to attract foreign funding. We alo confirm that a private credit hock boot GDP which i expected ince private credit typically expand conumption and invetment. Thi corroborate with reult of the finance and growth literature (e.g. Levine, 2005). While impule-repone function how the hort-term repone of each variable to a hock in another variable, the long-term cumulative impact of a hock i captured by the variance decompoition. Table 7, Panel A diplay the variance decompoition for our baeline model. Each cell in the table how what percent of variation in the row variable i explained by the column variable after the 10 quarter, i.e. 2.5 year. Note that for all variable their own hock explain mot of the variance (i.e. the diagonal of Table 7, Panel A contain the larget value). Row 4 of Table 7, Panel A report that a foreign liabilitie hock explain a relatively large part of the variance in private credit (9.7%). The depoit hock explain an even larger portion of the variance (14.2%), while a hock to GDP explain relatively little variance at 2.4%. Taken together, our PVAR model explain a ubtantial portion of the variation in private credit (i.e. 26.2% i explained by other variable and 73.8% i explained by it own hock). While our model explain maller portion of the variation in the other variable, it work reaonably well to fulfill our main objective, which i to explain the behavior of private credit. 23

25 Before we proceed with our ample plit, we conduct variou robutne check to our baeline model (reult of thee additional tet are available on requet). Firt, we include an additional lag of all variable in the model and find that our reult are robut, even with decreaing degree of freedom. 16 Second, we change the ordering of the variable where we interchange depoit and GDP and find our main reult till hold. Finally, we find that dropping large outlier (i.e. extreme obervation above 99th percentile and below 1t percentile) rather than winorizing the data doe not affect the reult. Therefore, in the remainder of thi paper, we ue the ordering of the baeline model and include 1 lag only. V.2. Sample Split: I ECA Different from the Ret of the World? In thi ection we formally tet whether private credit in the ECA region reponded differently compared to the ret of the world. To do that, we plit our full ample into ECA and non-eca ubample and run eparate PVAR for each ubample. The two ubample are roughly equal in ize: (ECA: 22 countrie, 947 obervation; non-eca: 19 countrie, 820 obervation). Figure 3 preent the reult. For pace conideration, we only preent impulerepone function and variance decompoition for private credit, our key variable. In other word, only the fourth row of graph from Figure 2 i preented for the plit ample. The firt and econd row in Figure 3 how the private credit repone to hock of all four variable for the ECA and non-eca ubample, repectively. Both row exhibit very imilar pattern to our 16 Adding an additional lag to the model ignificantly increae the number of coefficient that have to be etimated i.e. from 16 in our baeline model to 32 and reduce the number of obervation from 1,591 in the baeline model to 1,554 becaue one year of data i lot. The lo in degree of freedom i particularly relevant for model for ubample. 24

26 baeline finding and confirm our key reult: the ignificant and poitive repone of private credit to a foreign liabilitie hock. We alo oberve poitive and ignificant repone in both ample to GDP and depoit hock. The key quetion, however, i whether the magnitude of the private credit repone to a foreign liabilitie hock i ignificantly different between the ECA and non-eca ubample. Viual inpection of the 95 th percentile bound of the firt two row how that repone to a foreign liabilitie hock i ubtantially larger in the ECA ample: 1.71% v. 1.10% in non-eca. Table 6 report the actual point etimate for the impule-repone function in period zero: 1.53 and 0.89 in ECA and non-eca, repectively. The ECA repone i thu 0.64 percentage point higher (72%), which i an economically relevant difference. The difference i even more pronounced after 1 quarter: the repone in period 1 in ECA i 0.74, while in non-eca it i 0.23 (i.e. the repone i about 3 time higher in ECA). Thi pattern continue for everal quarter: in quarter 3 the ECA repone i at 0.28, while the non ECA i at Thu, the impact of foreign liabilitie hock in ECA i not only tronger in magnitude, but alo lat longer. To ae whether thi difference i tatitically ignificant, we calculate the impulerepone function of the difference between ECA and non-eca. Becaue the two ubample are independent, the impule-repone of the difference are equal to the difference in impulerepone. To calculate the confidence interval, we merge the ditribution of error produced by Monte-Carlo imulation with 200 repetition each for both ample and generate a new ditribution which i the difference between error generated in each of the repetition (i.e. each of the 200 error now contain the difference between the ditribution). From thi new ditribution we generate new 5 th and 95 th percentile bound. 25

27 Figure 3, row 3 preent the reult of the difference in impule repone ( Sample: Difference ). Row 3, column 1 confirm the difference in repone of private credit to foreign liabilitie hock i indeed ignificant (i.e. the zero line i outide of the confidence interval). Other impule-repone function in row 3 how that private credit doe not behave differently in ECA in repone to a GDP hock, while there i a lightly larger repone to a depoit hock after 1 period. Latly, we compute the variance decompoition in the ECA and non-eca ubample to tudy the cumulative longer-term impact of variou hock on private credit. Table 7, Panel B report the reult. To ave pace, we only report the decompoition for the private credit variable. The firt row i the baeline decompoition for the whole ample, replicated from Panel A. The econd and third row how the decompoition for the ECA and non-eca ubample, repectively. A foreign liabilitie hock explain 6.0% of the private credit variation in non-eca countrie and 12.9% of the variation in the ECA ubample, which i more than twice a large. Thee finding etablih the econd of our key reult: private credit in ECA ha been more heavily influenced by hock to foreign liabilitie. V.3. Further Sample Split: Explaining Difference in Private credit Repone After etablihing that private credit i ignificantly more reponive to foreign liabilitie hock in ECA compared to non-eca countrie, thi ection eek to identify the factor that could drive the difference. A dicued earlier, the banking ector in ECA i markedly different from thoe in non-eca countrie along our ample plit variable: ECA countrie exhibit 26

28 ignificantly higher foreign ownerhip, higher reliance on foreign funding and higher LDR ratio. To determine which of thee factor can explain the differential repone of private credit to foreign funding hock, we perform ample plit uing each of the three ample-plit variable one at a time. In each cae we plit the whole ample in two ubample of equal ize baed on the median value of the variable in quetion. The ample-plit approach i imilar to interacting each of the factor with the reponivene of private credit to foreign liabilitie. 17 Firt, we tet whether LDR i driving the repone difference. We calculate the pre-crii median LDR ratio in our full ample uing data from the fourth quarter of We then plit the full ample into high and low LDR ubample of equal ize baed on the median and rerun our baeline model for each ample. The impule-repone reult are preented in Figure 4 and the variance decompoition are preented in Table 7, Panel B. We find that the difference between the repone of private credit to a foreign liabilitie hock in both high LDR and low LDR ample i poitive and tatitically ignificant. 18 The variance decompoition how that in the high LDR ubample, a foreign liabilitie hock explain 14.1% of the variation in private credit, while it explain only 5.6% of the variation in the low LDR ample. Thi finding ugget that high LDR are aociated with a tronger repone of private credit to foreign funding hock. While thi doe not preent a direct proof, thi finding ugget that high LDR are at leat partially reponible for higher repone of private credit to foreign funding hock in ECA. 17 Note that uch interaction cannot be modeled directly in a VAR etting. 18 The ignificance i at about 5% in period 1 becaue the bottom 5 th percentile line i touching the zero line, while it i tronger in period 0 and period 2 and 3. 27

29 Next, we perform another ample plit to tet whether foreign funding dependence i driving the difference. Similarly, we ue pre-crii foreign funding value for all countrie from the fourth quarter of 2007 and plit the ample into high and low foreign funding ubample. The impule-repone reult are preented in Figure 5 and the variance decompoition are preented in Table 7, Panel B. We find that the repone of private credit to a foreign liabilitie hock i ignificantly higher in the high foreign funding um-ample. The variance decompoition how that in the high foreign funding ubample a foreign liabilitie hock explain 12.9% of the variation in private credit, while it only explain 6.9% in the low foreign funding ubample. Thee reult ugget that high reliance of countrie on foreign funding i aociated with tronger repone of private credit to foreign funding. Again, the reult are in line with the hypothei that high reliance on foreign funding i reponible for explaining the difference between the ECA and non-eca ample. Latly, we plit our full ample uing foreign ownerhip. Baed on the latet pre-crii data, we create high and low foreign ownerhip ubample. The impule-repone reult are preented in Figure 6 and the variance decompoition are preented in Table 7, Panel B. We find that the difference between the repone of private credit to a foreign liabilitie hock in the high and low foreign ownerhip ample are not tatitically ignificant (row 3). Thi i in contrat to previou finding of ample plit on LDR and foreign funding. The variance decompoition how that in the high foreign ownerhip ubample a foreign liabilitie hock explain 10.0% of the variation in private credit while it explain 9.1% in the low foreign ownerhip ubample. Thee two value are not materially different here, while there wa ubtantial difference in the two previou ample plit. Thee reult ugget that in the whole 28

30 ample foreign ownerhip i not aociated with a tronger repone of private credit to foreign liabilitie. Taken together, our finding ugget that high LDR and high foreign funding in ECA are aociated with tronger repone of private credit to foreign funding hock, while high foreign ownerhip doe not appear to drive thi difference. Thee reult ugget that ECA funding model i at the heart of ECA vulnerability to external financial hock rather than the trong preence of foreign bank in the region per e. However, while thee reult are uggetive of the reaon for why ECA i different, they do not directly demontrate thi becaue we have ued our whole ample to do the ample plit. In the next ection we preent a more direct evidence to explain the difference between ECA and non-eca. V.4. What factor explain difference of ECA and the ret of the world? In thi ection we provide more direct evidence on the factor that drive the difference in ECA and the ret of the world. While all three factor LDR, foreign funding and foreign ownerhip are higher in ECA, not all factor are preent in all countrie, a we dicued above. Thi allow u to invetigate which of the three factor i driving the difference between ECA and non-eca in term of the meaured repone of private credit to foreign funding hock. The intuition for our procedure i a follow. We create a truncated ECA ample which i contructed to reemble to the non-eca ample in one of the three indicator, without retricting the other two factor. For example, we make a truncated ECA ample by removing from ECA ample countrie with extra high LDR. In practice, we drop countrie with LDR that i higher than the median LDR in ECA. Thi truncated ECA ample i more imilar to the ret 29