Business cycle effects on commercial bank loan portfolio performance in developing economies

Available online at www.sciencedirect.com

Review of Development Finance 1 (2011) 150–165

Business cycle effects on commercial bank loan portfolio performance in developing economies夽 Jack Glen, Camilo Mondragón-Vélez ∗ International Finance Corporation, United States Available online 15 April 2011

Abstract This paper studies the effects of business cycles on the performance of commercial bank loan portfolios across major developing economies in the period 1996–2008. We measure loan performance via loan loss provisions (that is, recognized expenses related to expected losses in bank income statements). Our results indicate that while economic growth is the main driver of loan portfolio performance, interest rates have second-order effects. Furthermore, we find the relationship between loan loss provisions and economic growth to be highly non-linear only under extreme economic stress: GDP growth needs to decline by more than 6 percentage points (pp, in absolute terms) in order to generate an increase in loan loss provisions equivalent to median emerging market bank profits; while a decline of more than 10 pp in growth implies significant capital losses, of at least 20 percent, for the median emerging market bank. In addition, we find higher loan loss provisions are associated with private sector leverage, poor loan portfolio quality, and lack of banking system penetration and capitalization. © 2011 Production and hosting by Elsevier B.V. on behalf of Africagrowth Institute. JEL classification: E44; G21; O16 Keywords: Bank loan portfolio; Loan loss provisions; Loan loss reserves; Lending rate; Developing economies; Business cycle

1. Introduction Understanding bank performance gains renewed interest with every financial crisis and the 2009 global financial crisis was no exception. Although banks across industrialized countries captured most of the attention owing to their leading role in the origins of the crisis, they were not the only ones under scrutiny. Given the increasing importance of developing countries in the global economy and memories of the emerging market crises of 夽

The views expressed in this paper are those of the authors and do not necessarily represent those of the IFC, IFC Management or the World Bank Group. All errors and omissions are sole responsibility of the authors. ∗ Corresponding author at: Mail Stop MSN F6P-605, 1818 H Street NW, Washington DC, 20433, United States. Tel.: +1 202 473 8667. E-mail addresses: [email protected] (J. Glen), [email protected] (C. Mondragón-Vélez).

the 1990s, regulators, analysts and investors expressed concern about the strength of developing world banking systems in the face of global recession. In the aftermath of last year’s crisis, except for the particularities of foreign currency lending across eastern European countries, the banking systems across most of the remaining major emerging economies were in general resilient to the global downturn even though GDP growth rates in many of these countries dropped significantly. For instance, data from the IMF Global Financial Stability Report (April 2010) shows that while the level of non-performing loans (NPLs, measured as a percentage of total loans) in 2009 was more than 3.5 times larger than the level observed in 2007 for the U.S. and the U.K.; that ratio was less than 1.5 times larger in the case of Brazil, China and India. Moreover, 90 percent of (reported) developing economies across Asia, Latin America, Middle East and Africa experienced an increase in the level of NPLs in 2009 less than two times the level in 2007.1

1879-9337 © 2011 Production and hosting by Elsevier B.V. on behalf of Africagrowth Institute. Peer review under responsibility of Africagrowth Institute, Republic of South Africa. doi:10.1016/j.rdf.2011.03.002

Production and hosting by Elsevier

1 Loan performance deterioration was significantly higher across Emerging Europe. While only one quarter of countries saw their NPL ratio increasing by less than two times between 2007 and 2009, 60 percent of countries saw increases between two and ten times in that period. For a simulation model on NPLs for Emerging Europe see Box 1.2 and Annex 1.6 in IMF’s Global Financial Stability Report (April 2010).

J. Glen, C. Mondragón-Vélez / Review of Development Finance 1 (2011) 150–165

There are alternative explanations for recent banking performance across the developing world. One explanation points to an underlying governing relationship between loan portfolio performance and the business cycle; in most countries declines in GDP were significant though not extreme given that most developing countries (except Emerging Europe) faced the indirect effects of global recession rather than a domestic crisis. An alternative explanation focuses on the importance of stronger institutional frameworks reflecting past reforms and the lessons learned from past crisis episodes. Most surely, both explanations are relevant.2 However, this paper focuses on the first explanation by studying the effects of economic growth and interest rates on the performance of commercial bank loan portfolios across major developing economies during the period 1996–2008.3 Given the short time series of this sample, we exploit cross-sectional variation in the data to characterize an average relationship between bank loan portfolio performance and the business cycle across the developing world.4 Understanding these dynamics and quantifying the underlying relationship between loan performance measures and key macroeconomic indicators is of particular importance for risk analysis tools, including VaR (Value at Risk) and stress testing. Furthermore, such an estimate becomes critical to investors and analysts that do not have full access to detailed information about the structure and historical performance of bank loan portfolios by providing a simplified (though structured and statistically supported) approach to modeling loan performance sensitivities to macroeconomic scenarios and generating top-down analysis to flag potential vulnerabilities going forward.5 There is a large literature looking at macroeconomic determinants of bank loan portfolio performance. Most of these studies are country specific, many of which are focused on particular crisis episodes with a wide coverage across the industrialized world. The work can be grouped into three main methodological approaches. First, some authors use reduced-form linear models. Among this group Arpa et al. (2001) work with a sample of Austrian banks, Gerlach et al. (2005) uses data from banks in Hong Kong and Quagliariello (2004) studies the case of Italy. A second group uses VAR (Vector Auto Correlation) models. This group includes Babouˇcek and Janˇcar (2005) who work with data from the Czech Republic, and Hoggarth et al. (2005) for the case of England. A third group of studies focuses on the transmission mechanisms through the impacts on default and loss given default. For this literature see Altman et al. (2002), Pesaran et al. (2006), Segoviano (2006a,b); and Padilla and Segoviano (2006) for an application to stress testing. However, there is relatively little research looking at emerging markets overall. Aside from the cross-country estimations presented in the 2003 IMF Financial Soundness Indicators background paper and Annex

2 For an analysis of emerging markets during the 2009 financial crisis see Llaudes et al. (2010). 3 The period of analysis is dictated by data availability as of December 2010. 4 We define developing economies as those classified as Low Income, Low Middle Income and Upper Middle Income countries by the World Bank. 5 See Hoggarth et al. (2005) for details on top-down versus bottom-up stress test frameworks.

151

1.6 in the 2010 IMF Global Financial Stability Report, which presents a simulation of NPLs in Emerging European countries, to the best of our knowledge there is no other documented characterization of the relationship between loan performance and the business cycle for developing countries. In addition, the literature discussed above usually relies on linear relationships (or impulse response functions from linear VAR structures) between NPLs and macroeconomic determinants such as GDP, inflation, interest rates or exchange rates. However, data from past banking and financial crisis episodes across developing countries suggest the possibility of non-linearity under high economic stress (closely related to the so-called tail effects in the VaR, default and loss given default literature). Thus, the nature of the underlying relationship between loan performance and the business cycle remains an open question. It is worth noting that the dynamics of this relationship imply a series of two-way causalities. It is usually the case that initially a negative (positive) economic shock impacts loan portfolios of diverse qualities across the banking system. This is then followed by a contraction (expansion) of credit growth by banks, which in turn affects economic growth, jump-starting a new round of effects. Along these lines, the estimates documented in this paper pretend to capture the observed overall effects. We test alternative models to characterize this relationship, which include linear and non-linear specifications at the individual bank and banking system levels. In addition, we offer an alternative measure of loan portfolio performance through the use of loan loss provisions. As recently stated by John C. Dugan, U.S. Comptroller of the Currency, loan loss provisioning “allows banks to recognize an estimated loss on a loan or a portfolio of loans when the loss becomes likely, well before the amount of loss can be determined with precision and is actually charged off”.6 Given loan loss provisions are recognized expenses to increase loan loss reserves on the balance sheet, they provide a potentially valuable measure of bank reassessments about expected losses in their loan portfolio. Furthermore, while loan loss provisions and reserves are related to an assessment on the entire loan portfolio, not all NPLs may generate future losses (although the assigned probabilities of doing so are higher for these in the calculation of loan loss provisions and reserves). Section 2 describes the data and discusses relevant issues of using loan loss provisions as a measure of loan portfolio performance. We combine bank level data from the commercial database Bankscope with macroeconomic and banking system level data from the World Bank and the IMF. The Bankscope database contains financial statement information for more than 29,000 private and public banks globally over more than 15 years. We restrict the sample to 22 major developing economies for which there is available information on a diverse and representative group of banks. This group of countries accounts for 85% of the developing world’s GDP, as well as more than 80% of the developing economies commercial banking assets available in Bankscope. The sample

6 Remarks by John C. Dugan, Comptroller of the Currency before the Institute of International Bankers, March 2, 2009. “Loan Loss Provisioning and Procyclicality”.

152

J. Glen, C. Mondragón-Vélez / Review of Development Finance 1 (2011) 150–165

Table 1 Number of banks per country and year. 1996 Argentina Brazil Chile China Colombia Egypt India Indonesia Malaysia Mexico Nigeria Pakistan Peru Philippines Poland Romania Russia South Africa Thailand Turkey Ukraine Vietnam

67 80 19 1 23 20 23 63 36 7 36 9 17 4 9 1 17 5 8 10 5

1997 65 79 21 3 22 24 40 52 35 17 44 9 20 7 9 3 23 5 10 12 7

1998 63 80 23 3 21 26 44 57 32 20 44 11 20 8 8 6 17 4 13 11 6

1999 62 80 21 5 19 24 45 57 31 25 48 10 16 7 4 11 44 5 13 22 10

2000 61 85 21 6 22 24 44 47 24 25 48 9 15 3 2 14 74 6 16 18 17 3

2001 65 103 23 13 22 23 48 44 24 22 52 10 13 4 1 12 97 5 18 18 21 10

20002 62 104 22 28 22 22 52 39 26 23 47 11 12 7 16 110 3 18 20 20 8

2003 57 89 22 40 24 20 52 47 26 23 39 17 12 6 1 17 112 2 18 22 21 11

2004 57 87 21 47 23 20 53 51 27 24 26 19 12 20 13 12 452 11 17 13 14 16

2005 56 88 23 63 19 19 47 52 27 23 14 24 11 21 15 11 641 15 19 11 18 19

2006 53 77 21 59 14 10 41 41 27 22 16 23 10 21 15 10 822 13 19 11 20 16

2007

2008

55 101 1 106 13 18 41 52 27 28 19 23 11 20

56 85 22 92 14 18 40 50 29 29 17

22 871 18 19 25 41 32

21 907 18 19

13

41 35

Source: Bankscope (only includes banks with available financial information within countries with available economic data).

contains more than 11,300 bank-year observations in the period 1996–2008. Our estimates in Section 4 show that economic growth is the main driver of loan portfolio performance, while interest rates have second-order effects. The coefficients of the linear model specification for GDP growth and lending interest rates are consistent with past estimations from the IMF, as well as the findings of Glen (2005) on an analysis of interest coverage ratios for non-financial companies in developing countries. In addition, non-linear specifications in GDP growth provide the best statistical fit to the data. More importantly, the average underlying relationship implies that significant loan performance deterioration only occurs under extreme economic stress. Estimations based on bank level data indicate that GDP needs to decline by more than 6 percentage points (pp, in absolute terms) in order to generate an increase in loan loss provisions equivalent to median bank profitability in the sample (as measured by return on total assets). Similarly, a decline of more than 10 pp in the economy’s growth rate implies an exponential increase in loan loss provisions. For minor economic slowdowns the level of provisioning tends to be linear in GDP. We also find that higher private sector indebtedness, individual banks or banking system leverage (or lower capitalization) and accumulated loan loss reserves (a reflection of poor loan portfolio quality) are associated with higher levels of loan loss provisions. The paper is organized as follows: Section 2 describes the data; Section 3 documents stylized facts of the relationship between loan loss provisions, GDP growth and interest rates for selected developing countries; Section 4 presents the estimations and Section 5 concludes.

2. Sample data We combine bank level data from the commercial database Bankscope with macroeconomic and banking system level data from the World Bank and the IMF. The Bankscope database contains detailed financial statement information for more than 29,000 private and public banks around the world over more than 15 years. We focus on the period 1996–2008. While developing economies coverage has improved substantially over the years, the number of observations before 1996 is small across most emerging markets. On the other hand, there is a one to two year data collection lag. Therefore we truncate the analysis at year 2008 in order to prevent biases in the data in favor of early reporting banks. We restrict the sample to 22 major developing economies for which there is available information on a diverse and representative group of banks, as well as available macroeconomic series. Table 1 shows the number of banks per country and year included in the sample. This group of countries accounted for 85% of developing world GDP in 2006, and contains more than 11,300 bank-year observations representing more than 80% of developing economies commercial banking assets in Bankscope in this period.7 Our measure of loan portfolio performance is the ratio of loan loss provisions to gross loans. Loan loss provisions are defined in this context as recognized expenses in the income statement reflecting expected losses in the loan portfolio during the period, in contrast to loan loss reserves accumulated in the balance sheet.

7 These figures include only banks with available financial information for the period of analysis.

J. Glen, C. Mondragón-Vélez / Review of Development Finance 1 (2011) 150–165

153

Fig. 1. Serial correlations of loan loss provisions (as a percentage of gross loans) with GDP growth and lending rates for the period 1996–2008. Table 2 Balance sheet size, loan portfolio size and capitalization (year 2006). Balance sheet size total assets (USD billion)

Loan portfolio size gross loans/assets (%)

Capitalization total equity/assets (%)

Argentina Brazil Chile China Colombia Egypt India Indonesia Malaysia Mexico Nigeria Pakistan Peru Philippines Poland Romania Russia South Africa Thailand Turkey Ukraine Vietnam

82.0 741.7 113.4 3,366.1 46.0 61.3 657.2 147.9 296.0 235.5 48.5 65.4 26.0 71.9 129.4 26.2 238.7 214.2 228.5 54.7 25.1 19.7

39.1 37.7 71.8 51.4 65.3 35.2 58.2 48.5 55.4 53.0 30.6 59.4 59.6 38.5 51.1 55.5 63.4 73.4 71.8 52.1 77.8 47.3

10.9 9.9 8.2 5.9 10.8 5.7 5.7 10.5 7.2 13.2 13.6 8.2 9.4 11.2 10.5 9.7 13.3 6.3 8.3 12.0 10.4 8.6

Total

6,895.6

52.6

7.5

Source: Bankscope.

We chose this measure over non-performing loans (NPLs) for two reasons. On the one hand, the availability of NPL data in Bankscope is limited and not uniform across countries or banks. This is due to the fact that NPLs are not a line in the financial statements, but are in general disclosed within the notes to the financial statements. On the other hand, the definition of NPLs varies across countries and it may not be possible to access the data needed to make definitions consistent. Hence, even if the data were available, there is no guarantee in terms of consistency. An alternative source of data on provisions and NPLs (by country) are the IMF Financial Soundness Indicators in the statistical appendix of the Global Financial Stability Report. We use this data as a reference, but keep Bankscope as the primary source given the availability of bank level data. The estimations

in the paper are done at the bank and country levels. Table 2 shows aggregate figures for the size of the banking system (sum of total assets), relative size of the loan portfolio (gross loans to total assets) and capitalization (equity to total assets) in year 2006 for all countries included in the sample. The sample consists of 11,306 bank-year and 276 countryyear observations (including banks across 22 countries spanning 13 years).8 Although the panel is not fully balanced, we do not consider this as a significant source of bias in the estimation.9

8 Banks with Provisions to Gross Loans above 100% and below −100% were excluded from the sample. 9 The main results using bank level data are robust to the exclusion of Russian banks (which account for 37% of the bank-year observation in the sample).

154

J. Glen, C. Mondragón-Vélez / Review of Development Finance 1 (2011) 150–165

15.0

12.5 2002 10.0 7.5 5.0 1996

2001 2.5

2000

1999

0.0 -2.5 -15

-5

-10

0

1997 1998 2008 2007 2004 2006 2005 2003 GDP Growth (%) 5

10

15

Provisions / Gross Loans (%)

Provisions / Gross Loans (%)

Argenna 15.0

12.5 2002 10.0 7.5 5.0

1996 1999 2000 1997 1998 2004 2007 2005 2006

2.5 0.0

2001 2008 2003

Lending Rate (%)

-2.5 0

20

10

20

30

40

50

15.0

Provisions / Gross Loans (%)

Provisions / Gross Loans (%)

Russia 15.0 12.5 10.0 7.5 1998

5.0

1999 1996 2008 2005 2006 1997 2002 2007 2004 2001 2003 2000

2.5 0.0

GDP Growth (%)

-2.5 -15

-10

-5

0

5

10

15

12.5 10.0 7.5 1998

5.0 2005 2008 2006 2002 2007 20042003 2001

2.5 0.0

1999

1997 2000

Lending Rate (%)

-2.5 0

20

10

20

30

40

50

15.0

12.5 2001 10.0 7.5 5.0 1997 2003 2000 2002 1996 2008 2004 1998 2007 20062005 GDP Growth (%)

1999

2.5 0.0 -2.5 -15

-10

-5

0

5

10

15

Provisions / Gross Loans (%)

Provisions / Gross Loans (%)

Turkey 15.0

12.5 2001 10.0 7.5 5.0 2003

2.5 0.0

1999

2002 2000

2007 2005 2004 2006

1997

1998 1996 Lending Rate (%)

-2.5 0

20

10

20

30

40

50

60

70

80

90

100

Provisions / Gross Loans (%)

Provisions / Gross Loans (%)

Ukraine 15.0 1998 12.5 10.0 7.5 1997 5.0

2000

2008

2.5

2001 2002 2006 2003 1999 2005 2007 2004

1996

0.0

GDP Growth (%)

-2.5 -15

-10

-5

0

5

10

15

20

15.0 1998 12.5 10.0 7.5 1997 5.0

2008

2000

2001

1999

2005 2003 2002 2006 2004 2007

2.5 0.0

1996 Lending Rate (%)

-2.5 0

10

20

30

40

50

60

70

80

90

100

Fig. 2. Provisions, GDP growth and lending rates for Argentina, Russia, Turkey and Ukraine (1996–2008). (Russia not plotted in right panel for year 1996 (4.1% provisions to gross loans and 147% lending rate). Turkey not plotted in right panel for year 2008 due to lack of interest rate data (1.4% provisions to gross loans).)

With regard to the main macroeconomic effects included in the estimation, we use standard real gross domestic product growth rates (reported by the World Bank) and lending interest rates (reported by the IMF).10 The lending rate is not only the type of interest rate most related to commercial banking loan activ-

ities, but provides us with a comprehensive time series across countries.11 Other macroeconomic controls include the level of domestic credit to the private sector (a measure of private sector leverage), the level of domestic credit provided by the banking sector (a measure of financial sector penetration) and the current

10 The IMF’s International Financial Statistics publication defines the Lending Rate as the bank rate that usually meets the short- and medium-term financing needs of the private sector; which is normally differentiated according to creditworthiness of borrowers and objectives of financing.

11 We completed the IMF series with data from national sources (usually Central Banks) in some cases. These include Brazil and Vietnam for years 1995–1996, as well as Pakistan and Turkey for the entire series.

J. Glen, C. Mondragón-Vélez / Review of Development Finance 1 (2011) 150–165 China 5.0

Provisions / Gross Loans (%)

5.0

Provisions / Gross Loans (%)

155

4.0 3.0 2.0 2002 1997 2008 2000 2004 2006 2003 2005 2007 1999 2001 1996 1998

1.0 0.0

4.0 3.0 2.0 1998 1997 2002 1999 20002006 2008 2004 2005 2003 1996 2001 2007

1.0 0.0

GDP Growth (%) -5

0

5

10

5.0

15

2005 2006 1996

3.0 2004

1997

2007 2000 1998 1999 2001

2.0

2003 2002

1.0

5

10

15

Egypt 5.0

2008

4.0

0

20

Provisions / Gross Loans (%)

Provisions / Gross Loans (%)

Lending Rate (%)

-1.0

-1.0

0.0

2008 2005 2006

4.0

1996

3.0 2007 2004 1997 2003 2000 1998 2002 1999 2001

2.0 1.0 0.0

GDP Growth (%)

Lending Rate (%)

-1.0

-1.0 -5

0

5

10

15

20

20

0

5

10

15

20

Fig. 3. Provisions, GDP growth and lending rates for China and Egypt (1996–2008).

account balance (a measure of external leverage), all measured as a percentage of GDP (as reported by the World Bank).12 3. Stylized facts of loan loss provisions, GDP growth and interest rates We now provide some statistical evidence on the relationship between loan portfolio performance and the business cycle. The intuition behind this relationship is straightforward. For any loan portfolio originated by a bank in the past, the performance of that portfolio at present and going forward is conditional on current and future economic conditions. If for instance healthy growth in the economy goes on, the bank should not expect abnormal deterioration in their loan portfolio performance, but maybe some early or partial repayments within their customer base. However, if economic conditions deteriorate sufficiently firms may not be able to service their debt given lower revenues, while individuals affected by unemployment may stop paying their (credit card, auto or student loan, mortgage) bills once their savings are exhausted. On the other hand, interest rate hikes may affect the ability of firms or individuals to continue servicing their debt. The left panel of Fig. 1 shows the serial correlation between loan loss provisions and GDP growth for the period 1996–2008, and the right panel shows the serial correlation between provisions and lending rates. For most countries the correlation of loan loss provisions with GDP growth is negative (except for Egypt) and the correlation with lending rates is positive (except for Chile, China, Egypt and Vietnam). In addition, there is significant variation across countries.

12 As of January 2011, year 2009 official figures for these measures were not available.

We explore this heterogeneity by looking at the experience of a select group of countries during the period of analysis. Fig. 2 shows the level of provisions relative to GDP growth and lending rates for Argentina, Russia, Turkey and Ukraine. These countries experienced at least one crisis episode during 1996–2008, as reported in Laeven and Valencia (2008). Loan loss provisions in these countries increase significantly and spike around the reported year of the crisis (Argentina in 2001, Russia in 1998, Turkey in 2000 and Ukraine in 1998). Also note that in all cases provisions tend to remain at stable levels for all other years despite sizable differences in the pace of economic growth. Furthermore, the data in all these cases clearly suggest a negative relationship between provisions and GDP growth and a positive relationship with lending rates. However, the observed change in provisions relative to these macroeconomic variables differs significantly across countries. For instance, while Argentine banks had 11.2% loan loss provisions under −10.9% GDP growth and 51.7% interest rates in 2002, Turkish banks had a similar levels of provisions (11%) in 1998 with −5.7% GDP growth and 78.8% interest rates. In contrast, Russian banks experienced milder increases in loan loss provisions (up to 5.0%) during the 1998 crisis, despite experiencing −5.3% GDP growth and 41.8% interest rates. Fig. 3, on the other hand, shows the level of provisions relative to GDP growth and lending rates for China and Egypt. Recall that these countries were exception cases in the analysis of serial correlations described in Fig. 1. Furthermore, these two countries did not have a banking crisis episode in the period of analysis (as reported by Laeven and Valencia, 2008). The patterns in Fig. 3 show no clear relationship between provisions and GDP growth or interest rates. While loan loss provisions maintain a stable level in all years, GDP growth is always positive and interest rates vary in a limited range. Hence, these countries

156

J. Glen, C. Mondragón-Vélez / Review of Development Finance 1 (2011) 150–165

did not experience enough economic stress during the period of analysis to see a significant deterioration in loan performance and thus tend to behave as the group of countries presented in Fig. 2 do during no crisis times (detailed figures for provisions, GDP growth rates and lending interest rate levels for all countries in the sample are provided in Appendix A). The evidence presented in this section highlights important differences with regard to the magnitude of business cycle effects across countries. While the estimation in Section 4 may not fully explain this heterogeneity, it intends to characterize the dynamics of loan portfolio performance along the business cycle in emerging markets, as an initial step in our understanding of this phenomena going forward. In this regard, countries with crisis episodes during the period of analysis (which include Argentina, Brazil, Colombia, Indonesia, Malaysia, Philippines, Russia, Thailand, Turkey and Ukraine) contribute their loan performance variation under business cycle shocks; while the remaining countries act as a control for variations across non crisis times. Therefore, we should expect a differentiated relationship between provisions and the business cycle under significant economic stress than under mild slowdowns or continued growth times. 4. Estimation As mentioned before, we generate estimates using a sample of individual banks, as well as a sample of country banking systems. The relationship we estimate is of the general form: Prov NLoant = F (GDP ct , INTct , B - t, M - ct ) where Prov NLoant is the loan loss provisions to gross loans ratio for individual banks or banking systems at time t; GDPct is the growth rate of GDP for country c at time t; INTct is the lending interest rate within country c at time t; Bt is a set of variables describing characteristics of individual banks or the entire banking system; and Mc,t is a set of macroeconomic variables describing other dimensions of the economy over time.13 The functional form of F, as well as the specific set of variables included in B and M is unknown. Therefore, we explore alternative specifications with each of the samples. 4.1. Estimations using country level data We begin by presenting estimations at the country level, which provide a more balanced sample over time by avoiding bias due to over-sampling of banks within certain countries (see Table 1) and measure business cycle effects over the entire banking system. However, this comes at the cost of reducing sample size and underlying data variation. Country level ratios are generated as weighted averages rather than medians or means of the corresponding sample of individual banks, so as to reflect aggregate banking system characteristics.

13

We use the change in lending rates rather than the level, in order to control for differences in underlying inflation levels across developing countries during the period of analysis.

The first specification we test is a parsimonious linear model in GDP growth and lending rates, including one-period lags14 : Prov NLoanc,t = a + BGDPc,t + B1 GDPc,t−1 + γINTc,t + γ1 R INTc,t−1 + ξt

(1)

The estimates for this specification are included in the first column of Table 3 (all specifications were run with robust standard errors). GDP effects are negative as expected. While the contemporaneous GDP effect is significant, the lagged effect is not statistically significant. Similarly, interest rates are positive; and while the effect of the contemporaneous change in the lending rate (INT) is significant, the lagged (real) lending rate effect is not statistically significant. In addition, note that in terms of economic significance the effect of GDP growth (−0.408) is more than ten times larger than that of interest rates INT (+0.0328). This is consistent with the findings documented in the 2003 IMF Financial Soundness Indicators background paper on a global sample of countries, as well as with Glen (2005) on an analysis of interest coverage ratios for non-financial sector firms across developing countries. The statistical fit of this specification is relatively low (R2 = 0.28). A second specification (reported in column 2 of Table 3) includes macroeconomic (M) controls: current account to GDP ratio (CA/GDP), domestic credit to the private sector as a percentage of GDP (DomCredPriv/GDP), domestic credit provided by banks as a percentage of GDP (DomCredProv by Banks/GDP) and development stage by income group dummies (as defined by the World Bank): low middle (LMI) and upper middle income (UMI). Prov NLoanc,t = α + BGDPc,t + B1 GDPc,t−1 + γINTc,t + γ1 R INTc,t−1 + M - c,t + ξt

(2)

The estimates for GDP and INT related variables under this specification are similar to those in (1), without much improvement in overall fit (R2 = 0.31). With the exception of CA/GDP and the UMI dummy, none of the remaining macroeconomic covariates is statistically significant. The next model (reported in the third column of Table 3) includes banking system characteristics (B) in addition to macroeconomic variables (M): loan loss reserves to gross loans (LoanLossRes/GLoan), gross loans to total assets (GLoan/Total Assets) and total equity to total assets (Equity/Total Assets) ratios. Prov NLoanc,t = α + BGDPc,t + B1 GDPc,t−1 + γINTc,t + γ1 R INTc,t−1 + M - c,t + B- c,t + ξt

(3)

The importance of banking system characteristics is implied by the increase in overall fit (R2 = 0.54). There are also notable changes across some covariates under this specification. The 14 We consider the real lending rate (R INT = INT – CPI Inflation) as the lagged interest rate effect, rather than last year’s change in lending rates. In this way we measure the additional effect of a change in interest rates coming from a low or high real interest rate environment at the beginning of the period.

Table 3 Provisions to gross loans regressions on GDP and lending rates at the banking system level. Linear models with no fixed effectsa

GDPt

Linear model with country fixed effectsb

Polynomial model including country fixed effectsb

(2)

(3)

(4)

(5)

(6)

−0.4084** (0.19)

−0.4262** (0.193)

−0.2831*** (0.088)

−0.2206*** (0.075)

−0.4019*** (0.107)

−0.3441*** (0.107) −0.0337* (0.02) 0.0059** (0.0024) 0.0004** (0.0002) −0.00004*** (0.00001) −0.0789 (0.061) 0.0319** (0.014) 0.0205 (0.015) 0.0311 (0.033) 0.0687*** (0.026) −0.0338** (0.016) 0.0061 (0.677) −0.4995 (0.801) 0.4061*** (0.137) −0.0132 (0.023) −0.1248* (0.07) 2.4843 (1.777) 276 0.71

GDPt 2 GDPt 3

0.0045** (0.0021)

GDPt 4 GDPt 5 GDPt−1 INTt REALINTt−1

−0.1682 (0.168) 0.0328** (0.014) 0.0140 (0.011)

−0.1708 (0.158) 0.0375** (0.015) 0.0178 (0.012) 0.0571** (0.029) 0.0004 (0.014) −0.0042 (0.012) 0.1478 (0.267) −0.7118* (0.396)

4.9759*** (0.908) 276 0.28

5.4203*** (0.959) 276 0.31

CA/GDP DomCredPriv/GDP DomCredProv by Banks/GDP LMI dummy UMI dummy LoanLossRes/loans Gloans/total assets Equity/total assets Constant term No. observations R2 a b

−0.0694 (0.101) 0.0297** (0.013) 0.0189 (0.013) −0.0031 (0.023) 0.0173 (0.011) −0.0182** (0.009) 1.0459** (0.52) 0.968* (0.519) 0.2854*** (0.081) −0.0096 (0.02) −0.2122** (0.097) 3.9362*** (1.318) 276 0.54

−0.0324 (0.069) 0.0303** (0.014) 0.0212 (0.014) 0.051* (0.03) 0.085*** (0.024) −0.0425** (0.018) −0.0985 (0.786) −0.6061 (0.869) 0.4877*** (0.108) −0.0062 (0.027) −0.2037*** (0.067) 2.5892 (1.768) 276 0.68

−0.00002* (0.00001) −0.0394 (0.071) 0.0306** (0.014) 0.0205 (0.013) 0.0575* (0.03) 0.0783*** (0.025) −0.0383** (0.019) −0.0281 (0.756) −0.4734 (0.846) 0.4752*** (0.109) −0.0026 (0.026) −0.1947*** (0.065) 2.1702 (1.833) 276 0.69

J. Glen, C. Mondragón-Vélez / Review of Development Finance 1 (2011) 150–165

(1)

Robust standard errors in parentheses: ***p < 1%, **p < 5%, and *p < 10%. Country dummy coefficients not reported.

157

158

J. Glen, C. Mondragón-Vélez / Review of Development Finance 1 (2011) 150–165

contemporaneous GDP coefficient is now statistically significant at 1% but reduces its magnitude (to −0.28), while INT remains significant at 5% and reduces its economic significance as well (to +0.0297). However, the relative magnitude of GDP to interest rate effects is maintained. DomCredProv by Banks/GDP (a measure of banking sector penetration) is negative and significant, as well as the LMI and UMI dummies (positive and significant at 5% and 10%, respectively).15 With regard to banking system variables (B), larger loan loss reserves are associated with higher provisions. Given that loan loss reserves reflect loan portfolio quality, this result suggests that the lower the quality of the loan portfolio banks start with, the higher their vulnerability to business cycle effects. The negative (and significant) effect of Equity/Total Assets on provisions suggests that banking system capitalization is a barrier against business cycle vulnerability, as expected. GLoan/Total Assets (a measure of loan portfolio size) is not statistically significant. We test for country and time fixed effects using specification (3). For this purpose, we use a standard F test to compare the robust (including fixed effects) and simple OLS (not including fixed effects) models. We find evidence of country fixed effects but not of time fixed effects. Hence, we run an additional linear specification including country fixed effects (Fc ), which increases the R2 to 0.68 (estimates reported in column 4 of Table 3).16 Prov NLoanct = α + BGDPct + B1 GDPct−1 + γINTct + γ1 R INTct−1 + M - ct + B - ct + F- c + ξt (4) Under specification (4) the coefficient of GDP growth falls to −0.22. INT remains around 0.03, while their lags continue not to be significant.17 While DomCredPriv/GDP (a measure of private sector leverage) is positive and highly significant, DomCredProv by Banks/GDP is negative and significant at 5%. LoanLossRes/Loans and Equity/Total Assets preserve their signs and significance levels (with the former coefficient increasing to +0.49, from +0.28 in specification (3)). Development stage dummies are now negative and not significant. Although the sign and significance of GDP growth is consistent across linear specifications (1) through (4), the individual country facts presented in Section 2 suggest there might be nonlinear effects. Therefore, we estimate additional models that include a 5th order polynomial in GDP: Prov NLoanct = α + BGDPct + B2 GDPct2 + B3 GDPct3 + B4 GDPct4 + B5 GDPct5 + B1 GDPct−1 + γINTct + γ1 R INTct−1 + M - ct + B- ct + F- c + ξt (5) 15 The results on banking sector penetration and development stage are related to the findings of Loayza and Rancière (2006). 16 F tests run on non-linear specifications also show evidence of country fixed effects, but not of time fixed effects. 17 The order of magnitude of GDP growth and lending rate coefficients under specification (4) is consistent with the findings of the 2003 IMF Financial Soundness Indicators background paper and Glen (2005).

We run a constrained and unconstrained version of (5) and report the results in columns (5) and (6) of Table 3, respectively. The constrained specification assumes B2 = B4 = 0, so as to include only the terms of the polynomial that preserve the sign of GDP. The main results of the unconstrained specification are in general robust to the constrained model. Thus, we only comment on the estimates of the unconstrained model. While the overall statistical fit does not improve much from the linear specification (R2 = 0.71 relative to 0.68 under specification (4)), all polynomial terms are statistically significant. Moreover, the change in signs from the 2nd to the 5th order coefficients implies offsetting effects for some range of GDP growth rates. Lending interest rate effects remain positive and significant, and preserve their second-order magnitude (around +0.03). DomCredPriv/GDP and DomCredProv by Banks/GDP preserve their signs, order of magnitude and significance. The LMI and UMI dummies change sign and remain not significant. The effect of loan loss reserves on loan performance continues to be positive while that of capitalization remains negative. However, both show lower magnitudes relative to specification (4). In sum, country level estimations suggest negative and highly significant effects of economic growth on loan performance, and positive though second-order interest rate effects. In addition, high private sector leverage, poor loan portfolio quality and low banking system capitalization are associated with higher levels of loan loss provisions. On the other hand, non-linear GDP effects do not seem economically important at the banking system level. 4.2. Estimations using bank level data We now present individual bank level estimations, which fully exploit the variation in the data and increase the sample size significantly. However, this comes at the cost of larger volatility in the estimates as well as potential bias due to the structure of the sample in terms of number of banks per country-year (see Table 1 and footnote 9). We follow the same identification strategy as in Section 4.1. Table 4 shows the results of specifications (1) through (6) using bank level data; including standard errors clustered by country and year. The main results documented in Section 4.1 at the banking system level hold for individual banks, although the overall statistical fit is much lower for the latter (R2 ranging from 0.08 to 0.24 across specifications (1)–(6)): GDP growth effects are negative and strongly significant; interest rate effects are positive, second-order and not consistently significant; private sector leverage (as measured by DomCredPriv/GDP) is associated with higher loan loss provisions, while increasing depth of the banking system (as measured by DomCredProv by Banks/GDP) reduces provisions (although the effect is statistically significant only in the unconstrained non-linear specification); poor loan portfolio quality (as measured by LoanLossRes/Loans) is associated with higher provisioning, while increased capitalization is associated with lower levels of provisions. In addition, bank level estimations show a positive and significant effect (under specifications (4) and (5)) of CA/GDP (a measure of external leverage), and a positive (though second-

Table 4 Provisions to gross loans regressions on GDP and lending rates at the individual bank level. Linear models with no fixed effectsa

GDPt

Linear model with country fixed effectsb

Polynomial model including country fixed effectsb

(2)

(3)

(4)

(5)

(6)

−0.4614*** (0.167)

−0.4779*** (0.167)

−0.3758*** (0.125)

−0.4006*** (0.106)

−0.4678*** (0.14)

−0.2202*** (0.081) 0.0111 (0.015) −0.0013 (0.0015) 0.0002*** (0.0001) −0.00001 (0.00001) −0.0267 (0.036) 0.0086 (0.007) 0.0264*** (0.007) −0.0065 (0.019) 0.0343*** (0.011) −0.0297*** (0.009) 0.4339 (0.344) 0.0352 (0.43) 0.2252*** (0.039) 0.0232*** (0.007) −0.0353** (0.014) 1.1074 (0.758) 11,306 0.24

GDPt 2 GDPt 3

0.0025 (0.0024)

GDPt 4 GDPt 5 GDPt −1 INTt REALINTt −1

−0.0140 (0.077) 0.0148 (0.013) 0.0254*** (0.008)

−0.034 (0.065) 0.0222* (0.013) 0.0366*** (0.008) 0.0566** (0.027) 0.0088 (0.012) −0.0087 (0.009) 0.1813 (0.29) −0.1141 (0.417)

4.6859*** (0.752) 11,306 0.08

4.648*** (0.751) 11,306 0.09

CA/GDP DomCredPriv/GDP DomCredProv by Banks/GDP LMI dummy UMI dummy Loan los s res/loans Gloans/total assets Equity/total assets Constant term No. observations R2 a b

0.0565 (0.058) 0.0199* (0.011) 0.0397*** (0.008) 0.0317 (0.023) 0.0062 (0.01) −0.0075 (0.009) 0.4313 (0.341) 0.3331 (0.357) 0.2311*** (0.04) 0.0235*** (0.007) −0.0414** (0.017) 1.1756* (0.613) 11,306 0.20

0.0136 (0.048) 0.0136 (0.012) 0.032** (0.013) 0.0936** (0.041) 0.0571*** (0.022) −0.0228 (0.018) 1.2799* (0.756) 0.8388 (0.863) 0.2365*** (0.042) 0.0242*** (0.007) −0.04** (0.016) 1.1488 (1.254) 11,306 0.21

−0.00002 (0.00001) 0.0155 (0.046) 0.0145 (0.012) 0.0333** (0.013) 0.0986** (0.041) 0.0518*** (0.019) −0.0206 (0.017) 1.262* (0.703) 0.912 (0.815) 0.2354*** (0.042) 0.0245*** (0.007) −0.0397*** (0.015) 0.9198 (1.264) 11,306 0.22

J. Glen, C. Mondragón-Vélez / Review of Development Finance 1 (2011) 150–165

(1)

Standard errors clustered by country and year in parentheses ***p < 1%, **p < 5%, and *p < 10%. Country dummy coefficients not reported.

159

160

J. Glen, C. Mondragón-Vélez / Review of Development Finance 1 (2011) 150–165

order) relationship between bank relative loan portfolio size and loan loss provisions. Linear GDP effects are larger at the individual bank than at the banking system level; a natural consequence of individual bank data aggregation. Lending rate coefficients are positive and of second-order magnitude, although statistical significance is not fully consistent with banking system estimates. While INT is weakly significant only under specifications (2) and (3), the effect of R INT is significant across all specifications for individual banks. In addition, we found no evidence of fixed country or time effects when using the individual bank level sample under the linear or non-linear specifications.18 With regard to the non-linear specifications, the unconstrained non-linear model (specification (6) in Table 4) shows strong statistical significance only for the fourth order term in the polynomial and weak significance for the fifth order term. However, an F test on the joint significance of high-order terms supports the non-linear specification.19 4.3. Comparing linear and non-linear estimations Overall, individual bank level estimations support the main results at the banking system level. In short, these estimations suggest that GDP growth is a key driver of loan portfolio performance. However, the magnitude of GDP growth coefficients differs significantly across specifications and samples. Thus, this section provides a comparison of the relationship between loan loss provisions and GDP growth suggested by the (linear and non-linear) estimations using banking system and individual bank level data presented in Sections 4.1 and 4.2. Fig. 4 summarizes the effects of an absolute (percentage) change in the growth rate of GDP at the banking system (left panel) and individual bank (right panel) levels according to the last three specifications reported in Tables 3 and 4 (linear model, constrained non-linear model and full 5th order polynomial model). First, note that for both samples the non-linear constrained specification generates a similar profile to that of the linear model. However, the magnitude of the linear effect differs across samples. For instance, if GDP growth declines from 6% in year t − 1 to −2% the next year (that is, by 8 percentage points), loan loss provisions would increase by more than 3 percentage points (pp) for individual banks and around 1.75 pp at the banking system level. In other words, assuming the average bank in a specific country had a ratio of provisions to gross loans of 4% and the overall banking system had a ratio of 2% in year t; the linear model predicts loan loss provisions (as a percentage of total loans) will increase to 7% for the average bank and to less than 4% for the entire banking system. As expected, business cycle shocks to individual banks are larger than those observed across banking systems (maintaining other bank characteristics equal).

18 The results of F tests for country fixed effects are supported by the observed R2 across specifications (3) and (4). 19 The corresponding F tests run on the constrained non-linear model (specification 5 in Table 4) do not support the joint statistical significance of high-order terms.

On the other hand, the unconstrained non-linear model shows a relatively flat profile across both samples for changes in GDP growth above −6 pp, and predicts larger effects than those suggested by the linear model only for changes in GDP growth below −8 pp (at the individual bank level) and −10 pp (at the banking system level). Hence, loan loss provisions are relatively linear over the business cycle for GDP fluctuations of less than 6 pp; and will only show exponential deterioration for declines in GDP growth of more than 8 pp. Also note that the unconstrained non-linear profile remains relatively flat for large positive GDP fluctuations across samples. These results suggest that on average loan portfolio performance across major emerging markets attains abnormal deterioration only under extreme economic stress. There is an important caveat associated with the non-linear model. Driven by the small number of episodes with large GDP contractions in the period of analysis, confidence bands tend to widen at extreme GDP growth rate reductions. Therefore, we compute bootstrapped confidence bands for the unconstrained non-linear specification across both samples (reported in column 6 of Tables 3 and 4).20 The left panel of Fig. 5, which corresponds to the unconstrained non-linear estimation at the banking system level, confirms the flatness of the profile for mild fluctuations of GDP growth but provides little support to non-linearity for large economic contractions. The bootstrap estimation at the individual bank level (right panel of Fig. 5) also confirms the flatness of the profile under mild business cycle fluctuations, and provides some evidence in favor of non-linearity under high economic stress. Note that provisions decrease when the economy expands across both samples. In addition, the bootstrap confidence bands suggest that provisions at the individual bank level tend to always increase under economic contraction. This is not necessarily the case at the banking system level, where the profile tends to be flatter and confidence bands wider for negative changes in GDP growth. 4.4. Benchmarking the linear and non-linear models The results presented above suggest provisions are relatively inelastic to mild year-on-year fluctuations of GDP growth of as much as 6 percentage points. On the other hand, overshooting in loan performance deterioration may occur for year-on-year GDP contractions above 8% to 10%. Along these lines, the resilience of the banking sector across most major emerging markets during the past financial crisis comes as no surprise. According to GDP growth and loan loss provision figures for years 2008 and 2009 documented in Appendix A, only three countries in the sample saw a double-digit economic contraction during the 2009 global financial crisis: Romania (−17.9 pp), Russia (−13.1 pp) and Ukraine (−17.2 pp); and not surprisingly, these same countries had the largest increase in loan loss provisions (above

20 The confidence bands shown in Fig. 5 correspond to the 5th and 95th percentiles of the distribution of point estimates at a pre-defined grid of (changes in) GDP growth rates, generated through a bootstrap procedure of 1000 replications of full size random samples (with replacement).

J. Glen, C. Mondragón-Vélez / Review of Development Finance 1 (2011) 150–165

161

Esmaon at the Individual Bank Level

Esmaon at the Banking System Level 20 Change in Provisions (% Gross Loans)

Change in Provisions (% Gross Loans)

Linear model with country effects (4)

15

Linear model with country effects (4)

5th polynomial -odd terms only (5)

5th polynomial -odd terms only (5)

5th polynomial in GDP (6)

5th polynomial in GDP (6)

10

5

0

-5 Change in GDP Growth (%)

Change in GDP Growth (%)

-10 -12

-10

-8

-6

-4

-2

0

2

4

6

8

10

-12

12

-10

-8

-6

-4

-2

0

2

4

6

8

10

12

Fig. 4. Estimated effects of GDP growth on loan loss provisions. Esmaon at the Bank System Level

20

Esmaon at the Individual Bank Level

Change in Provisions (% Gross Loans)

Change in Provisions (% Gross Loans)

15

10

5

0

-5 Change in GDP Growth (%)

-10

Change in GDP Growth (%) -

-12

-10

-8

-6

-4

-2

0

2

4

6

8

10

-12

12

-10

-8

-6

-4

-2

0

2

4

6

8

10

12

Fig. 5. Bootstrap confidence bands for the unconstrained non-linear specification.

1.8 pp) during the crisis. Furthermore, Fig. 6 compares the actual change in loan loss provisions observed during the 2009 global financial crisis, with the linear (left panel) and non-linear (right panel) estimations at the banking system level (based only on observed GDP fluctuations). The linear model slightly overestimates the effect for countries that experienced mild business

15.0

cycle fluctuations, but generates an estimate with a similar order of magnitude for two out of the three countries that experienced significant economic stress. The non-linear model on the other hand, does a better job for most of the countries experiencing mild fluctuations, but shows extreme over-shooting in the case of Russia, Romania and Ukraine (with predicted changes in loan

15.0

2008-2009 Change in Loan Loss Provisions

2008-2009 Change in Loan Loss Provisions

2008-2009 Linear model Predicon

Fig. 6. 2008–2009 actual and predicted change in banking system loan loss provisions.

Vietnam

Turkey

Ukraine

Thailand

Russia

South Africa

Poland

Romania

Peru

Philippines

Nigeria

Pakistan

Mexico

Malaysia

India

Indonesia

Egypt

Colombia

Chile

China

Brazil

Vietnam

Turkey

Ukraine

Thailand

Russia

South Africa

Poland

Romania

Peru

Philippines

Nigeria

Pakistan

Mexico

Malaysia

India

Indonesia

-5.0

Egypt

-5.0

Colombia

0.0

Chile

0.0

China

5.0

Brazil

5.0

Argenna

10.0

Argenna

2008-2009 Non-Linear model Predicon

10.0

162

J. Glen, C. Mondragón-Vélez / Review of Development Finance 1 (2011) 150–165

Table 5 Median gross loans to total assets, return on assets and equity to total assets. 1996

1997

1998

1996

2000

2001

20002

2003

2004

2005

2006

2007

2008

2009

1996–2009

Gross loans to total assets (%) Low income 53.9 Low middle income 53.5 Upper middle income 56.9 Developing countries 55.3

50.0 53.2 56.7 53.7

47.8 54.9 56.9 53.4

42.8 50.5 54.4 49.5

43.5 52.1 52.4 49.4

44.5 51.3 55.9 49.9

44.8 52.3 49.6 49.3

46.5 53.8 52.6 50.6

49.2 57.8 51.3 53.5

51.5 58.1 53.1 56.1

53.6 59.1 53.5 57.6

54.3 61.0 55.0 58.6

56.7 61.9 57.3 60.0

55.6 61.0 55.5 58.2

49.8 57.5 54.7 55.1

ROA (%) Low income Low middle income Upper middle income Developing countries

1.2 1.1 0.9 1.1

1.1 0.9 0.7 0.9

1.3 1.1 1.0 1.1

1.4 0.9 1.0 1.1

1.4 1.0 1.1 1.1

1.3 1.0 1.1 1.1

1.6 1.2 1.2 1.3

1.4 1.4 1.2 1.4

1.6 1.3 1.4 1.4

1.7 1.3 1.5 1.4

1.5 1.3 1.3 1.3

1.4 1.3 1.1 1.3

1.3 10 1.0 1.1

1.4 1.2 1.1 1.2

8.7 12.2 9.9 10.2

9.8 12.3 10.4 10.8

9.3 12.1 10.4 10.9

9.8 12.2 11.3 11.2

9.9 12.3 12.3 11.8

10.0 11.9 12.1 11.4

10.2 14.1 11.1 12.2

10.5 13.6 10.6 12.2

10.4 13.6 10.5 12.4

10.4 13.2 9.6 12.0

11.2 13.3 9.6 12.1

10.9 11.9 10.5 11.2

10.1 12.9 10.5 11.5

1.2 1.3 1.0 1.2

Total equity to total assets (%) Low income 8.9 8.9 Low middle income 10.6 10.9 Upper middle income 9.9 9.6 Developing countries 10.0 9.9

loss provisions above 60 pp for the latter two). This issue is closely related to the wide confidence bands for extreme economic deterioration documented in Section 4.3, in addition to the particularities of foreign currency lending (not captured in the models) across Emerging Europe. Up to now we have presented our estimates in terms of expected changes in loan loss provisions; but what do these mean in terms of profitability and capitalization losses? To answer this question, we use a simple approximation and historical data to generate a measure of the estimated effects on loan loss provisions relative to total assets; so as to make these comparable to standard measures of profitability and capitalization such as returns to total assets (ROA) and total equity to total assets, respectively. We use financial ratios from a broad sample containing more than 19,500 bank-year observations across emerging markets, available from Bankscope for years 1996–2009. The top panel of Table 5 shows the median size of the loan portfolio relative to that of the balance sheet (as measured by the ratio of gross loans to total assets) across developing country groups (low, low middle and upper middle income, as defined by the World Bank) for the period 1996–2009. The medium panel shows median profitability (as measured by return on assets, ROA); and the bottom panel shows median capitalization (as measured by the ratio of total equity to total assets). While the relative size of the loan portfolio varies between 50% and 60% during the period of analysis (with an overall median of 55%), ROA varies between

1.00% and 1.35% for most years; and equity is around 10% to 12% of total assets. The estimated change in loan loss provisions (measured as a percentage of gross loans) documented in Sections 4.1–4.3 is expressed as follows:       Prov Provt Provt+1 Δ − = GLoans t,t+1 GLoanst+1 GLoanst Assuming that balance sheet structure remains around the historical median,   Gloanst+1 ∼ Gloanst ∼ Gloans = = median T.Assetst T.Assets T.Assetst+1 We can express the change in loan loss provisions (measured as a percentage of total assets) as:       Prov Provt+1 GLoanst+1 ∼ Δ × = T.Assets T.Assetst+1 T.Assetst+1     Prov GLoanst − × GLoanst T.Assetst Table 6 applies this expression to compute the approximate change in loan loss provisions in terms of total assets, for various levels of GDP fluctuations using the estimated individual bank effects presented in Section 4.2; and compares these to the historical median profitability and capitalization figures presented in Table 5. The results show that a 6% drop in GDP growth

Table 6 Estimated effects on loan loss provisions to total assets. Year on year change in GDP growth (%) −4.0 −6.0 −8.0 −10.0 −12.0

Estimated change in loan loss provisions (% total assets)

Estimated change in loan loss provisions (% net income)

Estimated change in loan loss provisions (% total equity)

Linear model

Non-linear model

Linear model

Non-linear model

Linear model

Non-linear model

0.9 1.3 1.8 2.2 2.6

0.7 1.3 2.4 4.2 7.2

72 108 144 179 215

54 105 192 341 588

8 11 15 19 23

6 11 20 36 63

J. Glen, C. Mondragón-Vélez / Review of Development Finance 1 (2011) 150–165

generates an increase in loan loss provisions equivalent to the median net income across emerging market banks, and consume about 10% of the median emerging market bank’s capital. On the other hand, a contraction in the growth rate of GDP of 10% or more implies sizable losses for the median bank, which would consume at least one fifth of its capital. 5. Conclusions This paper uses bank level and macroeconomic data to characterize the relationship between bank loan portfolio performance (measured through loan loss provisions) and the business cycle (measured mainly through GDP growth and lending rates), using a sample from major developing economies. We use linear as well as non-linear specifications across banking system aggregates and individual bank level data. Our results show that the main business cycle driver of loan performance is GDP growth, while interest rates have second-order effects. In addition, we find non-linear specifications in GDP to have a slightly better fit to the underlying data. Furthermore, the non-linear relationship implies that only extreme economic stress (GDP growth falling by more than 10 pp) can generate exponential loan performance deterioration. Moreover, the evidence supports non-linear effects under extreme economic contractions more for individual banks than at the banking system level (which show wide confidence bands for large contractions in the growth rate of GDP). In addition, while higher banking system capitalization and penetration are positively correlated with loan performance, high private sector leverage and poor loan portfolio quality are associated with higher loan loss provisions. When comparing the banking system estimates with major emerging markets performance during the 2009 global finan-

163

cial crisis, we find that the linear specification tends to slightly over-estimate the effects for countries with mild economic fluctuations, but generates a similar order of magnitude for loan loss provisions observed in economies under high economic stress. On the other hand, the non-linear specification provides better estimates for countries with mild fluctuations, but tends to overshoot for countries that experienced extreme economic contraction. Finally, we use historical banking ratios across a broader sample of emerging markets to frame our results for individual banks relative to standard measures of banking profitability and capitalization. We find that while a (year-on-year) drop in economic growth of 6% is equivalent to the profitability of the median emerging market bank, a drop of 10% (or more) implies significant losses that would consume at least one fifth of the median bank’s capital. Given the increasing role of emerging markets in the global economy, understanding the dynamics of the banking sector in the developing world is important. In this regard, additional research along the lines of simplified approaches to support the development of top-down bank stress testing and risk analysis frameworks in these countries is critical as these approaches complement more comprehensive (bottom-up) models that require detailed knowledge and information about the underlying loan portfolios of banks, which are not always available to investors and analysts worldwide.

Appendix A. Provisions to gross loans, real GDP growth and lending interest rates Tables A1–A3.

Table A1 Loan loss provisions to gross loans (%).

Argentina Brazil Chile China Colombia Egypt India Indonesia Malaysia Mexico Nigeria Pakistan Peru Philippines Poland Romania Russia South Africa Thailand Turkey Ukraine Vietnam

1996

1997

1998

1999

2000

2001

20002

2003

2004

2005

2006

2007

2008

2009

4.1 6.2 0.5 0.3 2.2 3.0 1.0 0.2 0.6 1.6 1.6 3.1 2.5 0.6 0.7 1.6 4.1 1.1 1.6 1.7 1.7

2.0 4.6 0.7 0.9 2.3 2.0 1.3 1.9 2.0 2.5 1.4 0.6 2.8 1.1 1.2 7.7 2.7 1.0 4.7 3.2 6.0

2.1 6.5 1.3 1.2 3.3 1.2 1.2 41.0 3.8 2.4 3.4 3.1 2.7 3.8 0.8 12.6 5.0 1.7 6.2 0.8 14.6

2.7 6.2 1.7 0.8 5.1 1.1 1.4 36.1 2.4 10.1 5.4 1.5 3.6 3.2 0.6 5.6 1.9 1.3 7.4 2.1 2.2

2.4 2.2 0.9 0.8 3.6 1.6 1.8 2.4 1.5 1.8 3.6 1.6 3.1 1.8 1.2 4.2 0.4 1.1 0.4 1.6 3.8 3.4

4.0 2.8 1.1 0.6 2.1 1.1 1.7 3.8 2.0 2.4 4.5 1.5 2.4 1.1 3.1 2.3 0.8 1.4 −0.3 11.0 2.9 3.0

11.2 4.0 1.2 0.9 1.5 1.1 1.8 1.7 1.4 2.0 3.8 2.3 2.1 1.3

−0.9 3.3 0.9 0.8 1.6 1.7 1.9 1.7 1.0 1.6 2.2 1.4 1.5 1.1 0.4 1.4 1.3 1.8 1.0 2.6 1.8 1.0

−0.3 2.9 0.8 0.8 1.3 2.3 0.6 0.9 0.9 1.2 2.2 0.7 1.2 1.4 1.0 1.3 1.3 0.7 0.7 0.9 1.3 0.8

−0.4 3.7 0.6 0.7 1.4 4.2 0.4 1.1 1.0 1.6 1.7 0.9 1.1 1.3 0.2 1.3 1.9 0.5 0.5 0.9 1.2 1.4

−0.4 4.5 0.8 0.7 1.5 3.9 0.5 1.7 1.0 2.1 1.6 0.8 1.0 1.7 0.2 0.5 1.8 0.6 1.3 0.7 1.6 0.4

1.0 3.6 7.9 0.6 3.1 2.2 0.6 1.3 0.9 3.4 1.6 2.0 1.2 0.8 0.1 0.6 1.4 0.8 1.7 1.0 1.2 1.0

1.6 3.8 1.4 0.9 4.0 4.9 0.6 1.7 0.6 5.4 8.1 2.6 1.5 0.5 0.8 1.3 2.1 1.4 1.0 1.4 4.4 1.4

2.1 4.8 1.8 0.5 4.4 0.7 0.7 2.0 0.7 5.1 7.2 2.3 1.7 0.9 2.0 3.1 5.8 1.8 0.9 2.2 10.1 0.5

Source: Bankscope.

0.7 1.3 4.5 1.3 1.9 1.6 1.2

164

J. Glen, C. Mondragón-Vélez / Review of Development Finance 1 (2011) 150–165

Table A2 Real GDP growth (%).

Argentina Brazil Chile China Colombia Egypt India Indonesia Malaysia Mexico Nigeria Pakistan Peru Philippines Poland Romania Russia South Africa Thailand Turkey Ukraine Vietnam

1996

1997

1998

1999

2000

2001

20002

2003

2004

2005

2006

2007

2008

2009

5.5 2.2 7.4 10.0 2.1 5.0 7.6 7.6 10.0 5.1 4.3 4.9 2.5 5.9 6.2 4.0 −3.6 4.3 5.9 7.4 −10.0

8.1 3.4 6.6 9.3 3.4 5.5 4.1 4.7 7.3 6.8 2.7 1.0 6.9 5.2 7.1 −6.1 1.4 2.7 −1.4 7.6 −3.0

3.9 0.0 3.2 7.8 0.6 4.0 6.2 −13.1 −7.4 4.9 1.9 2.6 −0.7 −0.6 5.0 −4.8 −5.3 0.5 −10.5 2.3 −1.9

−3.4 0.3 −0.8 7.6 −4.2 6.1 7.4 0.8 6.1 3.9 1.1 3.7 0.9 3.4 4.5 −1.2 6.4 2.4 4.5 −3.4 −0.2

−0.8 4.3 4.5 8.4 4.4 5.4 4.0 4.9 8.9 6.6 5.4 4.3 3.0 6.0 4.3 2.1 10.0 4.2 4.8 6.8 5.9 6.8

−4.4 1.3 3.4 8.3 1.7 3.5 5.2 3.6 0.5 −0.2 3.1 2.0 0.2 1.8 1.2 5.7 5.1 2.7 2.2 −5.7 9.2 6.9

−10.9 2.7 2.2 9.1 2.5 2.4 3.8 4.5 5.4 0.8 1.6 3.2 5.0 4.5 5.1 4.7 3.7 5.3 6.2 5.2 7.1

8.8 1.2 3.9 10.0 3.9 3.2 8.4 4.8 5.8 1.4 10.3 4.9 4.0 4.9 3.9 5.2 7.3 3.0 7.1 5.3 9.4 7.3

9.0 5.7 6.0 10.1 5.3 4.1 8.3 5.0 6.8 4.1 10.6 7.4 5.0 6.4 5.3 8.4 7.2 4.6 6.3 9.4 12.1 7.8

9.2 3.2 5.6 11.3 4.7 4.5 9.3 5.7 5.3 3.2 5.4 7.7 6.8 5.0 3.6 4.2 6.4 5.3 4.6 8.4 2.7 8.4

8.5 4.0 4.6 12.7 6.7 6.8 9.4 5.5 5.9 4.9 6.2 6.2 7.7 5.3 6.2 7.9 8.2 5.6 5.2 6.9 7.3 8.2

8.7 6.1 4.6 14.2 6.9 7.1 9.6 6.3 6.5 3.3 6.5 5.7 8.9 7.1 6.8 6.0 8.5 5.5 4.9 4.7 7.9 8.5

6.8 5.1 3.7 9.6 2.7 7.2 5.1 6.0 4.7 1.5 6.0 1.6 9.8 3.7 5.0 9.4 5.2 3.7 2.5 0.7 2.1 6.3

0.9 −0.2 −1.5 9.1 0.8 4.7 7.7 4.6 −1.7 −6.5 5.6 3.6 0.9 1.1 1.7 −8.5 −7.9 −1.8 −2.3 −4.7 −15.1 5.3

Source: World Bank.

Table A3 Lending interest rate (%).

Argentina Brazil Chile China Colombia Egypt India Indonesia Malaysia Mexico Nigeria Pakistan Peru Philippines Poland Romania Russia South Africa Thailand Turkey Ukraine Vietnam

1996

1997

1998

1999

2000

2001

20002

2003

2004

2005

2006

2007

2008

2009

10.5 93.3 17.4 10.1 42.0 15.6 16.0 19.2 9.9 36.4 19.8 14.5 31.5 14.8 26.1 55.1 146.8 19.5 13.4 99.2 79.9

9.2 78.2 15.7 8.6 34.2 13.8 13.8 21.8 10.6 22.1 17.8 15.1 30.9 16.3 25.2 72.5 32.0 20.0 13.7 99.4 49.1

10.6 86.4 20.2 6.4 42.2 13.0 13.5 32.2 12.1 26.4 18.2 16.0 32.6 16.8 24.5 55.3 41.8 21.8 14.4 79.5 54.5

11.0 80.4 12.6 5.9 25.8 13.0 12.5 27.7 8.6 23.7 20.3 15.0 35.1 11.8 16.9 65.6 39.7 18.0 9.0 86.1 55.0

11.1 56.8 14.8 5.9 18.8 13.2 12.3 18.5 7.7 16.9 21.3 13.7 30.0 10.9 20.0 53.9 24.4 14.5 7.8 51.2 41.5 10.6

27.7 57.6 11.9 5.9 20.7 13.3 12.1 18.6 7.1 12.8 23.4 13.8 25.0 12.4 18.4 45.4 17.9 13.8 7.3 78.8 32.3 9.4

51.7 62.9 7.8 5.3 16.3 13.8 11.9 18.9 6.5 8.2 24.8 12.0 20.8 9.1

19.2 67.1 6.2 5.3 15.2 13.5 11.5 16.9 6.3 7.0 20.7 7.9 21.0 9.5 7.3 25.4 13.0 15.0 5.9 42.8 17.9 9.5

6.8 54.9 5.1 5.6 15.1 13.4 10.9 14.1 6.1 7.4 19.2 7.3 24.7 10.1 7.6 25.6 11.4 11.3 5.5 29.1 17.4 9.7

6.2 55.4 6.7 5.6 14.6 13.1 10.8 14.1 6.0 9.7 18.0 9.1 25.5 10.2 6.8 19.6 10.7 10.6 5.8 23.8 16.2 11.0

8.6 50.8 8.0 6.1 12.9 12.6 11.2 16.0 6.5 7.5 16.9 11.0 23.9 9.8 5.5 14.0 10.4 11.2 7.4 19.0 15.2 11.2

11.1 43.7 8.7 7.5 15.4 12.5 13.0 13.9 6.4 7.6 16.9 11.8 22.9 8.7

19.5 47.3 13.3 5.3 17.2 12.3 13.3 13.6 6.1 8.7 15.5 12.9 23.7 8.8

15.7 44.7 7.3 5.3 13.0 12.0 12.2 14.5 5.1 7.1 18.4 14.5 21.0 8.6

13.4 10.0 13.2 7.1 20.1 13.9 11.2

15.0 12.2 15.1 7.0

17.3 15.3 11.7 6.0

17.5 15.8

20.9

35.4 15.7 15.8 6.9 53.7 25.4 9.1

Source: IMF (and central bank or national sources for selected countries).

References Altman, E., Bharath, S., Saunders, A., 2002. Credit ratings and the BIS capital adequacy reform agenda. Journal of Banking & Finance Volume 26 (5), 909–921. Arpa, M., Giulini, I., Ittner, A., Pauer, F., 2001. The influence of macroeconomic developments on Austrian banks: implications for banking supervision. Bank of International Settlements Papers 1, 91–116.

Babouˇcek, I., Janˇcar, M., 2005. Effects of Macroeconomic Shocks to the Quality of the Aggregate Loan Portfolio. Czech National Bank Research Department, working papers 2005/01. Gerlach, S., Peng, W., Shu, C., 2005. Macroeconomic conditions and banking performance in Hong Kong SAR: a panel data study. In: Investigating the Relationship between the Financial and Real Economy, vol. 22. Bank for International Settlements, pp. 481– 497.

J. Glen, C. Mondragón-Vélez / Review of Development Finance 1 (2011) 150–165 Glen, J., 2005. Corporate Restructuring: Lessons from Experience. In: Pomerleano, M., Shaw, W. (Eds.), Debt and firm vulnerability. The World Bank, pp. 237–259. Hoggarth, G., Logan, A., Zicchino, L., 2005. Macro stress tests of UK banks. In: Investigating the Relationship between the Financial and Real Economy, vol. 22. Bank for International Settlements, pp. 392–408. International Monetary Fund, 2010. Global Financial Stability Report. International Monetary Fund, Washington, DC. International Monetary Fund, 2003. Financial Soundness Indicators— Background Paper. International Monetary Fund, Washington, DC. Laeven, L., Valencia, F., 2008. Systemic Banking Crises: A New Database, IMF Working Papers, 08/224, pp. 1–78. Llaudes, R., Salman, F., Chivakul, M., 2010. The Impact of the Great Recession on Emerging Markets, IMF Working Paper 10/237. Loayza, N., Rancière, R., 2006. Financial development, financial fragility and growth. Journal of Money, Credit and Banking 38 (4), 1051–1076.

165

Padilla, P., Segoviano, M., 2006. Portfolio Credit risk and Macroeconomic Shocks: Applications to Stress Testing under Data Restricted Environments, International Monetary Fund Working Paper 06/283. Pesaran, M., Schuermann, T., Treutler, B., Weiner, S., 2006. Macroeconomic dynamics and credit risk: a global perspective. Journal of Money, Credit and Banking 38 (5), 1211–1261. Quagliariello, M., 2004. Banks’ Performance over the Business Cycle: A Panel Analysis on Italian Intermediaries. University of York, Department of Economics Discussion Papers 04/17. Segoviano, M., 2006a. The Conditional Probability of Default Methodology. Financial Markets Group, London School of Economics, Discussion Paper 558. Segoviano, M., 2006b. The Consistent Information Multivariate Density Optimizing Methodology. Financial Markets Group, London School of Economics, Discussion Paper 557.