The IMF Financial Soundness Indicators as
predictors of the 2008 banking crisis
Bachelor’s thesis Adriaan Willems 10119035
adriaan.willems@student.uva.nl
Faculty of Economics and Business Economics and Finance
Supervisor: Eglė Jakučionytė February 16, 2017
2
Statement of Originality
This document is written by Adriaan Willems who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is
original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
3
Table of contents
1 – Introduction 4
2 – Literature review 6
2.1 – Defining banking crises 6
2.2 – General macroeconomic indicators of banking crises 7
2.3 – IMF Financial Soundness Indicators 12
2.4 – Conclusion of literature review 15
3 – Methodology 16
3.1 – Sample 16
3.2 – Measure of crisis 17
3.3 – FSIs 17
3.4 – Macroeconomic control variables 19
3.5 – Model specification 20 3.6 – Heteroscedasticity 22 3.7 – Serial correlation 22 3.8 – Model variations 23 4 – Results 24 5 – Conclusion 28 Bibliography 30 Appendices
Appendix 1: Countries included in the dataset 34
Appendix 2: Overview of systemic banking crises 35
Appendix 3: Description of FSIs 36
Appendix 4: Data sources for macroeconomic variables 37
4
1 - Introduction
The 2008 financial crisis that raged through North America and, subsequently, Europe has had profound implications on long-held financial beliefs and theories on how regulation of finance should be structured, and has led to a vibrant debate among institutions, academics and practitioners (Caprio et al., 2014). However, the question of which policies to implement cannot be properly addressed without analyzing the determinants of the financial crisis of 2008 (Caprio et al., 2014). Aside from being useful in shaping future financial policy and structure, a better understanding of these determinants can lead to better predictions of future crises (Frankel & Saravelos, 2012). At the 2009 G20 group of nations summit in London, a call was issued to the International Monetary Fund, the IMF, “to provide early warning of macroeconomic and financial risks and the actions needed to address them.” (Frankel & Saravelos, 2012). In response to this call, the IMF launched a new online database of so-called Financial Soundness Indicators, or FSIs (IMF press release, 2009), intending to support macro-prudential analysis and financial sector surveillance.
Previous research into the effectiveness of the IMF FSI indicators exists, utilizing a dataset of crisis dummies indicating whether a country is experiencing a systemic banking crisis compiled and maintained by Laeven and Valencia (2012), but results have varied (Čihák & Schaeck, 2007; Costa Navajas & Thegeya, 2013). This thesis adds to the existing literature by updating the dataset of crisis dummies, originally compiled by Laeven and Valencia (2012), to the fourth quarter of 2015, as well as by examining the IMF FSIs using quarterly data as opposed to yearly data, which was used in previous research (Čihák & Schaeck, 2007; Costa Navajas & Thegeya, 2013). Following previous literature that investigates potential predictors of banking crises (Demirgüç-Kunt & Detragiache, 1998; Čihák & Schaeck, 2007; Costa Navajas & Thegeya, 2013), this thesis utilizes a multivariate pooled logit model to ascertain which FSIs are correlated with the probability of a country experiencing a systemic banking crisis at a given time. The following research question is answered in this thesis: are the IMF Financial Soundness Indicators able to predict the 2008 financial crisis?
A selection of six FSIs is examined, and the return on equity variable is found to have a negative coefficient that is robust to lag structures, which could mean return on equity is a leading indicator of the 2008 banking crisis. This result mirrors findings by previous literature, although in one case the coefficient found for the return on equity variable, when lagged, is positive, thus contradicting the coefficient found in this thesis (Čihák & Schaeck, 2007; Costa Navajas & Thegeya, 2013). The regulatory capital to risk-weighted assets is found to have a positive coefficient. However, this coefficient loses its significance when introduced to different lag structures, therefore implying
5
that the capital to risk-weighted assets variable is not a useful leading indicator of the 2008 banking crisis.
This thesis is structured as follows: in chapter 2, a literature review is conducted. In chapter 3, the research methodology used in this thesis is discussed. In chapter 4, the results of various model estimations are discussed. In chapter 5, a conclusion is made.
6
2 - Literature review
Literature concerning the following topics will be discussed in this literature review. Firstly, various definitions of banking crises will be discussed. Secondly, literature examining the usefulness of broad macroeconomic variables as indicators of banking crises will be discussed. Thirdly, the IMF Financial Soundness Indicators and research into their usefulness will be discussed. And lastly, a short conclusion will be made.
2.1 - Defining banking crises
According to Frankel and Saravelos (2012, p.p. 3-4), the definitions of a financial crisis and the severity of incidence vary wildly, with past literature using both discrete and continuous measures to define a crisis. Another problem in defining a banking crisis is that differently defined crises tend to occur more and more simultaneously in recent years: Bordo et al. (2001) conclude that so-called twin crises, where a banking crisis and a currency crisis occur simultaneously, have become much more commonplace in recent years, as well as banking crises occurring more and more frequently in recent years. Surveying 21 countries, they find only one banking crisis in the 25 years after 1945, but as many as 19 since 1970 (Bordo et al., 2001). Frankel and Rose (1996, p. 2) define a currency crash dummy variable as a depreciation of the nominal exchange rate of at least 25 per cent that is also at least a 10 per cent increase in the rate of nominal depreciation from the previous year.
Frankel and Saravelos (2012) distinguish between two approaches in defining a financial crisis. In their literature review, they (Frankel & Saravelos, 2012) find that literature uses both discrete and continuous measures to define a crisis. Discrete measures take on the form of binary variables, which define a crisis as occurring when a certain threshold value of a chosen economic or financial variable has been breached (Frankel & Saravelos, 2012; see also Demirgüç-Kunt &
Detragiache, 1998; Mehrez & Kaufman, 2000; Beck, Demirgüç-Kunt & Levine, 2006; Čihák & Schaeck, 2007; Costa Navajas & Thegeya, 2013; Caprio et al., 2014). Other literature measures crisis intensity on a continuous scale, utilizing measures such as drops in GDP, drops in the equity market, exchange rates, and speculative pressure indices (Frankel & Saravelos, 2012; see also Rose & Spiegel, 2010, 2011, 2012; Claessens et al., 2010; Lane & Milesi-Ferretti, 2011). Some literature uses more exotic continuous measures, such as the difference between 2008 GDP forecasts and actual GDP (Berkmen et al., 2012).
According to Laeven and Valencia (2010, p. 3), banking crises, although having differed in terms of underlying causes, triggers, and economic impact, share many commonalities. They find that more recent banking crises have a significantly larger economic cost than past crises, both in terms of output losses and increases in public debt, and conclude that these differences reflect an
7
increase in the size of financial systems and the fact that recent crises are concentrated in high-income countries (Laeven & Valencia, 2010). Building on earlier work by Caprio et al. (2005), Laeven and Valencia (2010) compose a database, stretching from 1970 to 2009, of dummy variables indicating whether a country is in a banking crisis at a specified time. A banking crisis is considered systemic if two conditions are met: there are significant signs of financial distress in the banking system, as indicated by bank runs, losses in the banking system, and bank liquidations; and there are significant banking policy intervention measures in response to significant losses in the banking system (Laeven & Valencia, 2012, p. 6). They consider the policy interventions in the banking sector to be significant if at least three out of the following six measures occur: extensive liquidity support of at least 5 per cent of deposits and liabilities to nonresidents; at least 3 per cent of GDP in bank restructuring costs; significant bank nationalizations; significant guarantees put in place; at least 5 per cent of GDP of significant bank purchases; and deposit freezes and bank holidays (Laeven & Valencia, 2012 p.p. 6-7). In addition, a list of so-called borderline cases is maintained of incidences that almost meet these criteria, but still show clear signs of distress in the banking sector (Laeven & Valencia, 2012, p. 8). Laeven & Valencia (2012, p. 10) define the end of a banking crisis as the year before two conditions hold: real GDP growth and real credit growth are positive for at least two consecutive years. All crises durations are also truncated at five years (Laeven & Valencia, 2012, p. 10).
2.2 - General macroeconomic indicators of banking crises
Research into the determinants of banking crises has been ongoing for some time. Results of some studies of macroeconomic determinants dated before the 2008 banking crisis are briefly discussed below, followed by a discussion of literature dated after the 2008 crisis.
Demirgüç-Kunt and Detragiache (1998), using a multivariate logit model, attempt to link macroeconomic environment factors to banking crises that occurred between 1980 and 1994. They find that crises tend to erupt when the macroeconomic environment is weak, in particular when GDP growth is low and inflation is high (Demirgüç-Kunt & Detragiache, 1998). They also find that high real interest rates are associated with systemic banking sector problems, as well as vulnerabilities in the balance of payments (Demirgüç-Kunt & Detragiache, 1998).
Mehrez and Kaufman (2000) investigate how transparency affects the probability of a country experiencing a financial crisis. After constructing a model in which banks cannot distinguish between aggregate shocks, government policy, and firms’ quality, they (Mehrez & Kaufman, 2000) then run a multivariate probit model using data on 56 countries from 1977 to 1997. Mehrez and
8
Kaufman (2000) conclude that when a lack of transparency in government policy occurs, the probability of experiencing a financial crisis is higher, due to banks having incentives to raise credit above the optimal level.
Using data on 69 countries from 1980 to 1997, Beck, Demirgüç-Kunt and Levine (2006) investigate the relationship between bank concentration, banking competition and national
institutions on one side and the likelihood of a country experiencing a systemic banking crisis on the other side. They conclude that countries with more concentrated banking sectors are less likely to experience banking crises (Beck, Demirgüç-Kunt & Levine, 2006). When examining the
macroeconomic control variables used in this study, Beck, Demirgüç-Kunt and Levine (2006)
corroborate some of the results found by Demirgüç-Kunt and Detragiache (1998): real interest rates as well as GDP growth are found to be correlated with the probability of a country experiencing a banking crisis.
Rose and Spiegel (2010, 2011, and 2012) conducted several studies of macroeconomic predictors of the 2008 banking crisis. In their earliest paper, which was actually published last, Rose and Spiegel (2012) attempt to model the causes of the 2008 banking crisis together with its
manifestations, using a Multiple Indicator Multiple Cause, or MIMIC, model. Using cross-sectional data of 107 countries, Rose and Spiegel (2012) initially conclude that only one of the tested macroeconomic predictors, the size of the equity market run-up prior to the crisis, is a robust predictor of crisis severity. Contrary to intuition, they (Rose & Spiegel, 2012) find that all other macroeconomic fundamentals, including regulatory framework, financial conditions, and macroeconomic, institutional and geographic features of countries, are not robustly significant predictors of crisis severity. A possible explanation for these weak results is the poor availability of measures of crisis incidence, since the data used in this study was collected in early 2009, possibly not adequately capturing the full extent of the financial crisis (Rose & Spiegel, 2012). Another
explanation might be that Rose and Spiegel (2012) do not consider contagion effects due to exposure to the US. In a second study, Rose and Spiegel (2010) include contagion effects through financial and real linkages, such as trade linkage, foreign asset exposure, and sudden stops in international credit. However, Rose and Spiegel (2010, p. 22) are unable to find strong evidence of contagion effects, with countries that held a disproportionate amount of American securities seemingly experiencing lower crisis severity. They conclude that they are skeptical of the ability of early warning systems with respect to the occurrence of international financial crises (Rose & Spiegel, 2010, p. 22). Updating their work another time, Rose and Spiegel (2011) replace the previously used non-linear MIMIC model with an ordinary least squares model as it has been favored by other literature, add new
9
measures of crisis severity, and extend their dataset to include data from the year 2009, but arrive at the same conclusions as they have in previous studies.
In their study of macroeconomic determinants of the 2008 banking crisis, Claessens et al. (2010) group countries based on the time they were first affected by the crisis, thereby hoping to capture different transmission channels, such as national imbalances, financial exposure to the US, and trade linkage. They argue that different transmission channels are reflected in the timing of being affected by the crisis, and group countries by the time they were first affected by the crisis (Claessens et al., 2010). Three measures of economic performance during the crisis are used: crisis duration, if a decline in GDP occurred, the severity of income loss following the crisis and the change in average growth rate in the crisis years compared to pre-crisis periods (Claessens et al., 2010, p. 282). They find that only a few of the initial conditions have statistically significant and robust relationships with their performance measures, with house price appreciation, bank credit growth prior to the crisis, and the size of the current account deficit being significant predictors for all three performance indicators (Claessens et al., 2010). Trade openness, measured by exports plus imports to GDP, is found to be positively associated with two out of three performance measures (Claessens et al., 2010, p. 285). However, Claessens et al. (2010, p. 285) conclude that their measure for financial sector development, the private credit to GDP ratio, and their measures for mortgage market development and wholesale funding dependence are not significant for any of the economic performance measures. According to Claessens et al. (2010, p. 285), these results imply that the level of financial deepening or the structure of the financial system does not cause increased vulnerability to the banking crisis, but rather the occurrence of rapid financial deepening combined with sharp rises in asset prices. When trying to explain financial sector performance of individual countries, Claessens et al. (2010, p. 286) find very few variables to be statistically significant. The ratio of private credit to GDP is found to be positively related to an increase in the financial stress indicator of countries (Claessens et al., 2010, p. 286). They conclude that most other variables, including credit growth, wholesale funding dependence, and foreign bank exposure are not significantly related to the financial stress indicator, which they argue implies that the financial shocks of the banking crisis were global and systemic in nature, affecting all financial markets more or less equally (Claessens et al., 2010, p.286).
Lane and Milesi-Ferretti (2011) find much stronger evidence in their cross-country survey. They find a strong link between pre-crisis domestic financial factors, such as private credit growth, and external imbalances, such as current account deficits, on one side and the decline in growth rate of economic output and domestic demand during the crisis on the other side (Lane & Milesi-Ferretti, 2011). They also find that variables such as trade openness and the manufacturing share are
10
correlated with the declines of output and domestic demand during the crisis, although the
conditional correlations are not significant in all cases (Lane & Ferretti, 2011). Lane and Milesi-Ferretti (2011) emphasize that the “advanced economies nature” of the 2008 banking crisis is highlighted by the negative correlation between GDP per capita and decline in output growth that is found. They also find weak evidence that countries with pegged exchange rate regimes experienced weaker output growth during the banking crisis (Lane & Milesi-Ferretti, 2011).
Berkmen et al. (2012) use the difference between GDP forecasts prior to the banking crisis and actual GDP growth in 2009 as the dependent variable in their study. Panel data on 43 countries was used, including both emerging and developing economies (Berkmen et al., 2012). Using a wide range of explanatory variables, Berkmen et al. (2012, p. 44) attempt to capture different transmission mechanisms: trade linkages, financial linkages, underlying vulnerabilities and financial structure, and the overall policy framework. To capture the various trade channels, Berkmen et al. (2012, p. 44) used three groups of variables such as trade openness variables (exports to GDP and exports plus imports to GDP), trade composition variables (share of commodities and manufactured products in total exports), and direction of trade variables (the share of trade with advanced economies). The financial channel is reflected by variables such as financial openness, capital account restrictions, and the stock of bank lending from advanced economies (Berkmen et al., 2012, p. 44). The channel of underlying vulnerability is captured by variables such as national savings to GDP, an overall balance to GDP, inflation, and the Milken Institute Opacity Index (Berkmen et al., 2012). Lastly, the policy framework channel is captured by variables such as a structural government balance, a degree of exchange-rate flexibility and a primary gap (Berkmen et al., 2012). Berkmen et al. (2012) conclude that the main avenue of transmission is the financial channel, particularly through high leverage and short-term debt, and that the transmission effects are stronger when countries use pegged exchange rates and in emerging economies. Berkmen et al. (2012) also conclude that, when specifically looking at emerging economies, the trade channel plays a significant role. Other significant explaining variables include primary fiscal gaps, current account imbalances, and public debt (Berkmen et al., 2012).
Frankel and Saravelos (2012) conduct a literature review of more than 80 different studies on early warning indicators. The primary focus of their literature review is to identify the causes and symptoms of financial crises that are the most consistent across different timeframes, countries, and crises (Frankel & Saravelos, 2012). They note several problems in some of the literature discussed earlier in this thesis (Frankel & Saravelos, 2012). Frankel & Saravelos (2012, p. 11) argue that the timeframe used by Rose and Spiegel in their earliest study (2012) is inappropriate, since crisis incidence was measured over the 2008 calendar year, even though the crisis did not become severe
11
until September 2008, and global output and financial markets continued to contract sharply well into 2009. These accuracy problems persist in the follow-up study done by Rose & Spiegel (2011), where they update the data sample to include the year 2009, because annual data is used, resulting in more imprecise estimations (Frankel & Saravelos, 2012, p. 11). Similar criticism is directed at Lane and Milesi-Ferretti (2011), who use annual data as well (Frankel & Saravelos, 2012, p. 12). Frankel & Saravelos (2012) also criticize Berkmen et al. (2012) for only considering forecast differences, not actual performance. After reviewing literature, Frankel and Saravelos (2012) conduct a regression analysis of their own, using 50 annual macroeconomic and financial variables. All independent variables are dated from 2007 or earlier, minimizing endogeneity issues (Frankel & Saravelos, 2012, p. 13). The following crisis measures are used: a nominal local currency percentage change versus the US dollar, equity market returns, the percentage change of real GDP between the second quarter of 2008 and the second quarter of 2009, a percentage change in industrial production, and a dummy variable indicating whether a country has requested funds from the IMF. They conclude that
international reserves and real exchange rate overvaluation are useful leading indicators of the 2008 banking crisis, with international reserves being robust to a number of different crisis incidence definitions as well as to the addition of more independent variables, when using exchange market pressure index as a measure of crisis severity (Frankel & Saravelos, 2012). A number of other variables is concluded to be useful as leading indicators during the 2008 banking crisis as well, although the robustness of these variables across different crisis incidence measures was not as strong (Frankel & Saravelos, 2012). Lower past credit growth, lower external debt and short term debt, and larger current accounts were found to be indicative of lower crisis incidence (Frankel & Saravelos, 2012).
Caprio et al. (2014) run a cross-country regression on 13 potential determinants of the probability that a country experiences a crisis during 2008, as reported by the database compiled by Laeven and Valencia (2010). Primarily focusing on banking indicators, they find that countries with a higher level of net interest margin, a higher level of concentration in the banking sector, more restrictions on banking activities, and/or a higher level of private monitoring had a lower probability of being in crisis during 2008 (Caprio et al., 2014). Countries with a higher credit to deposit ratio are found to be more likely to experience a crisis during 2008 (Caprio et al., 2014). Caprio et al. (2014) argue that these results imply that more traditional banking systems had a lower probability to be in crisis in 2008. A higher level of interest margin is interpreted as representing a stronger incentive for banks to undertake traditional banking activity, e.g. loans, instead of riskier non-traditional activities, e.g. trading in securities (Caprio et al., 2014). They recommend ending broader regulatory framework that boosts the rewards of securitization, and that measures concerning stricter capital requirements
12
are to be considered carefully, as it is argued that these measures are not a cure-all (Caprio et al., 2014).
Table 1: Significant explaining variables found in various studies.
Significant explanatory variables
De m ir gü ç-Ku n t an d De tra giach e (19 98 ) Be ck, De m irgü ç-Ku n t an d L ev in e (2006) Cl ae ss en s et al. (2 010) La n e a n d M ile si -Fe rre tt i ( 2011 ) Be rk m en et al. (20 12) Fran kel a n d Sa ra ve los (2012 ) Real GDP growth * * GDP/capita * * *
Real interest rate * *
Inflation * *
Real exchange rate * * *
Imports *
Trade to GDP * *
Trade openness *
Change in terms of trade *
Current account balance * * *
Bank deposits/GDP *
Private credit growth * * * * *
Domestic credit to private sector/GDP *
House prices *
Region or country dummy * * *
Industry concentration *
2.3 - IMF Financial Soundness Indicators
The IMF Financial Soundness Indicators (FSIs) are aggregate measures of a country’s financial sector, intended to add to a regulatory authority’s macro-prudential surveillance toolkit, dating back to 1999 when the IMF and the World Bank launched the Financial Sector Assessment Program (FSAP) to monitor the fragility of financial systems (Costa Navajas & Thegeya, 2013). Although data on several FSIs was already collected as a part of the FSAP, inconsistency in the methodology of data collection was a hindrance to cross-country comparisons (Costa Navajas & Thegeya, 2013). In 2000, the IMF, in collaboration with the International Accounting Standards Board, the Bank for International
Settlements, the Basel Committee for Banking Supervision, and other international organizations, drafted a guide intended to arrive at a single, uniform methodology for the compilation of FSIs (Costa
13
Navajas & Thegeya, 2013; see also IMF, 2000; IMF, 2007). A full set of guidelines was published in the IMF FSI Compilation Guide (2006). An amended set of guidelines was published in 2008 (IMF, 2008). FSIs are split into two sets: a set of 12 core indicators and a set of 28 encouraged indicators (Costa Navajas & Thegeya, 2013). According to Costa Navajas & Thegeya (2013, p. 6), the fundamental value of the FSIs is their ability to potentially indicate overall distress within a banking sector, and it should therefore be possible for FSIs to not only be useful tools for monitoring banking sector health, but also to be useful as leading indicators for the incidence of banking crises.
The first research into the effectiveness of the IMF FSIs as predictors of banking crises was done by Čihák and Schaeck (2007), using a dataset consisting of observations from 100 countries spanning the period between 1994 and 2004. The dependent variable, a dummy variable indicating whether a country is experiencing a systemic banking crisis at a specific time, was compiled from two different datasets (Čihák & Schaeck, 2007, p. 12): a survey of systemic banking crises by Demirgüç-Kunt (2005), which uses criteria very similar to Laeven and Valencia (2012); and a database compiled by Honohan and Laeven (2005), which has since been updated by Laeven and Valencia (2012). Using a multivariate logit model, they (Čihák & Schaeck, 2007) find weak evidence that the Capital Asset Ratio and the Non-performing Loans Ratio signal systemic banking problems, as well as evidence that the Return on Equity of financial institutions and corporate leverage are good indicators of a build up towards systemic banking problems. They also corroborate the hypothesis that banking crises tend to occur in countries that are vulnerable to capital outflows and in less developed economies (Čihák & Schaeck, 2007). A limitation of this study is that not all countries that reported FSI data adhered to the standards laid out in the IMF Compilation Guide on Financial Soundness Indicators (2006), which means that, in this case, some FSI’s are not strictly comparable across countries (Čihák & Schaeck, 2007, p. 5; Costa Navajas & Thegeya, 2013, p. 4).
Babihuga (2007) explores the macroeconomic determinants of several FSI’s and finds that these FSI’s fluctuate strongly with the business cycle and the inflation rate. Babihuga (2007) also finds that short-term interest rates and the real exchange rate are important determinants. She also finds that less developed economies, with relatively undeveloped financial systems, tend to have much higher capital ratios than their more developed counterparts during economic downturns (Babihuga, 2007, p. 21).
Sun (2011) investigates several FSI’s at the individual bank level, using the expected frequency of defaults as the dependent variable, and concludes that leverage ratios are the most reliable indicator but that the Return on Assets FSI also provides predictive power. However, Sun (2011) finds that Capital to Asset ratios and Non-performing Loans have little predictive power at the
14
individual bank level. Sun (2011) also finds that inflation and other global macroeconomic conditions, such as global excess liquidity and the financial stress index, are of significant influence.
Building on Čihák and Schaeck (2007), Costa Navajas and Thegeya (2013) conduct a study of the effectiveness of FSI’s as predictors of banking crises, using a multivariate logit model containing FSI’s, broad macroeconomic indicators, and institutional indicators. The dependent variable in their (Costa Navajas & Thegeya, 2013) dataset is a binary variable indicating the occurrence of a systemic banking crisis as described by Laeven and Valencia (2012), supplemented by the definition of a banking shock as used by Boyd et al. (2009): a significant annual decrease in gross loans outstanding. Their (Costa Navajas & Thegeya, 2013) results indicate significant correlation between some FSIs and banking crises, with contemporaneous Capital Asset Ratio and Return on Equity FSIs exhibiting significant negative correlation with the occurrence of banking crises. The lagged Return on Equity FSI is also found to be a significant leading indicator of crises (Costa Navajas & Thegeya, 2013). The coefficients found for contemporary Capital Asset Ratio and Return on Equity FSIs are negative, indicating that a decline of each of these FSIs indicates a potential crisis (Costa Navajas & Thegeya, 2013, p. 16). The large coefficient found for contemporary Capital Asset Ratio also implies that even small changes in this FSI could have large effects on the probability of default (Costa Navajas & Thegeya, 2013, p. 16). The coefficient for lagged Return on Equity is positive, which is consistent with theory that periods preceding systemic banking crises yield higher returns and higher joint risk within the banking sector (Costa Navajas & Thegeya, 2013, p. 16). Weak evidence that an increase in Non-interest Expenses to Gross Income correspond with a higher probability of systemic banking crises is also found (Costa Navajas & Thegeya, 2013, p. 16). Several macroeconomic control variables are found to have significant positive coefficients, including inflation, monetization and the ratio of broad money to international reserves (Costa Navajas & Thegeya, 2013, p. 16). However, Costa Navajas and Thegeya conclude that none of the broad macroeconomic variables are consistently significant in all models, and that included institutional indicators are not significant in any of the models (Costa Navajas & Thegeya, 2013, p. 16). A comparison with the earlier work by Čihák and Schaeck (2007) is also made, using the same model specifications, with results showing that Capital Asset Ratio and Non-performing Loans ratio as well as lagged Capital Asset Ratio and lagged Return on Equity have significant explanatory power (Costa Navajas & Thegeya, 2013, p. 18). In contrast with Čihák and Schaeck (2007), Costa Navajas and Thegeya (2013, p. 18) find a positive coefficient for lagged Return on Equity and a negative coefficient for Non-performing Loans Ratio.
15
Table 2: IMF FSIs that have are found to be significant explaining variables of the probability of a country being in a financial crisis, including the signs that are found.
Significant explanatory FSIs Čihák and Schaeck (2007)
Costa Navajas and Thegeya (2013)
Costa Navajas and Thegeya (2013)*
Capital to Risk Weighted Assets - (weak) - -
Nonperforming Loans to Total Gross Loans + (weak) - -
Return on Equity - -
Lagged Capital Asset Ratio -
Lagged Return on Equity - + +
* when using Čihák’s and Schaeck’s (2007) model specifications for comparison
2.4 – Conclusion of literature review
In conclusion, a wide array of literature exists investigating both the 2008 banking crisis and earlier financial crises. Two distinct methods are used to define a banking crisis, with literature using both discrete and continuous measures. Of the examined potential macroeconomic indicators, credit growth, real exchange rate and current account balance stand out as they appear significant in several studies. Other significant variables include country dummies, development measures, such as a GDP per capita variable or region dummies, and international trade. In literature examining IMF FSIs it is concluded that Capital to Risk Weighted Assets, Nonperforming Loans and Return to Equity are significant in explaining banking crises.
16
3 - Methodology
The methodology section contains descriptions of the data sample, a measure of crisis, FSI variables, macroeconomic control variables and the model specification in this study. The countries included in the data sample are discussed, followed by the selected FSIs and macroeconomic control variables. The regression model is then specified and discussed, followed by the relevant tests for
heteroscedasticity and serial correlation. Lastly, model variations utilized in robustness checks are discussed.
3.1 - Sample
The sample contains data on 44 countries and ranges from the first quarter of 2005 until the fourth quarter of 2015, resulting in a time series of 44 quarters. Of the countries that experienced a
systemic banking crisis in the Laeven and Valencia (2012) database, Iceland is not included because it does not report FSIs, and Kazakhstan, Mongolia and Nigeria are excluded due to lack of data. Croatia, Estonia, Georgia, and Lithuania are dropped due to severe gaps in observations or a general lack of observations. Ukraine is dropped because of significant changes in methodology (Costa Navajas & Thegeya, 2013). Germany, France and Russia were excluded because they only report yearly FSIs. The final sample therefore contains data on 38 countries. Due to limitations pertaining data availability, a balanced panel could not be compiled using quarterly data, and this dataset contains gaps. The resulting dataset contains only crisis countries that were involved in the 2008 financial crisis. A detailed overview of the included countries is given in Appendix 1.
Figure 1: map overview of countries included in the data sample used in this thesis. Included countries are colored dark grey.
17
3.2 - Measure of crisis
In this thesis, the Laeven and Valencia (2012) database of dummy variables indicating whether a country is experiencing a systemic banking crisis will be used (literature utilizing this database or earlier versions of it includes Demirgüç-Kunt & Detragiache, 1998; Mehrez & Kaufman, 2000; Beck, Demirgüç-Kunt & Levine, 2006; Čihák & Schaeck, 2007; Costa Navajas & Thegeya, 2013; Caprio et al., 2014). One problem with this database is that it is outdated: unlike earlier crises, no end dates are included for banking crises that occurred after 2004. The database used in this thesis is therefore updated with crisis end dates for each country, using the criteria that Laeven and Valencia (2012, p. 17) universally apply in their database: the end of a crisis is defined as the year before both real GDP growth and real credit growth have been positive for at least two consecutive years. In all cases, crisis duration is truncated at 5 years. A detailed overview of the Laeven and Valencia (2012) database is provided in Appendix 2.
3.3 - FSIs
Following Čihák and Schaeck (2007) and Costa Navajas and Thegeya (2013), this thesis examines a selection of core FSIs. This selection is mainly based on data availability and multicollinearity problems, as several FSIs capture the same risk category (Čihák & Schaeck, 2007; Costa Navajas & Thegeya, 2013). Similar to Čihák and Schaeck (2007) and Costa Navajas and Thegeya (2013), year on year changes in FSIs are not used because this would further reduce the number of available data points. Regulatory capital to risk-weighted assets, non-performing loans to total gross loans and return on equity are included because these FSIs are found to be significant in earlier studies (Čihák & Schaeck, 2007; Costa Navajas & Thegeya, 2013). Additionally, non-performing loans net of
provisions to capital, interest margin to gross income and non-interest expenses to gross income are included due to their availability as well as covering several different risk categories (Costa Navajas & Thegeya, 2013). An overview of the FSIs that are selected for this thesis is provided in table 3.
Table 3: descriptive statistics of FSIs selected as independent variables.
Variable Unit Frequency N Mean Std. Dev. Min Max Interest margin to gross income fraction quarterly 1144 .6018338 .1701692 -2.943336 1.491222 Non-interest expenses to gross income fraction quarterly 1145 .5809412 .1813088 -3.034569 1.866093 Non-performing loans net of provisions fraction quarterly 1113 .1667539 .2492946 -.129761 2.188017 Non-performing loans to total gross loans fraction quarterly 1123 .0473359 .0511442 .0008181 .3699481 Regulatory capital to risk-weighted assets fraction quarterly 1147 .1511114 .0287177 .0019506 .2433176 Return on equity fraction quarterly 1134 .1256178 .104827 -.9761606 .4170191
18
Figure 2: evolution of the averages of selected FSIs for crisis countries and non-crisis countries. Crisis countries are defined as countries having experienced a systemic banking crisis during the period between 2005 and 2015, as reported by Laeven and Valencia (2012).
As can be seen in figure 2, crisis countries had, on average, lower regulatory capital to risk-weighted assets than non-crisis countries, but all countries responded with a significant increase in capital levels after the banking crisis started in 2007. This trend is also noted by Costa Navajas and Thegeya (2013). Non-performing loans net of provisions and non-performing loans to total gross loans rose very rapidly in crisis countries and are consistently at higher levels after 2008. It is of note that the non-performing loans to total loans ratio is actually lower in crisis countries before the crisis erupted. One explanation for this is that crisis countries tend to be more developed countries, as is emphasized in several studies (Lane & Milesi-Ferretti, 2011): developed countries tend to have lower levels of non-performing loans to total gross loans before the outbreak of the crisis. It is
19
financial infrastructure than developing countries, therefore allowing businesses and households to plan their income more safely and manage finances more efficiently, therefore reducing the amount of non-performing loans (Mileris, 2014, p. 28). Similar to Costa Navajas and Thegeya (2013), return on equity tends to be higher in non-crisis countries, with both levels of return on equity already declining before the crisis started. One possible explanation is that banking profitability in crisis countries declined sharply before the crisis. Another possible explanation is that asset value declined due to write offs of, for example, subprime mortgages. A third explanation is that banks in crisis countries took on more leverage, as Moussu and Petit-Romec (2014) argue that the focus on return on equity as a performance measure drove banks into taking on excessive leverage. Non-interest expenses and interest margin to gross income are at higher levels for crisis countries, and the
difference between levels of crisis countries and levels of non-crisis countries increases after the start of the crisis. As the interest margin is related to bank leverage, one explanation for the higher
interest margin in crisis countries is that banks in crisis countries are driven to reduce their leverage as a result of the 2008 banking crisis, thus raising the interest margin.
3.4 - Macroeconomic control variables
As it is concluded earlier in this thesis, a number of general macroeconomic variables are found to be significant in literature. The selection of macroeconomic control variables is based on data
availability and the need to capture as many different transmission channels as possible, in order to avoid omitted variable bias. The real GDP growth rate is included because it is found to be significant in several studies (Demirgüç-Kunt & Detragiache, 1998; Beck, Demirgüç-Kunt and Levine, 2006). Beck, Demirgüç-Kunt and Levine (2006), Claessens et al. (2010), and Lane and Milesi-Ferretti (2011) conclude that GDP per capita is a significant variable, and several other studies find country or region dummies to be significant (Demirgüç-Kunt & Detragiache, 1998; Berkmen et al., 2010; Frankel & Saravelos, 2012), suggesting that country development has a role in the 2008 banking crisis, and therefore GDP per capita is included in this study. The inflation rate (Demirgüç-Kunt & Detragiache, 1998; Frankel & Saravelos, 2012), the growth rate of private credit (Beck, Demirgüç-Kunt and Levine, 2006; Claessens et al., 2010; Lane & Milesi-Ferretti, 2011; Frankel & Saravelos, 2012), the growth rate of house prices (Claessens et al., 2010), and the change in real exchange rate (Demirgüç-Kunt & Detragiache, 1998; Berkmen et al., 2012; Frankel & Saravelos, 2012) are included for similar reasons. The five year sovereign CDS spread was originally included, but later dropped because of severe gaps in time series.
20
Table 4: macroeconomic control variables, descriptive statistics.
Variable Measure Unit Frequency N Mean Std. Dev. Min Max
Real GDP growth Year on year growth rate
fraction quarterly 1678 .0261744 .039146 -.1616385 .2843082
Inflation Year on year growth rate of GDP deflator
fraction quarterly 1698 .0289445 .0347934 -.1348809 .2113177
ln (GDP per capita) Interpolated yearly average Log of 2011 I$ quarterly 1558 10.19352 .6474914 8.07498 11.47948 Current account to GDP
Current level fraction quarterly 1700 .0323078 .0996247 -.2995888 .498068
Private credit to GDP
Year on year growth rate
fraction quarterly 1642 .034658 .0903168 -.3325153 .5352021
Real exchange rate Year on year growth rate
fraction quarterly 1710 .002803 .065557 -.3097409 .2980113
External debt to GDP
Year on year growth rate
fraction quarterly 1651 .0326912 .1209425 -.2711619 .5401933
House price index Year on year growth rate
fraction quarterly 1494 .0158783 .0780395 -.4486 .39
3.5 - Model specification
Similar to Demirgüç-Kunt and Detragiache (1998), Čihák and Schaeck (2007), and Costa Navajas and Thegeya (2013), a multivariate pooled logit model is estimated:
𝐶𝑅𝐼𝑆𝐼𝑆𝑖,𝑡 = 𝛼 + 𝛽1∗ 𝑁𝐼𝐸𝑖,𝑡+ 𝛽2∗ 𝐼𝑀𝑖,𝑡+ 𝛽3∗ 𝑁𝑃𝐿𝑃𝑖,𝑡+ 𝛽4∗ 𝑁𝑃𝐿𝑇𝑖,𝑡+ 𝛽5∗ 𝑅𝐶𝐴𝑖,𝑡+ 𝛽6∗
𝑅𝑂𝐸𝑖,𝑡+ 𝛽7∗ 𝐺𝐷𝑃𝑖,𝑡−4+ 𝛽8∗ 𝐼𝑁𝐹𝑖,𝑡−4+ 𝛽9∗ 𝐶𝐴𝑃𝑖,𝑡−4+ 𝛽10∗ 𝐶𝐴𝑖,𝑡−4+ 𝛽11∗ 𝑃𝐶𝑅𝐸𝐷𝑖,𝑡−4+ 𝛽12∗
𝑅𝐸𝑋𝑖,𝑡−4+ 𝛽13∗ 𝐸𝑋𝑇𝑖,𝑡−4+ 𝛽14∗ 𝐻𝑂𝑈𝑆𝐸𝑖,𝑡−4+ 𝜀𝑖,𝑡 (1.1)
Where CRISIS is the dummy variable indicating whether a country is in a systemic banking crisis, as defined by Laeven and Valencia (2012), index i indicates a specific country and index t indicates a specific date, and 𝜀 is the regression error term. All macroeconomic control variables used are year on year growth rates and are lagged one year. The independent variables in this regression are specified in table 5.
21
Table 5: independent variables in the estimated logit regression.
NIE FSI: Non-interest expenses to gross income IM FSI: Interest margin to gross income NPLP FSI: Non-performing loans net of provisions NPLT FSI: Non-performing loans to total gross loans RCA FSI: Regulatory capital to risk-weighted assets ROE FSI: Return on equity
GDP Real GDP growth INF Inflation
CAP Natural logarithm of GDP per capita CA Current account to GDP
PCRED Private credit to GDP growth REX Real exchange rate growth EXT External debt to GDP growth HOUSE House price index growth
As Costa Navajas and Thegeya (2013) point out, maximum likelihood estimates are inconsistent for a fixed effects logit model with fixed time T, as information about incidental parameters stops accumulating after a certain number have been taken. Hsiao (2014, p.p. 237-238) demonstrates that the maximum likelihood estimates are inconsistent by observing a logit model with one regressor over two periods. Chamberlain (1980) estimates a fixed effects logit model with conditional time invariant effects. However, Costa Navajas and Thegeya (2013) point out that the conditional distribution of such a model would lead to the exclusion of all countries that did not experience a crisis at all, leading to a biased sample. A random effects model was considered, in order to include countries that did not experience a crisis during the examined period, but as Demirgüç-Kunt and Detragiache (1998, p. 90) point out, such a model would produce unbiased
estimates only if the random effects are uncorrelated with the regressors, which is very unlikely to
be true in practice. Therefore, a random effects model is not used. Due to the complex nature of the 2008 banking crisis, the fact that fixed effects or random effects models are not utilized in this thesis means that it is fairly certain that omitted variable bias will occur. Cramer (2005, p. 5) shows that omitting variables in a logit model depresses the partial slope coefficients of the model estimation towards zero, following Wooldridge (2002), who shows that this occurs in probit models. Therefore, the results of the performed regression in this thesis will have to be interpreted while keeping the likelihood of omitted variable bias in mind. Another problem occurs when including time dummies in the model, as the maximum likelihood estimates become unbounded for any time T when no crisis
22
occurs at all across the panel (Costa Navajas & Thegeya, 2013). Poirier and Rudd (1998) demonstrate that logit estimates are consistent but inefficient if time dependence exists, and that standard errors of regression coefficients are wrong (Costa Navajas & Thegeya, 2013). Costa Navajas and Thegeya (2013, p. 17) use robust standard errors clustered by country to correct this problem. Kézdi (2004) shows that the cluster-robust standard error estimator converges to the true standard error as the number of clusters approaches infinity, not the number of observations. Kézdi (2004) concludes that 50 clusters of roughly equal size is a sufficient number for accurate inference of standard errors when using the clustered standard error estimator, but that a small number of clusters, or very unbalanced cluster sizes, can lead to worse performance than that of the non-clustered robust standard error estimator.
3.6 Heteroscedasticity
Wooldridge (2002, p.p. 463-465) describes a variable addition method of the likelihood maximization approach which can be used to test binary response models, such as probit and logit, against models of a more general functional form. In this case, this approach (Wooldridge, 2002, p.p. 463-465) can be used to test the model used in this thesis for heteroscedasticity. First, the null model is estimated, assuming homoscedasticity. Secondly, the fitted linear indices are obtained. Thirdly, an augmented model is estimated including the subset of linear indices. Lastly, the joint significance of these interactions is tested against the null model using the standard Wald test for exclusion restrictions. The resulting test statistic is approximately 𝜒2 distributed with degrees of freedom equal to the number of independent variables in the specified models.
Table 6: modified Wald test statistic for heteroscedasticity. 𝐻0: 𝜎𝑖2= 𝜎2
𝜒2(14) 31,85
𝑃𝑟𝑜𝑏. > 𝜒2 0,0042
From the resulting test statistic, the null hypothesis is rejected and it is concluded that heteroscedasticity is present.
3.7 - Serial correlation
Wooldridge (2002, p.p. 176-177) describes a test for first order serial correlation, where the error term is being tested for an autoregressive process. A model is estimated using the residuals from a first-difference regression (Drukker, 2003). By using the first differences, effects at the individual level are removed (Drukker, 2003). Wooldridge (2002) observes that if the residuals are not serially
23
correlated, the correlation between the first difference residual at time t and the first difference residual at time t-1 must equal -0,5. The procedure therefore consists of regressing the first difference residuals on the lagged first difference residuals and testing whether the resulting regression coefficient is equal to -0,5 (Drukker, 2003). According to Drukker (2003), the procedure requires few assumptions, and has strong size and power properties. The procedure also accounts for within-panel correlation and conditional heteroscedasticity by using clustered standard errors at the panel level (Drukker, 2003).
Table 7: Wooldridge test statistic for first order serial correlation. 𝐻0: 𝜌(∆𝜀𝑖,𝑡; ∆𝜀𝑖,𝑡−1) = −0,5
𝐹(1,33) 437,863
𝑃𝑟𝑜𝑏. > 𝐹 0,0000
The test statistic results in the rejection of the null hypothesis, and it is concluded that first order serial correlation is present in the model. Costa Navajas and Thegeya (2013, p. 15) run regressions on FSIs against a time trend by country and use the residuals from these regressions in their final logit model, in order to address serial correlation. However, it is beyond the scope of this thesis to address this issue. Following Costa Navajas & Thegeya (2013), this thesis applies robust standard errors clustered by country. It must be noted that the dataset used in this thesis contains data on 38 countries, which is short of the ideal number of panels as specified by Kézdi (2004), as well as unbalanced panel data. As a result, the clustered standard errors may suffer from a bias, and are inflated when compared to non-clustered robust standard errors.
3.8 - Model variations
Following Costa Navajas and Thegeya (2013), the regression is estimated without FSIs as explaining variables, followed by an estimation that includes FSIs. The regression model is then expanded with FSIs that are lagged one quarter, and FSIs that are lagged one year. Due to multicollinearity, it is not possible to estimate regression models containing both contemporaneous FSIs and lagged FSIs (Costa Navajas & Thegeya, 2013). Because to the performing loans net of provisions and the non-performing loans to gross total loans display VIFs that are larger than 4, indicating that collinearity is present, additional robustness checks are also conducted by estimating several model variations that omit one of these variables.
24
4 - Results
Firstly, in order to examine the sensitivity of FSI ratios to different lag structures, several models are estimated containing lagged variables. Model (1) is estimated using only the one year lagged
macroeconomic control variables. Model (1) is then expanded with contemporaneous FSIs, resulting in model (2). Model (3) contains FSIs that are lagged one quarter, and model (4) contains FSIs that are lagged one year. The macroeconomic control variables are lagged one year in all model specifications, following Costa Navajas and Thegeya (2013, p. 21), as the basic model variation, variation (1), performs better when lagged values are used according to the Akaike Information Criterion. The regression results of these models are presented in table 8 on the next page.
In order to check for variance inflation due to collinearity of the non-performing loans net of provisions and the non-performing loans to total gross loans FSIs, models (5) through (11) are estimated, using the same lag structure of models (1) through (4), but sequentially omitting one variable or the other. For reasons of brevity, the results of model specifications (5) through (11) are presented in appendix 5.
25
Table 8: Regression outputs of model variations (1) through (4).
Variables (1) (2) (3) (4)
Real GDP growth (t-4) -14.10754 -9.647966 -8.102713 -7.268502 (8.130355) (7.658085) (8.607439) (9.646081) Inflation (t-4) 13.43278 19.80066 * 23.1521 * 24.80943 * (8.431621) (7.732) (9.595204) (9.859908) Log of GDP per capita (t-4) 3.367614 ** 3.379019 ** 3.461386 ** 3.579677 ***
(1.151587) (1.161673) (1.07595) (1.018308) Current Account to GDP (t-4) -5.912567 -8.446288 -8.16182 -8.033566 *
(3.466848) (4.440484) (4.226348) (3.994056) Private credit to GDP growth (t-4) -5.721091 -1.83078 -2.121101 -2.683299
(4.143855) (4.111721) (3.942082) (4.41543) Real exchange rate growth (t-4) .0201602 -2.746386 -2.705279 -1.459249 (2.047652) (2.286611) (2.227236) (2.729704) External debt to GDP growth (t-4) .488407 1.654842 2.550219 1.41021
(1.057393) (1.284516) (1.380971) (1.758123) House price growth (t-4) -15.84129 *** -10.09979 * -10.85119 * -12.73491 **
(4.341811) (4.693998) (4.954877) (4.90382)
FSIs FSIs (t-1) FSIs (t-4)
Interest margin to gross income .2459359 .5630562 .3337276 (1.807618) (1.613027) (1.522227) Non-interest expenses to gross income .0398361 -.0886324 .451985
(1.654242) (1.514699) (1.336368) Non-performing loans net of provisions 2.769962 2.570597 1.805397
(2.146627) (2.159879) (2.0677) Non-performing loans to total gross loans -10.96602 -10.63999 -9.958157
(6.810979) (7.029719) (8.79371) Regulatory capital to risk-weighted assets 35.11429 ** 32.05532 * 16.50578 (13.18479) (12.5735) (10.20315) Return on equity -8.114215 ** -8.486578 ** -8.110829 ** (3.062526) (3.174116) (3.122426) Constant -36.89283 ** -41.92439 ** -42.51699 *** -41.59177 *** (12.11973) (12.90868) (12.1361) (10.96792) Pseudo R-squared .317 .367 .364 .340 𝜒2 40.23 *** 92.20 *** 71.77 *** 91.53 *** N 1291 887 895 882
Akaike Information Criterion 829.2987 548.2418 540.9409 519.3894
Note: All macroeconomic control variables are lagged one year. Contemporaneous FSIs are used in model variation (2), FSIs that are lagged one quarter are used in model variation (3), and FSIs that are lagged one year are used in model variation (4). The dependent variable is the crisis dummy as specified by Laeven and Valencia (2012) from Q1 2005 until Q4 2015. Clustered standard errors by country are in parentheses. Concerning significance, *** indicates p<0.001, ** indicates p<0.01, and * indicates p<0.05.
26
Of the investigated contemporaneous FSIs, regulatory capital to risk-weighted assets and return on equity show significant coefficients. The positive coefficient found for regulatory capital to risk-weighted assets suggests that countries experiencing a crisis require their banks to hold higher capital reserves, as is illustrated in figure 2. When lagged one quarter, this positive coefficient is still significant, but when lagged one year it loses its significance, which could imply that holding higher capital reserves is a reaction to the 2008 banking crisis more than it is a cause of the 2008 banking crisis. This result also contradicts Čihák and Schaeck (2007) and Costa Navajas and Thegeya (2013), with both these studies finding a negative sign instead of a positive sign. However, Čihák and Schaeck (2007) and Costa Navajas and Thegeya (2013) also find that this coefficient loses its significance when lagged one year. Contrary to the regulatory capital to risk-weighted assets variable, the negative sign found for return on equity is robust to lag structures, as the coefficient for return on equity remains negative and significant when lagged one quarter and when lagged one year. This could imply that as equity profitability deteriorates, the likelihood of experiencing a banking crisis increases, which would make return on equity a leading indicator of the 2008 banking crisis. Both Čihák and Schaeck (2007) and Costa Navajas and Thegeya (2013) find a similar negative sign for return on equity as well. However, Costa Navajas and Thegeya (2013) find a positive sign for return on equity when it is lagged one year, which could indicate that higher risk associated with higher past returns is a leading
indicator instead. The coefficients found for the regulatory capital to risk-weighted assets and return on equity FSIs retain their signs and significance in models (5) through (11), where the
non-performing loans net of provisions and the non-non-performing loans to total gross loans FSIs are omitted sequentially. None of the four other FSIs that are included in the regression models are found to have significant coefficients.
Of the included macroeconomic control variables, GDP per capita, the inflation rate, and house price growth are found to have significant coefficients. The sign for GDP per capita is found to be positive and significant in all model specifications, which corroborates the findings of existing literature (Čihák & Schaeck, 2007; Claessens et al., 2010; Lane & Milesi-Ferretti, 2011). The 2008 banking crisis occurred mostly in highly developed economies, and as Claessens et al. (2010) and Lane and Milesi-Ferretti (2011) argue it could have been systemic in nature, affecting the entire first world. When FSI variables are not included in the models, the inflation rate is not found to be significant. However, when FSI variables are included, the inflation rate is significant with a positive coefficient, and remains so when lagged. The coefficient found for the house price growth variable is significant and negative in all model specifications, implying that a decline in house prices is a leading indicator of the 2008 banking crisis. This result corroborates the results found by Claessens et al. (2010) and is in line with the existing theory that a house price bubble played a role in the 2008
27
financial crisis: as the bubble collapsed, house prices started to decline, leading to a higher insolvency of mortgage takers (Mian & Sufi, 2014), eventually contributing to the eruption of the 2008 banking crisis.
Similar to previous literature using logit regression models, the performance of the estimated logit regression models is assessed based on Akaike’s Information Criterion, or AIC, and the model 𝜒2 statistic (Demirgüç-Kunt & Detragiache, 1998; Čihák & Schaeck, 2007; Costa Navajas & Thegeya, 2013). The null hypothesis when considering the model 𝜒2 is that all partial slope coefficients are equal to zero, and is rejected for all model specifications. The AIC is a measure of the relative quality of a specified model, for a given set of data (Čihák & Schaeck, 2007; Costa Navajas & Thegeya, 2013). Models with a lower AIC are preferred, and the AIC penalizes the addition of regressors (Čihák & Schaeck, 2007; Costa Navajas & Thegeya, 2013). As can be seen in table 8, model variation (1), which does not include the FSI variables, has a higher AIC value than model variations (2) through (4), which do include FSI variables. Of all the specified model variations, variation (4), which utilizes FSI variables that are lagged one year, performs best according to the AIC. This implies that the FSIs jointly add predictive power to the model, even though not all of the found coefficients are significant, which suggests that FSIs are useful macro-prudential tools. The fact that model variation (4), with lagged FSIs, performs best implies that the six FSIs that are examined are, jointly, most useful when used as forward looking indicators instead of as contemporaneous indicators.
It must be noted that the regression estimations are likely to be affected by omitted variable bias, because the utilized model specification does not employ fixed effects to capture these omitted variables, and because it would be very difficult to account for all variables that had an effect on the outbreak of the 2008 banking crisis, due to its complex nature. This is an unfortunate limitation that accompanies the estimation of a logit model without conditional fixed effects or random effects. Since international banking exposure was an important transmission channel for the spread of the 2008 banking crisis (Claessens et al., 2010), the omission of the geographical distribution of loans FSI variable might have left out an important transmission channel of the 2008 banking crisis. Financial deregulation, low interest rates, and high trade openness are deemed to be co-determinants of the 2008 banking crisis (Claessens et al., 2010, Caprio et al., 2014), and it is hypothesized that more developed countries show higher financial deregulation, lower interest rates and higher trade openness. Leaving these variables out of the model specifications could have caused the strong significance of the GDP per capita variable.
28
5 - Conclusion
This thesis investigates the usefulness of selected IMF Financial Soundness Indicators as predictors of the 2008 banking crisis. From the performed literature review, it is concluded that of the previously investigated FSIs, the capital to risk-weighted assets ratio, the non-performing loans to total gross loans ratio, and the return on equity ratio have been found to be significant in predicting previous banking crises. Of the previously investigated general macroeconomic variables, the credit growth, the real exchange rate, the current account balance and the GDP per capita level have been found to be significant and are included in the econometric model that is examined in this thesis. The Laeven and Valencia (2012) dataset containing dummy variables indicating whether a country is experiencing a systemic banking crisis is updated to Q4 2015 in this thesis, and is subsequently used in an
econometric examination of the 2008 banking crisis by looking at the period from Q1 2005 until Q4 2015. Six core FSIs are included as independent variables in the econometric analysis in this thesis, and their usefulness as predictors of the 2008 banking crisis is investigated.
A multivariate pooled logit regression is estimated using selected FSIs and macroeconomic control variables. Of the examined FSIs, the regulatory capital to risk-weighted assets variable and the return on equity variable are shown to have significant coefficients. The coefficients for the regulatory capital to risk-weighted assets ratio contradict earlier research done by Čihák and Schaeck (2007) and Costa Navajas and Thegeya (2013), who find negative coefficients instead of the positive coefficients found in this thesis. The usefulness of the regulatory capital to risk-weighted assets ratio as a predictor of the 2008 banking crisis is highly questionable, as its coefficient loses its significance when lagged one year. The possibility exists that capital requirements were raised in reaction to the 2008 banking crisis. The negative coefficients found for the return on equity FSI are robust to different lag structures and it is concluded that return on equity could be a leading indicator for the 2008 banking crisis, implying that declining profitability in a country’s banking sector forewarns the 2008 banking crisis. This result corroborates earlier work by Čihák and Schaeck (2007), but
contradicts earlier work by Costa Navajas and Thegeya (2013) as they find a positive sign when lagging the return to equity variable one year. The possibility exists that one or more regression coefficients are incorrectly classified, due to omitted variable bias. However, this is a limitation that accompanies all logit models that do not use conditional fixed effects or random effects.
It is concluded that whilst IMF FSIs, and the return on equity variable in particular, are useful tools to include in the macro-prudential toolkit, the systemic nature of this crisis and the fact that this crisis occurred almost exclusively in the developed world should not be discounted. This is underlined by the coefficients for the GDP per capita levels that are found to be strongly significant,
29
as well as previous literature heavily stressing the systemic and “advanced economies” nature of the 2008 banking crisis (Claessens et al., 2010; Lane & Milesi-Ferretti, 2011).
A limitation to the research that examines IMF FSIs is the lack of available data points. The dataset used in this thesis contains quarterly unbalanced panel data with time gaps. Previous work suffers from the same limitations, even when examining yearly data instead of quarterly data (Čihák & Schaeck, 2007; Costa Navajas & Thegeya, 2013). Further research would therefore strongly benefit from a more complete historical coverage of FSI data, as well as increased coverage of FSI data across countries.
30
Bibliography
Babihuga, R., (2007). “Macroeconomic and Financial Soundness Indicators: An Empirical Investigation”, IMF Working Paper 07/115. Washington: International Monetary Fund. Beck, T., Demirgüç-Kunt, A., Levine, R., (2006). Bank Concentration, Competition, and Crises: First
Results. Journal of Banking & Finance, Vol. 30, Issue 5, p.p. 1581-1603.
Berkmen, S. P., Gelos, G., Rennhack, R., & Walsh, J. P., (2012). The global financial crisis: explaining cross-country differences in the output impact. Journal of International Money and Finance, 31, p.p. 42-59.
Bordo, M., Eichengreen, B., Klingebiel, D., & Martinez-Peria, M.S., (2001). Is the Crisis Problem Growing More Severe?. Economic Policy, Vol. 32, pp. 51–82.
Boyd, J., De Nicolo, G., & Loukoianova, E., (2009). Banking Crises and Crisis Dating: Theory and Evidence. IMF Working Paper 09/141. Washington: International Monetary Fund. Brunnermeier, M.K., (2010). Deciphering the Liquidity and Credit Crunch 2007-2008. Journal of
Economic Perspectives, Vol. 23, no. 1, p.p. 77-100.
Caprio, G., Klingebiel, D., Laeven, L., & Noguera, G., (2005). “Appendix: Banking Crisis Database,” in Patrick Honohan and Luc Laeven (eds.), Systemic Financial Crises: Containment and
Resolution. Cambridge, U.K.: Cambridge University Press.
Caprio, G. Jr., D'Apice, V., Ferri, G., & Puopolo, G. W., (2014). Macro-financial determinants of the great financial crisis: Implications for financial regulations. Journal of Banking and Finance, 44, p.p. 114-129.
Chamberlain, G., (1980). Analysis of Covariance with Qualitative Data. Review of Economic Studies Vol. 47, p.p. 225-238.
Čihák, M., & Schaeck, K., (2007), How Well Do Aggregate Bank Ratios Identify Banking Problems. IMF Working Paper 07/275. Washington: International Monetary Fund.
Claessens, S., Dell’Ariccia, G., Igan, D., & Laeven, L., (2010). Cross-country experiences and policy implications from the global financial crisis. Economic Policy, April 2010, p.p. 267-293. Costa Navajas, M., & Thegeya, A., (2013). Financial Soundness Indicators and Banking Crises. IMF
31
Cramer, J.S., (2005). Omitted Variables and Misspecified Disturbances in the Logit Model. Amsterdam School of Economics Discussion Paper 2005/05. Amsterdam: University of Amsterdam.
Demirgüç-Kunt, A., & Detragiache, E., (1998). The Determinants of Banking Crises in Developing and Developed Countries. Staff Papers, International Monetary Fund, Vol. 45, No. 1, pp. 81-109.
Demirgüç-Kunt, A., (2005). Cross-Country Empirical Studies of Systemic Bank Distress: A Survey. IMF Working Paper 05/96. Washington: International Monetary Fund.
Drukker, D. M., (2003). Testing for serial correlation in linear panel-data models. The Stata Journal, 3(2), p.p. 168-177.
Frankel, J., & Rose, A., (1996). Currency Crashes in Emerging Markets: An Empirical Treatment. Journal of International Economics 41, no. 3/4, p.p. 351-366.
Frankel, J., & Saravelos, G., (2012). Can leading indicators assess country vulnerability? Evidence from the 2008-2009 global financial crisis. Journal of International Economics, 87, p.p. 216-231. Hawkins, J. & Klau, M., (2000). Measuring potential vulnerabilities in emerging market economies. BIS
Working Papers 91, Bank for International Settlements.
Honohan, P., & Laeven, L., (2005) (eds.). Systemic Financial Crises: Containment and Resolution. Cambridge, U.K.: Cambridge University Press.
Hsiao, C., (2014). Analysis of Panel Data (Third edition). Cambridge, U.K.: Cambridge University Press.
IMF, (2001). Financial Soundness Indicators: Policy Paper.
http://www.imf.org/external/np/mae/fsi/2001/eng/pp.pdf, accessed November 27 2016. IMF, (2006). Financial Soundness Indicators: Compilation Guide.
http://www.imf.org/external/pubs/ft/fsi/guide/2006/pdf/fsiFT.pdf, accessed November 27 2016.
IMF, (2007). Financial Soundness Indicators: Experience with the Coordinated Compilation Exercise
and Next Steps. http://www.imf.org/External/NP/pp/2007/eng/101807a.pdf, accessed
32
IMF, (2008). Amendments to the Financial Soundness Indicators (FSIs): Compilation Guide.
http://www.imf.org/external/pubs/ft/fsi/guide/2008/pdf/071408.pdf, accessed November 27 2016.
IMF, (2009). Press Release: IMF Launches New Online Database of Financial Soundness Indicators. July 31, 2009. http://www.imf.org/en/news/articles/2015/09/14/01/49/pr09275, accessed November 27 2016.
Kézdi, G., (2004). Robust Standard Error Estimation in Fixed-Effects Panel Models. Hungarian Statistical Review Special (9), p.p. 96-116.
Laeven, L., & Valencia, F., (2010). Resolution of Banking Crises: The Good, the Bad, and the Ugly. IMF Working Paper 10/146. Washington: International Monetary Fund.
Laeven, L., & Valencia, F., (2012). Systemic Banking Crises: An Update. IMF Working Paper 12/163. Washington: International Monetary Fund.
Lane, P. R., & Milesi-Ferretti, G. M., (2011). The cross-country incidence of the economic crisis. IMF Economic Review, 59(1), p.p. 77-110.
Mehrez, G., & Kaufman, D., (2000). Transparency, Liberalization, and Banking Crises. Policy Research Working Paper 2286, the World Bank.
Mian, A., & Sufi, A., (2014). House of Debt. Chicago, IL, U.S.A.: The University of Chicago Press. Mileris, R., (2014). Macroeconomic Factors of Non-Performing Loans in Commercial Banks.
Ekonomika, Vol. 93, Issue 1, p.p. 22-39.
Moussu, C., & Petit-Romec, A., (2014). ROE in Banks: Myth and Reality. ESCP Working Paper. https://ssrn.com/abstract=2374068. Accessed February 14 2017.
Poirier, D., & Rudd, P., (1988). Probit with Dependent Observations. Review of Economic Studies Vol. 55, p.p. 593-614.
Rose, A. K., & Spiegel, M. M., (2010). Cross country causes and consequences of the 2008 crisis: International linkages and American exposure. Pacific Economic Review, 15(3), p.p. 340-363. Rose, A. K., & Spiegel, M. M., (2011). Cross country causes and consequences of the crisis: An update.
European Economic Review, 55, p.p. 309-324.
Rose, A. K., & Spiegel, M. M., (2012). Cross country causes and consequences of the 2008 crisis: Early warning. Japan and the World Economy, 24, p.p. 1-16.
33
Sun, T., (2011). Identifying Vulnerabilities in Systemically Important Financial Institutions in a Macro-financial Linkages Framework. IMF Working Paper 11/11. Washington: International Monetary Fund.
Wooldridge, J. M., (2002). Econometric analysis of cross-section and panel data. Cambridge, MA, U.S.A.: MIT Press.
