Causes of Banking Crises: Deregulation, Credit Booms and Asset Bubbles, Then and Now Saktinil Roy (Athabasca University) David M. Kemme (University of Memphis) Objectives • How similar are the global financial crisis of 2008-2009 and historical banking crises around the world? • What factors are most common in the runup to banking crises? • What factors triggered the global financial crisis of 2008-2009? Background and Emerging Literature • The collapse of housing bubbles and the global financial crisis of 2008 naturally prompted the emergence of a literature that seeks to identify and explain the factors that triggered the crisis. • In the literature as well as in debates in policy circles, the most commonly cited causes are: • Lax regulation and oversight • Risky financial innovations and excessive risks taken by financial institutions Background and Emerging Literature (contd..) • Capital and credit flows in the aftermath of the Asian financial crises in 1997 • Interdependencies in assets that increased the risk of contagion and reduced the ability of risk sharing • The low interest rate policy of the Federal Reserve and other central banks in early 2000s helped to inflate the housing bubbles by encouraging easy availability of mortgage loans • Increase in risky mortgage debts has been partly an effect of the rise in income inequality over the last three decades. Background and Emerging Literature (contd…) • Reinhart and Rogoff (2008, 2009), Krugman (2007, 2009), Bordo (2008), Gorton (2010), Almunia et. al (2009), Roubini and Mihm (2010), and Rajan (2010) are only few of the authors contributing in this area • A rekindled interest in banking crises has also inspired an empirical literature that examines probable common patterns across banking crises based on their predictability Background and Emerging Literature (contd…) • Following the literature of early warning systems for financial crises, Barrell et. al. (2010), Jorda et. al. (2011), Roy and Kemme (2011) and others have investigated robust similarities in the run-ups to the global financial crisis and historical banking crises. • The present paper also examines robust similarities across the global financial crisis and historical banking crises based on their predictability. • We consider a larger set of crisis indicators, a longer historical period and a greater number of historical crisis episodes in advanced countries -- present fresh results that are not always similar to those in the extant literature. Model and Data • Two most popular approaches to predicting financial crises: • i) “Signal method” (Kaminsky et. al. , 1998, 1999, 2000)– examines individual indicators and issues alarms if they indicate anomalous behaviour exceeding a threshold; • ii) Logit/probit models (Berg and Pattillo 1999, Demirguc-Kunt and Detragiachem,1998; and Davis and Karim, 2008) examine the joint effects of several indicators on the probability of crisis. Alarms are issued if the estimated probability of crisis exceeds a pre-specified threshold; Statistical significance of each indicator can be tested. • We examine several specifications of bivariate and multivariate panel logit models to examine the joint effects of different combinations of indicators on the probability of crisis and to better capture fat tails in the data Model and Data (contd..) • We ask, specifically, if the historical crises and the current global crisis can be predicted to occur within a period of three years. • The corresponding modeling strategy requires that all three observations prior to a crisis episode are labeled as “pre-crisis” and all observations prior to these three years are labeled as “tranquil.” • For regression purposes, the dependent variable representing crisis probability is assigned the actual value one when the observation is “precrisis” and zero when the observation is “tranquil.” Model and Data (contd..) • The two essential criteria for assessing crisis similarity are: • Percentage of “pre-crisis” years correctly called (conditional probability of an alarm given a crisis within three years) • True alarms as a percentage of total alarms (conditional probability of a crisis within three years given an alarm) Model and Data (contd..) • The Choice of Threshold • The first preference under all conditions: 50% • No bias towards pre-crisis or tranquil observations • However, for a rare event, such as banking crisis, only a few alarms are generated • Candelon, Dumitrescu and Hurlin (2009) recommend several alternative criteria for the determination of an optimum threshold • The minimum bias toward either pre-crisis or tranquil observations: a threshold that simultaneously and conditionally maximizes the percentage of pre-crisis observations correctly called and the percentage of tranquil observations correctly called in the within sample exercises. • We apply both 50% and optimum thresholds and compare the results Model and Data (contd..) Crisis indicators • • • • • • • • • • • Current account as a percentage of GDP (CA) Growth rate of per capita real GDP (GGDP) Public debt as a percentage of GDP Real house prices (RHP) Real share prices (RSP) Income inequality (IE) Central bank real interest rate (RIR) M2/reserves (M2/R) Bank liquidity (LIQ) Currency appreciation (APP) Private sector debt as a percentage of GDP (PVD) Model and Data (contd..) • Data: Historical crisis episodes in advanced countries since World War II --Same as in Reinhart and Rogoff (2009), depending on the availability of data. • Due to a large number of missing observations, following the econometric literature, employ a “choice-based” sample: • Make the sample distribution symmetric across historical crisis experiences. • Test similarities with within sample and out of sample prediction exercises. • However, no suitable quantitative measure to identify a banking crisis (Reinhart and Rogoff, 2009) – quite unlike in the case for currency crisis. • Following Reinhart and Rogoff (2009), identify the crises by the dates on which they were announced. Model and Data (contd..) For the out of sample exercise consider a representative sample of crises: • USA (2007); Britain (2008); Ireland (2008); Spain (2008) • All four countries had spectacular prior housing bubbles and suffered the most severe banking crises – also found in Reinhart and Rogoff (2008), inter alia. • To check the robustness of results, we also predict out of sample the relatively tranquil environments of Australia, Canada and New Zealand around 2008-2009. • To further check the robustness of results from logit model specifications and to identify which indicators primarily contributed to the global financial crisis, we report Granger causality test results on a selected set of indicators. Results from Estimation and Forecasting Significance of Coefficients • In the bivariate specifications : current account, real GDP growth, real interest rate, income inequality, M2/reserve, private sector debt, real share prices and real house prices are significant at 5% or 10% level. • In the multivariate specifications: real GDP growth, income inequality, private sector debt, real share prices and real house prices are significant at 5% or 10% level in most specifications Results from Estimation and Forecasting (contd..) Within Sample Predictions Bivariate specifications: At 50% threshold, PVD, RSP, and RHP predict 37%-63% of the pre-crisis years correctly and issue alarms that are correct 47%-66% of the time. At 50% threshold, ALL other indicators fail either completely or nearly so. At optimum threshold, the best performances are by PVD, RSP, and RHP: 69 – 76% of the pre-crisis years are correctly called and 52-55% of the alarms are true. At optimum threshold, the only other notable performance is by income inequality (IE): 67% of the pre-crisis years are correctly called and 42% true alarms. Results from Estimation and Forecasting (contd..) • Multivariate Specifications • At 50% threshold, the specification containing only PVD, RSP and RHP predicts 61% of the pre-crisis years correctly and generates 75% true alarms. • At optimum thresholds, the same specification predicts 81% of the pre-crisis years correctly and generates 64% true alarms. • Adding any other variable to this combination of three variables does not improve prediction performance significantly. • The worst performance is by the specification which has ALL variables except PVD, RSP and RHP. Results from Estimation and Forecasting (contd..) Out of Sample Predictions • Bivariate specifications At 50% threshold, PVD, RSP, and RHP predict 67%-100% of the pre-crisis years correctly and issue alarms that are true 30%-100% of the time. At 50% threshold, ALL other indicators fail either completely or nearly so. At optimum threshold, PVD, RSP, and RHP predict 100% of the pre-crisis years correctly and issue alarms that are correct 30%-100% of the time. At optimum threshold, notably CA also does very well: 100% of the pre-crisis years are correctly called and 75% of the alarms are true. At optimum threshold all other indicators do much worse. Results from Estimation and Forecasting (contd..) • Multivariate Specifications • The best performance is by the specification that has PVD, RSP, RHP and IE. • At 50% threshold, 100% of the pre-crisis years correctly called and 86% of the alarms are true. • At optimum threshold, 100% of the pre-crisis years are correctly called and 75% of the alarms are true. • No other specification does nearly as well. • The worst performance is by the specification that has ALL indicators except PVD, RSP and RHP. Results from Estimation and Forecasting (contd..) Check Robustness of Results • Log probability scores, Quadratic probability scores, Global squared bias, McFadden R-squared, AIC, SIC and HIC confirm the results. • Out of sample predictions of the tranquil observations in the tranquil countries (Canada, Australia and New Zealand) and crisis countries (US, UK, Ireland and Spain) are consistent. Granger Causality Test Results • For crisis countries, if there is any causality between CA and RHP or PVD then it is from RHP or PVD to CA. • For tranquil countries quite the opposite is true. • In the US and Ireland, any possible causality between IE and RHP or PVD is from the former to the later. Conclusions Similarities of banking crises are most robust in terms of prior stock and real estate bubbles in an era of increasing financial deregulation where private sector debt continues to expand. Lends credence to Shiller’s (2005, 2008) dictum that, often spurred by “new era” stories of indefinite future increases of asset prices, the over-optimism of market participants— and sometimes also of policy makers and economists— sets the stage for a financial crisis, particularly in a period of regulatory imprudence. A sustained rise in income inequality is the most important other factor that demonstrates a similarity between the current global crisis and historical banking crises. Conclusions (contd.) Sustained surges in capital inflow, increasing GDP share of public debt, falling growth rate of real GDP per capita, the central bank’s low interest rate policy, and even rising income inequality, when taken independently from bubbles in equity and housing markets, largely fail to demonstrate similarities across historical banking crises and the current global crisis. The important implication: on most occasions of a systemic banking crisis— even as catastrophic as to roil the whole global financial system— there need not be a prior sustained surge in capital inflow, or the central bank having slashed the key interest rate for a prolonged period, or even structural shocks restraining economic growth. Conclusions (contd..) • It also follows that increasing financial deregulation over three decades that helped to increase private sector debt and to form a spectacular housing bubble contributed most to the global financial crisis of 2008. • The rise in income inequality over the same period was the next most important factor that contributed to the crisis. • Large capital inflows in the aftermath of the Asian financial crisis might have contributed to the global financial crisis, but it was a secondary factor.