Make It or Break It? Assessing Credit Booms in Developing Countries

Make It or Break It? Assessing Credit Booms in Developing Countries Abstract In this paper, we argue that the assessment of systemic risk associated with rapid credit growth should be conducted with different tools at different stages of financial development. In particular, using a signal detection framework, we show that, when the level of financial development is low, the most used leading indicator of banking crisis, namely the credit-to-GDP gap, has poor predictive performance, as it is unable to disentangle episodes of financial deepening from bubble-like credit booms. On the contrary, a new class of indicators, comparing credit levels to structural benchmarks, has good capacity to correctly predict financial crises in economies in the early stages of financial development. 1. Background The past 10 years witnessed a global trend towards increased financial deepening in developing countries (Figure 1). According to an extensive theoretical and empirical literature on the finance-growth nexus, this development is expected to support economic growth, by promoting a higher mobilisation of savings, improving the allocation of capital, enhancing risk management, and facilitating transactions (Rajan and Zingales, 1998; Levine et al., 2000; Levine, 2005). However, there is also a vast literature on financial crises and early warning systems (EWS), mainly based on the experience of advanced economies, which warns that fast acceleration in credit expansion may entail tradeoffs in terms of financial stability (Gourinchas et al., 2001; Cottarelli et al., 2005; Mendoza and Terrones, 2008; Dell’Ariccia et al., 2012; Crowe et al., 2013). Figure 1: View largeDownload slide Rapid Credit Growth in Developing Countries. Source: World Bank World Development Indicators Database. Figure 1: View largeDownload slide Rapid Credit Growth in Developing Countries. Source: World Bank World Development Indicators Database. Evidence from recent empirical literature seems to suggest that the negative effects of credit growth are less likely to manifest in developing countries. Several studies show in fact that credit booms are less frequently associated with systemic banking crises in those countries (Barajas et al., 2007; Dell’Ariccia et al., 2012; Arena et al., 2015; Meng and Gonzalez, 2017). This finding is likely to reflect the fact that, in economies in the early stages of financial development, rapid credit expansions are mostly connected to a healthy process of financial development rather than to the buildup of financial vulnerabilities, but the more the financial sector develops in size and sophistication, the more the risks of additional financial deepening exceed the benefits. Nonetheless, excessive credit growth can be a source of risk also in countries at an early stage of financial development, as there are limits to a country’s capacity to absorb financial deepening at each point in time. The systemic financial crisis experienced in Nigeria in 2009 offers a recent lesson on the dangers of credit booms in developing economies. The crisis followed a consolidation of the banking sector in the years 2005–06,1 which spurred a large credit expansion (Figure 2)—further fuelled by large oil-related inflows and a loose monetary policy stance. The acceleration in credit was too rapid to be absorbed in productive sectors of the economy and significant flows were channelled to non-priority sectors and to the capital markets. When the stress generated by the global financial crisis burst the equity bubble and oil prices collapsed, the credit boom ended in a systemic banking crisis (Sanusi, 2010). Figure 2: View largeDownload slide Nigeria: A Boom Bust Episode. Source: World Bank World Development Indicators Database. Figure 2: View largeDownload slide Nigeria: A Boom Bust Episode. Source: World Bank World Development Indicators Database. The issue is then understanding whether the fast pace of credit growth reflects a process of financial deepening that is beneficial for the real economy (make it), or if instead, it is indicative of systemic risk buildup, detrimental to loan quality and banking system stability (break it). In this paper, we aim to investigate this issue more in detail. We will show that the choice of the metric used to measure excess credit is critical to disentangle episodes of financial deepening from bubble-like credit booms. In particular, our results suggest that traditional leading indicators of systemic risk buildup, which compare credit levels to their long-term trends, are unable to tell the difference between good and bad booms in countries at a low level of financial development. On the contrary, indicators of excess credit comparing credit levels to a structural benchmark have a median good predictive performance in those countries. The findings in this paper are important from a policy perspective. First, assessing when credit growth can pose risks to financial stability is relevant for the activation and calibration of macro-prudential tools. An incorrect assessment of the risks associated to a boom might lead to an erroneous activation or calibration of macro-prudential tools, with the risk of either hindering financial development or, on the opposite side of the spectrum, leaving systemic risk buildup undetected. Second, as the analysis suggests that factors other than excess credit are frequently the source of financial stress in countries at a low level of financial development, alternative early warning indicators must be sought for these countries. The risk is that, by using a leading indicator of crises based on excess credit, sources of systemic risk peculiar to developing countries would pass unnoticed if they do not manifest themselves as credit booms. This paper relates to two strands of literature. First, it is associated with the literature on the nexus between boom-bust episodes and financial crises (Mendoza and Terrones, 2008; Reinhart and Rogoff, 2009; Jorda et al., 2011; Dell’Ariccia et al., 2012; Gourinchas and Obstfeld, 2012; Laeven and Valencia, 2013), and to the empirical research on EWS and indicators of financial crises (Kaminsky and Reinhart, 1999; Borio and Drehmann, 2009). Our work, however, differs from this empirical research in the fact that it focuses on the relationship between financial stability and financial development and proposes different leading indicators of systemic risk buildup depending on the stage of financial development. The paper is also connected to the literature on benchmarking financial systems (Beck et al., 2008; Al-Hussainy et al., 2011), as it uses the financial possibility frontier concept to distinguish between episodes of financial deepening and occurrences of systemic risk buildup in countries at an early stage of financial development. The rest of the paper is organised as follows. Section 2 classifies measures of excess credit into two major categories: cyclical and structural indicators. Section 3 tests the predictive performance of financial crises of the two classes of indicators across income levels and regions. Finally, Section 4 summarises the findings and suggests future research. 2. Metrics of Excess Credit How much credit is too much? There are two possible ways of approaching this question. The first entails assessing credit levels along the time dimension. Based on this approach, credit would be considered excessive if it is significantly above historical values. The second option consists in evaluating credit levels along the cross-sectional dimension. In this case, credit would be regarded as excessive if it is higher than the level recorded in economies with similar structural characteristics. These two approaches permit to define two classes of metrics of excessive credit, which we will refer hereafter as cyclical and structural indicators, respectively. 2.1 Cyclical indicators Based on the evidence of advanced economies and large emerging markets, the early warning literature relates the concept of excess credit to the notion of financial cycles and suggests that ‘peaks in the financial cycle (i.e. booms) are closely associated with systemic banking crises’ (Borio, 2012). A number of empirical and theoretical papers provide explanations for why lending booms may lead to financial stress. One chain of causation links credit booms and banking crises to excessive risk taking during the upswing of the financial cycle. This, in turn, may be stimulated by accommodative monetary policies, especially those in place for extended periods (Gambacorta et al., 2009). These dynamics tend to be amplified by a financial accelerator mechanism, where the supply of credit increases and credit standards are loosened pari passu with an improvement in collateral values (Kiyotaki and Moore, 1997; Bernanke et al., 1999; Gilchrist et al., 2009; Schularick et al., 2012). In general, the boom phase is captured by deviations of a credit measure from its historical trend, thereby defining a gap. Methodologies, however, differ substantially in the choice of the credit measure and in the computation of the trend. Among others, the early warning literature assigns a prominent role to the gap computed as the difference between the credit-to-GDP ratio and its long-term trend, calculated with a backward-looking Hodrick–Prescott (HP) filter with a high smoothing parameter. Empirical research from the Basel Committee on Banking Supervision has shown that, indeed, this gap (BIS gap hereafter) is a valuable leading indicator of systemic banking crises in advanced economies (Borio and Lowe, 2002; Drehmann et al., 2010, 2011), and, as such, Basel III has endorsed it as a guide to set the countercyclical capital buffer (BCBS, 2010). To date, the predictive performance of the BIS gap has been mainly tested on advanced economies. When the analysis has been extended to developing countries and emerging markets, results are mixed. A study by the IMF (2011), which assessed the performance of the BIS gap on a very large sample including advanced economies, emerging markets and low-income countries (169 countries in total), found that, for both emerging markets and low-income countries, the BIS gap did not perform well as a signalling variable. Drehmann and Tsatsaronis (2014), instead, using a sample including both advanced economies and emerging markets, found evidence that in emerging markets the BIS gap remains a good indicator of financial stress, albeit the performance is not as strong as in advanced economies. Emerging markets included in this sample, however, have mainly a developed financial sector. As highlighted in the background section, and argued by a number of commentators (Gerš and Seidler, 2012; IMF, 2014), rapid credit growth in low-income countries and emerging markets may reflect improved economic fundamentals and financial deepening. In this case, the signalling power of the BIS gap would not be compromised only if financial deepening occurs at a steady pace, as this would be embedded in the long-term trend and would not impact on the gap (Drehmann and Tsatsaronis, 2014). If, instead, financial deepening takes the form of sudden and rapid increases in credit growth, these would not be captured in the trend and might be signalled by the gap as buildup of financial vulnerability. 2.2 Structural indicators The literature on benchmarking financial systems offers an alternative approach for measuring excess credit, based on the concept of financial possibility frontier. According to this research strand, the development of a country’s financial system is critically influenced by structural factors that are invariant in the short term and often lie outside the purview of policy makers (Beck et al., 2008; Al-Hussainy et al., 2011). Those factors impose an upper limit to financial deepening in an economy at a given point in time, represented by the financial possibility frontier. This is the efficient (and safe) level of financial development, typically derived from a panel regression that estimates the relationship between a measure of financial development Y and a set of structural indicators X: Yi,t=f(Xi,t)+ϵi,t, where i and t are the country and time indexes, and ϵ is an error term. The difference between actual and predicted levels of financial development provides important information on the status of a country’s financial sector. A negative gap would signal an inefficient financial sector that does not operate at capacity. Instead, overshooting the predicted level of financial development would be associated with overheating pressures and financial stress (Barajas et al., 2013). The gap with respect to the frontier is, thus, a structural indicator of financial performance, as opposed to the gap with respect to the long-term trend, which is a cyclical measure. To date the capacity of structural indicators to predict financial crises has not been investigated. Barajas et al. (2013), however, found evidence that periods with a positive gap (i.e. levels of private credit to GDP above the benchmark) are more likely to be associated with booms that end in low growth episodes or even banking crises (bad booms). Instead, zero gaps (when a country’s private credit to GDP is close to its structural benchmark) have the lowest incidence of bad booms. In addition, they found that large changes in the gap, especially positive changes, are associated with a higher likelihood that the boom will be sub-par or end in a crisis. These findings are promising and suggest that the class of structural indicators could represent an alternative to cyclical indicators as predictors of financial crisis. In addition, these indicators might provide a better measure of excess credit in developing countries, as they account for financial deepening while flagging whether financial development is consistent with a country’s structural characteristics. 3. Systemic Risk Assessment across Levels of Financial Development In this section, we compare the predictive performance of the two classes of indicators across levels of financial development. The goal is to assess whether excess credit should be measured with different metrics at different stages of financial development. 3.1 Data Our analysis is conducted on a sample that includes 79 countries belonging to all income levels and regions (Table 1). Table 1: Sample Composition Country Income Region Data Source First Observation Last Observation Algeria UMI MENA IFS 1970 2011 Argentina UMI LAC IFS 1970 2011 Australia HI EAP BIS 1970 2011 Austria HI ECA BIS 1970 2011 Bangladesh LI EAP IFS 1974 2011 Belgium HI ECA BIS 1971 2011 Benin LI SSA IFS 1970 2011 Bolivia LMI LAC IFS 1970 2011 Brazil UMI LAC IFS 1970 1985 Burkina Faso LI SSA IFS 1970 2011 Burundi LI SSA IFS 1970 2011 Cameroon HI MNA IFS 1970 2011 Canada LMI SSA BIS 1970 2011 Cape Verde UMI SSA IFS 1980 2011 Centr. Afric. Rep. LI SSA IFS 1970 2011 Chad LI SSA IFS 1970 2011 Chile HI LAC IFS 1970 2011 Congo, Dem. Rep. LI SSA IFS 1970 1995 Congo, Rep. LMI SSA IFS 1970 2011 Costa Rica UMI LAC IFS 1970 2011 Cote d’Ivoire LMI SSA IFS 1970 2011 Denmark HI ECA BIS 1970 2011 Dominican Republic UMI LAC IFS 1970 2011 Ecuador UMI LAC IFS 1970 2011 Egypt, Arab Rep. LMI MENA IFS 1970 2011 El Salvador LMI LAC IFS 1970 2011 Finland HI ECA BIS 1971 2011 France HI ECA BIS 1970 2011 Germany HI ECA BIS 1970 2011 Ghana LMI SSA IFS 1970 2011 Greece HI ECA BIS 1971 2011 Guatemala LMI LAC IFS 1970 2011 Guyana LMI LAC IFS 1970 2011 Honduras LMI LAC IFS 1970 2011 Iceland HI ECA IFS 1970 2011 India LMI EAP BIS 1970 2011 Indonesia LMI EAP BIS 1976 2011 Ireland HI ECA BIS 1972 2011 Italy HI ECA BIS 1970 2011 Japan HI EAP BIS 1970 2011 Jordan UMI MENA IFS 1970 2011 Kazakhstan LMI ECA IFS 1993 2011 Kenya LI SSA IFS 1970 2011 Korea, Rep. HI EAP BIS 1970 2011 Madagascar LI SSA IFS 1970 2011 Malaysia UMI EAP IFS 1970 2011 Mali LI SSA IFS 1970 2011 Mauritania LI MENA IFS 1970 2011 Mexico UMI LAC BIS 1991 2011 Mongolia UMI LAC IFS 1991 2011 Nepal LI EAP IFS 1970 2011 Netherlands HI ECA BIS 1970 2011 New Zealand HI EAP IFS 1970 2010 Nicaragua LMI LAC IFS 1970 2011 Niger LI SSA IFS 1970 2011 Nigeria LMI SSA IFS 1970 2011 Norway HI ECA BIS 1970 2011 Panama LMI LAC IFS 1970 2011 Paraguay LMI LAC IFS 1970 2011 Peru UMI LAC IFS 1970 2011 Philippines LMI EAP IFS 1970 2011 Portugal HI ECA BIS 1970 2011 Senegal LMI SSA IFS 1970 2011 Sierra Leone LI SSA IFS 1970 2011 Singapore HI EAP BIS 1971 2011 South Africa LMI SSA BIS 1970 2011 Spain HI ECA BIS 1972 2011 Sri Lanka LMI EAP IFS 1970 2011 Swaziland LMI SSA IFS 1973 2011 Sweden HI ECA BIS 1970 2011 Switzerland HI ECA BIS 1970 2011 Thailand UMI EAP BIS 1970 2011 Togo LI SSA IFS 1970 2011 Tunisia UMI MENA IFS 1970 2011 Turkey UMI ECA IFS 1970 2011 UK HI ECA BIS 1970 2011 USA HI MNA BIS 1970 2011 Uruguay HI LAC IFS 1970 2011 Venezuela, RB UMI LAC IFS 1970 2011 Country Income Region Data Source First Observation Last Observation Algeria UMI MENA IFS 1970 2011 Argentina UMI LAC IFS 1970 2011 Australia HI EAP BIS 1970 2011 Austria HI ECA BIS 1970 2011 Bangladesh LI EAP IFS 1974 2011 Belgium HI ECA BIS 1971 2011 Benin LI SSA IFS 1970 2011 Bolivia LMI LAC IFS 1970 2011 Brazil UMI LAC IFS 1970 1985 Burkina Faso LI SSA IFS 1970 2011 Burundi LI SSA IFS 1970 2011 Cameroon HI MNA IFS 1970 2011 Canada LMI SSA BIS 1970 2011 Cape Verde UMI SSA IFS 1980 2011 Centr. Afric. Rep. LI SSA IFS 1970 2011 Chad LI SSA IFS 1970 2011 Chile HI LAC IFS 1970 2011 Congo, Dem. Rep. LI SSA IFS 1970 1995 Congo, Rep. LMI SSA IFS 1970 2011 Costa Rica UMI LAC IFS 1970 2011 Cote d’Ivoire LMI SSA IFS 1970 2011 Denmark HI ECA BIS 1970 2011 Dominican Republic UMI LAC IFS 1970 2011 Ecuador UMI LAC IFS 1970 2011 Egypt, Arab Rep. LMI MENA IFS 1970 2011 El Salvador LMI LAC IFS 1970 2011 Finland HI ECA BIS 1971 2011 France HI ECA BIS 1970 2011 Germany HI ECA BIS 1970 2011 Ghana LMI SSA IFS 1970 2011 Greece HI ECA BIS 1971 2011 Guatemala LMI LAC IFS 1970 2011 Guyana LMI LAC IFS 1970 2011 Honduras LMI LAC IFS 1970 2011 Iceland HI ECA IFS 1970 2011 India LMI EAP BIS 1970 2011 Indonesia LMI EAP BIS 1976 2011 Ireland HI ECA BIS 1972 2011 Italy HI ECA BIS 1970 2011 Japan HI EAP BIS 1970 2011 Jordan UMI MENA IFS 1970 2011 Kazakhstan LMI ECA IFS 1993 2011 Kenya LI SSA IFS 1970 2011 Korea, Rep. HI EAP BIS 1970 2011 Madagascar LI SSA IFS 1970 2011 Malaysia UMI EAP IFS 1970 2011 Mali LI SSA IFS 1970 2011 Mauritania LI MENA IFS 1970 2011 Mexico UMI LAC BIS 1991 2011 Mongolia UMI LAC IFS 1991 2011 Nepal LI EAP IFS 1970 2011 Netherlands HI ECA BIS 1970 2011 New Zealand HI EAP IFS 1970 2010 Nicaragua LMI LAC IFS 1970 2011 Niger LI SSA IFS 1970 2011 Nigeria LMI SSA IFS 1970 2011 Norway HI ECA BIS 1970 2011 Panama LMI LAC IFS 1970 2011 Paraguay LMI LAC IFS 1970 2011 Peru UMI LAC IFS 1970 2011 Philippines LMI EAP IFS 1970 2011 Portugal HI ECA BIS 1970 2011 Senegal LMI SSA IFS 1970 2011 Sierra Leone LI SSA IFS 1970 2011 Singapore HI EAP BIS 1971 2011 South Africa LMI SSA BIS 1970 2011 Spain HI ECA BIS 1972 2011 Sri Lanka LMI EAP IFS 1970 2011 Swaziland LMI SSA IFS 1973 2011 Sweden HI ECA BIS 1970 2011 Switzerland HI ECA BIS 1970 2011 Thailand UMI EAP BIS 1970 2011 Togo LI SSA IFS 1970 2011 Tunisia UMI MENA IFS 1970 2011 Turkey UMI ECA IFS 1970 2011 UK HI ECA BIS 1970 2011 USA HI MNA BIS 1970 2011 Uruguay HI LAC IFS 1970 2011 Venezuela, RB UMI LAC IFS 1970 2011 Note: LI, LMI, UMI, HI stand for low income, lower middle income, upper middle income and high income, respectively. SSA, MENA, LAC, EAP, SA, NA and ECA stand for sub-Saharan Africa, Middle-East and North Africa, Latina America and Caribbean, East Asia and Pacific, South Asia, North America, and Europe and Central Asia, respectively. Income level classification is from the World Bank and refers to 2013. View Large Table 1: Sample Composition Country Income Region Data Source First Observation Last Observation Algeria UMI MENA IFS 1970 2011 Argentina UMI LAC IFS 1970 2011 Australia HI EAP BIS 1970 2011 Austria HI ECA BIS 1970 2011 Bangladesh LI EAP IFS 1974 2011 Belgium HI ECA BIS 1971 2011 Benin LI SSA IFS 1970 2011 Bolivia LMI LAC IFS 1970 2011 Brazil UMI LAC IFS 1970 1985 Burkina Faso LI SSA IFS 1970 2011 Burundi LI SSA IFS 1970 2011 Cameroon HI MNA IFS 1970 2011 Canada LMI SSA BIS 1970 2011 Cape Verde UMI SSA IFS 1980 2011 Centr. Afric. Rep. LI SSA IFS 1970 2011 Chad LI SSA IFS 1970 2011 Chile HI LAC IFS 1970 2011 Congo, Dem. Rep. LI SSA IFS 1970 1995 Congo, Rep. LMI SSA IFS 1970 2011 Costa Rica UMI LAC IFS 1970 2011 Cote d’Ivoire LMI SSA IFS 1970 2011 Denmark HI ECA BIS 1970 2011 Dominican Republic UMI LAC IFS 1970 2011 Ecuador UMI LAC IFS 1970 2011 Egypt, Arab Rep. LMI MENA IFS 1970 2011 El Salvador LMI LAC IFS 1970 2011 Finland HI ECA BIS 1971 2011 France HI ECA BIS 1970 2011 Germany HI ECA BIS 1970 2011 Ghana LMI SSA IFS 1970 2011 Greece HI ECA BIS 1971 2011 Guatemala LMI LAC IFS 1970 2011 Guyana LMI LAC IFS 1970 2011 Honduras LMI LAC IFS 1970 2011 Iceland HI ECA IFS 1970 2011 India LMI EAP BIS 1970 2011 Indonesia LMI EAP BIS 1976 2011 Ireland HI ECA BIS 1972 2011 Italy HI ECA BIS 1970 2011 Japan HI EAP BIS 1970 2011 Jordan UMI MENA IFS 1970 2011 Kazakhstan LMI ECA IFS 1993 2011 Kenya LI SSA IFS 1970 2011 Korea, Rep. HI EAP BIS 1970 2011 Madagascar LI SSA IFS 1970 2011 Malaysia UMI EAP IFS 1970 2011 Mali LI SSA IFS 1970 2011 Mauritania LI MENA IFS 1970 2011 Mexico UMI LAC BIS 1991 2011 Mongolia UMI LAC IFS 1991 2011 Nepal LI EAP IFS 1970 2011 Netherlands HI ECA BIS 1970 2011 New Zealand HI EAP IFS 1970 2010 Nicaragua LMI LAC IFS 1970 2011 Niger LI SSA IFS 1970 2011 Nigeria LMI SSA IFS 1970 2011 Norway HI ECA BIS 1970 2011 Panama LMI LAC IFS 1970 2011 Paraguay LMI LAC IFS 1970 2011 Peru UMI LAC IFS 1970 2011 Philippines LMI EAP IFS 1970 2011 Portugal HI ECA BIS 1970 2011 Senegal LMI SSA IFS 1970 2011 Sierra Leone LI SSA IFS 1970 2011 Singapore HI EAP BIS 1971 2011 South Africa LMI SSA BIS 1970 2011 Spain HI ECA BIS 1972 2011 Sri Lanka LMI EAP IFS 1970 2011 Swaziland LMI SSA IFS 1973 2011 Sweden HI ECA BIS 1970 2011 Switzerland HI ECA BIS 1970 2011 Thailand UMI EAP BIS 1970 2011 Togo LI SSA IFS 1970 2011 Tunisia UMI MENA IFS 1970 2011 Turkey UMI ECA IFS 1970 2011 UK HI ECA BIS 1970 2011 USA HI MNA BIS 1970 2011 Uruguay HI LAC IFS 1970 2011 Venezuela, RB UMI LAC IFS 1970 2011 Country Income Region Data Source First Observation Last Observation Algeria UMI MENA IFS 1970 2011 Argentina UMI LAC IFS 1970 2011 Australia HI EAP BIS 1970 2011 Austria HI ECA BIS 1970 2011 Bangladesh LI EAP IFS 1974 2011 Belgium HI ECA BIS 1971 2011 Benin LI SSA IFS 1970 2011 Bolivia LMI LAC IFS 1970 2011 Brazil UMI LAC IFS 1970 1985 Burkina Faso LI SSA IFS 1970 2011 Burundi LI SSA IFS 1970 2011 Cameroon HI MNA IFS 1970 2011 Canada LMI SSA BIS 1970 2011 Cape Verde UMI SSA IFS 1980 2011 Centr. Afric. Rep. LI SSA IFS 1970 2011 Chad LI SSA IFS 1970 2011 Chile HI LAC IFS 1970 2011 Congo, Dem. Rep. LI SSA IFS 1970 1995 Congo, Rep. LMI SSA IFS 1970 2011 Costa Rica UMI LAC IFS 1970 2011 Cote d’Ivoire LMI SSA IFS 1970 2011 Denmark HI ECA BIS 1970 2011 Dominican Republic UMI LAC IFS 1970 2011 Ecuador UMI LAC IFS 1970 2011 Egypt, Arab Rep. LMI MENA IFS 1970 2011 El Salvador LMI LAC IFS 1970 2011 Finland HI ECA BIS 1971 2011 France HI ECA BIS 1970 2011 Germany HI ECA BIS 1970 2011 Ghana LMI SSA IFS 1970 2011 Greece HI ECA BIS 1971 2011 Guatemala LMI LAC IFS 1970 2011 Guyana LMI LAC IFS 1970 2011 Honduras LMI LAC IFS 1970 2011 Iceland HI ECA IFS 1970 2011 India LMI EAP BIS 1970 2011 Indonesia LMI EAP BIS 1976 2011 Ireland HI ECA BIS 1972 2011 Italy HI ECA BIS 1970 2011 Japan HI EAP BIS 1970 2011 Jordan UMI MENA IFS 1970 2011 Kazakhstan LMI ECA IFS 1993 2011 Kenya LI SSA IFS 1970 2011 Korea, Rep. HI EAP BIS 1970 2011 Madagascar LI SSA IFS 1970 2011 Malaysia UMI EAP IFS 1970 2011 Mali LI SSA IFS 1970 2011 Mauritania LI MENA IFS 1970 2011 Mexico UMI LAC BIS 1991 2011 Mongolia UMI LAC IFS 1991 2011 Nepal LI EAP IFS 1970 2011 Netherlands HI ECA BIS 1970 2011 New Zealand HI EAP IFS 1970 2010 Nicaragua LMI LAC IFS 1970 2011 Niger LI SSA IFS 1970 2011 Nigeria LMI SSA IFS 1970 2011 Norway HI ECA BIS 1970 2011 Panama LMI LAC IFS 1970 2011 Paraguay LMI LAC IFS 1970 2011 Peru UMI LAC IFS 1970 2011 Philippines LMI EAP IFS 1970 2011 Portugal HI ECA BIS 1970 2011 Senegal LMI SSA IFS 1970 2011 Sierra Leone LI SSA IFS 1970 2011 Singapore HI EAP BIS 1971 2011 South Africa LMI SSA BIS 1970 2011 Spain HI ECA BIS 1972 2011 Sri Lanka LMI EAP IFS 1970 2011 Swaziland LMI SSA IFS 1973 2011 Sweden HI ECA BIS 1970 2011 Switzerland HI ECA BIS 1970 2011 Thailand UMI EAP BIS 1970 2011 Togo LI SSA IFS 1970 2011 Tunisia UMI MENA IFS 1970 2011 Turkey UMI ECA IFS 1970 2011 UK HI ECA BIS 1970 2011 USA HI MNA BIS 1970 2011 Uruguay HI LAC IFS 1970 2011 Venezuela, RB UMI LAC IFS 1970 2011 Note: LI, LMI, UMI, HI stand for low income, lower middle income, upper middle income and high income, respectively. SSA, MENA, LAC, EAP, SA, NA and ECA stand for sub-Saharan Africa, Middle-East and North Africa, Latina America and Caribbean, East Asia and Pacific, South Asia, North America, and Europe and Central Asia, respectively. Income level classification is from the World Bank and refers to 2013. View Large The cyclical measure of excess credit tested in this section is the BIS gap, the most widely used leading indicator of financial crisis in the EWS literature. The gap for each country is computed as the percentage deviation of the credit-to-GDP ratio from its long-term trend. The series of private credit-to-GDP is derived from two alternative sources, depending on data availability. For 27 countries,2 mainly advanced economies and large emerging markets, we use the new BIS series of total credit to the private sector, adjusted for structural breaks.3 Quarterly data have been averaged to form an annual series. For the remaining countries, we use the annual series of domestic credit to the private non-financial sector from the World Bank Global Financial Development Database.4 We require that the series of each country does not include data gaps and that data are available for at least 10 years prior to a crisis. The estimates of the long-term trend are obtained by using a one-sided (backward-looking) HP filter, which allows us to obtain real-time estimates of the gap.5 The smoothing parameter λ is set equal to 1600.6 Finally, to obtain more robust estimates of the trend, we start the computation 10 years after the beginning of each series so that the HP filter can use a minimum of 10 observations for each data point estimate of the trend. As far as regards structural indicators, we use a gap (frontier gap hereafter) computed as the percentage deviation of a country’s credit-to-GDP ratio from its financial possibility frontier (or benchmark), as retrieved from FinStats. This is a tool developed by the World Bank and updated every year, that implements the methodology in Beck et al. (2008) and estimates frontiers for the quasi-totality of countries in the world (177 countries) through a pooled quantile (median) regression (Feyen et al., 2015). Frontiers for each country are obtained by regressing private credit to GDP on a set of structural characteristics, including GDP per capita and its square (to account for potential non-linearities linking economic and financial development, see e.g., Arcand et al., 2015), population size and density, the age dependency ratio and year-fixed effects. Regressors also include a number of dummies to control for additional structural factors, including a dummy for natural resource exporters, as worldwide evidence shows that resource rich countries tend to have comparatively smaller financial sectors than other countries at similar levels of income, reflecting the fact that oil revenues can boost GDP out of proportion with the country’s overall level of economic and financial development. The identification and timing of banking crises are based on the systemic banking crises database by Laeven and Valencia (2013). The authors define a banking crisis as an event that satisfies two conditions: (a) significant signs of financial distress in the banking system (as indicated by significant bank runs, losses in the banking system, and/or bank liquidations) and (b) significant policy intervention measures in response to large losses in the banking sector.7 Given the paucity of banking crises based on this definition, we extend the Laeven and Valencia (2013) database based on Reinhart (2010) that uses an alternative definition of banking crisis, as an event that satisfies one of the following two conditions: (a) bank runs that lead to the closure, merging, or takeover by the public sector of one or more financial institutions or (b) if there are no runs, the closure, merger, takeover, or large-scale government support of the banking sector (Reinhart and Rogoff, 2009). Table 2 lists the crisis episodes used in this paper (first column), satisfying either Reinhart (2010) (second column of the Table) or Laeven and Valencia (2013) (third column) definitions. Table 2: Systemic Banking Crises, 1980–2010 Country Paper1 R2 LV3 Start End Start End Start End Algeria 1990 1994 1990 1992 1990 1994 Argentina 1980 1982 1980 1982 1980 1982 Argentina 1989 1991 1989 1990 1989 1991 Argentina 1995 1996 1995 1996 1995 1995 Argentina 2001 2003 2001 2003 2001 2003 Australia 1989 1992 1989 1992 n.a. n.a. Austria 2008 2011 2008 2010 2008 2011 Bangladesh 1987 1987 n.a. n.a. 1987 1987 Belgium 2008 2011 2008 2010 2008 2011 Benin 1988 1992 n.a. n.a. 1988 1992 Bolivia 1986 1987 1986 1987 1986 1986 Bolivia 1994 1997 1994 1997 1994 1994 Bolivia 1999 1999 1999 1999 n.a. n.a. Brazil 1985 1985 1985 1985 n.a. n.a. Brazil 1990 1994 1990 1990 1990 1994 Burkina Faso 1990 1994 n.a. n.a. 1990 1994 Burundi 1994 1998 n.a. n.a. 1994 1998 Canada 1983 1985 1983 1985 n.a. n.a. Cape Verde 1993 1993 n.a. n.a. 1993 1993 CAR 1980 1982 1976 1982 1976 1976 CAR 1988 1999 1988 1999 1995 1996 Chad 1983 1983 n.a. n.a. 1983 1983 Chad 1992 1996 n.a. n.a. 1992 1996 Chile 1981 1985 1982 1984 1981 1985 Congo, Dem. Rep. 1983 1983 n.a. n.a. 1983 1983 Congo, Dem. Rep. 1991 1994 n.a. n.a. 1991 1994 Congo, Dem. Rep. 1994 1998 n.a. n.a. 1994 1998 Congo, Rep. 1992 1994 n.a. n.a. 1992 1994 Costa Rica 1987 1991 1987 1987 1987 1991 Costa Rica 1994 1996 1994 1996 1994 1995 Cote dIvoire 1988 1992 1988 1991 1988 1992 Denmark 1987 1992 1987 1992 n.a. n.a. Denmark 2008 2011 2008 2010 2008 2011 Dominican Rep 1996 1996 1996 1996 n.a. n.a. Dominican Rep 2003 2004 2003 2003 2003 2004 Ecuador 1981 1986 1981 1981 1982 1986 Ecuador 1998 2002 1998 2002 1998 2002 Egypt 1980 1983 1981 1983 1980 1980 Egypt 1990 1995 1990 1995 n.a. n.a. El Salvador 1989 1990 1989 1989 1989 1990 Finland 1991 1995 1991 1994 1991 1995 France 1994 1995 1994 1995 n.a. n.a. France 2008 2011 2008 2010 2008 2011 Germany 1991 1994 1991 1994 n.a. n.a. Germany 2008 2011 n.a. n.a. 2008 2011 Ghana 1982 1989 1982 1989 1982 1983 Ghana 1997 1997 1997 1997 n.a. n.a. Greece 1991 1995 1991 1995 n.a. n.a. Greece 2008 2011 2008 2010 2008 2011 Guatemala 1990 1990 1990 1990 n.a. n.a. Guatemala 2001 2001 2001 2001 n.a. n.a. Guatemala 2006 2006 2006 2006 n.a. n.a. Guyana 1993 1993 n.a. n.a. 1993 1993 Honduras 1999 1999 1999 1999 n.a. n.a. Honduras 2001 2002 2001 2002 n.a. n.a. Iceland 1985 1986 1985 1986 n.a. n.a. Iceland 1993 1993 1993 1993 n.a. n.a. Iceland 2008 2011 2007 2010 2008 2011 India 1993 1998 1993 1998 1993 1993 Indonesia 1992 1992 1992 1992 n.a. n.a. Indonesia 1994 1994 1994 1994 n.a. n.a. Indonesia 1997 2002 1997 2002 1997 2001 Ireland 2007 2011 2007 2010 2008 2011 Italy 1990 1995 1990 1995 n.a. n.a. Italy 2008 2011 n.a. n.a. 2008 2011 Japan 1992 2001 1992 2001 1997 2001 Jordan 1989 1991 n.a. n.a. 1989 1991 Kazakhstan 2008 2011 n.a. n.a. 2008 2011 Kenya 1985 1989 1985 1989 1985 1985 Kenya 1992 1995 1992 1995 1992 1994 Korea 1983 1983 1983 1983 n.a. n.a. Korea 1985 1988 1985 1988 n.a. n.a. Korea 1997 2002 1997 2002 1997 1998 Madagascar 1988 1988 n.a. n.a. 1988 1988 Malaysia 1985 1988 1985 1988 n.a. n.a. Malaysia 1997 2001 1997 2001 1997 1999 Mali 1987 1991 n.a. n.a. 1987 1991 Mauritania 1984 1984 n.a. n.a. 1984 1984 Mexico 1981 1985 1981 1982 1981 1985 Mexico 1994 2000 1994 2000 1994 1996 Mongolia 2008 2011 n.a. n.a. 2008 2011 Nepal 1988 1988 n.a. n.a. 1988 1988 Netherlands 2008 2011 2008 2010 2008 2011 New Zealand 1987 1990 1987 1990 n.a. n.a. Nicaragua 1987 1996 1987 1996 n.a. n.a. Nicaragua 1990 1993 n.a. n.a. 1990 1993 Nicaragua 2000 2002 2000 2002 2000 2001 Niger 1983 1985 n.a. n.a. 1983 1985 Nigeria 1992 1995 1992 1995 1991 1995 Nigeria 1997 1997 1997 1997 n.a. n.a. Nigeria 2009 2011 n.a. n.a. 2009 2011 Norway 1987 1993 1987 1993 1991 1993 Panama 1988 1989 1988 1989 1988 1989 Paraguay 1995 1999 1995 1999 1995 1995 Paraguay 2002 2002 2002 2002 n.a. n.a. Peru 1983 1990 1983 1990 1983 1983 Peru 1999 1999 1999 1999 n.a. n.a. Philippines 1981 1987 1981 1987 1983 1986 Philippines 1997 2001 1997 2001 1997 2001 Portugal 2008 2011 2008 2010 2008 2011 Senegal 1988 1991 n.a. n.a. 1988 1991 Sierra Leone 1990 1994 n.a. n.a. 1990 1994 Singapore 1982 1983 1982 1983 n.a. n.a. South Africa 1989 1989 1989 1989 n.a. n.a. Spain 1980 1985 1977 1985 1977 1981 Spain 2008 2011 2008 2010 2008 2011 Sri Lanka 1989 1993 1989 1993 1989 1991 Swaziland 1995 1999 n.a. n.a. 1995 1999 Sweden 1991 1995 1991 1994 1991 1995 Sweden 2008 2011 n.a. n.a. 2008 2011 Switzerland 2008 2011 2008 2009 2008 2011 Thailand 1980 1987 1980 1987 1983 1983 Thailand 1996 2001 1996 2001 1997 2000 Togo 1993 1994 n.a. n.a. 1993 1994 Tunisia 1991 1991 n.a. n.a. 1991 1991 Turkey 1982 1985 1982 1985 1982 1984 Turkey 1991 1991 1991 1991 n.a. n.a. Turkey 1994 1994 1994 1994 n.a. n.a. Turkey 2000 2001 2000 2000 2000 2001 UK 1984 1984 1984 1984 n.a. n.a. UK 1991 1991 1991 1991 n.a. n.a. UK 1995 1995 1995 1995 n.a. n.a. UK 2007 2011 2007 2009 2007 2011 USA 1984 1991 1984 1991 1988 1988 USA 2007 2011 2007 2010 2007 2011 Uruguay 1981 1985 n.a. n.a. 1981 1985 Uruguay 2002 2005 n.a. n.a. 2002 2005 Venezuela 1980 1986 1978 1986 n.a. n.a. Venezuela 1993 1998 1993 1994 1994 1998 Country Paper1 R2 LV3 Start End Start End Start End Algeria 1990 1994 1990 1992 1990 1994 Argentina 1980 1982 1980 1982 1980 1982 Argentina 1989 1991 1989 1990 1989 1991 Argentina 1995 1996 1995 1996 1995 1995 Argentina 2001 2003 2001 2003 2001 2003 Australia 1989 1992 1989 1992 n.a. n.a. Austria 2008 2011 2008 2010 2008 2011 Bangladesh 1987 1987 n.a. n.a. 1987 1987 Belgium 2008 2011 2008 2010 2008 2011 Benin 1988 1992 n.a. n.a. 1988 1992 Bolivia 1986 1987 1986 1987 1986 1986 Bolivia 1994 1997 1994 1997 1994 1994 Bolivia 1999 1999 1999 1999 n.a. n.a. Brazil 1985 1985 1985 1985 n.a. n.a. Brazil 1990 1994 1990 1990 1990 1994 Burkina Faso 1990 1994 n.a. n.a. 1990 1994 Burundi 1994 1998 n.a. n.a. 1994 1998 Canada 1983 1985 1983 1985 n.a. n.a. Cape Verde 1993 1993 n.a. n.a. 1993 1993 CAR 1980 1982 1976 1982 1976 1976 CAR 1988 1999 1988 1999 1995 1996 Chad 1983 1983 n.a. n.a. 1983 1983 Chad 1992 1996 n.a. n.a. 1992 1996 Chile 1981 1985 1982 1984 1981 1985 Congo, Dem. Rep. 1983 1983 n.a. n.a. 1983 1983 Congo, Dem. Rep. 1991 1994 n.a. n.a. 1991 1994 Congo, Dem. Rep. 1994 1998 n.a. n.a. 1994 1998 Congo, Rep. 1992 1994 n.a. n.a. 1992 1994 Costa Rica 1987 1991 1987 1987 1987 1991 Costa Rica 1994 1996 1994 1996 1994 1995 Cote dIvoire 1988 1992 1988 1991 1988 1992 Denmark 1987 1992 1987 1992 n.a. n.a. Denmark 2008 2011 2008 2010 2008 2011 Dominican Rep 1996 1996 1996 1996 n.a. n.a. Dominican Rep 2003 2004 2003 2003 2003 2004 Ecuador 1981 1986 1981 1981 1982 1986 Ecuador 1998 2002 1998 2002 1998 2002 Egypt 1980 1983 1981 1983 1980 1980 Egypt 1990 1995 1990 1995 n.a. n.a. El Salvador 1989 1990 1989 1989 1989 1990 Finland 1991 1995 1991 1994 1991 1995 France 1994 1995 1994 1995 n.a. n.a. France 2008 2011 2008 2010 2008 2011 Germany 1991 1994 1991 1994 n.a. n.a. Germany 2008 2011 n.a. n.a. 2008 2011 Ghana 1982 1989 1982 1989 1982 1983 Ghana 1997 1997 1997 1997 n.a. n.a. Greece 1991 1995 1991 1995 n.a. n.a. Greece 2008 2011 2008 2010 2008 2011 Guatemala 1990 1990 1990 1990 n.a. n.a. Guatemala 2001 2001 2001 2001 n.a. n.a. Guatemala 2006 2006 2006 2006 n.a. n.a. Guyana 1993 1993 n.a. n.a. 1993 1993 Honduras 1999 1999 1999 1999 n.a. n.a. Honduras 2001 2002 2001 2002 n.a. n.a. Iceland 1985 1986 1985 1986 n.a. n.a. Iceland 1993 1993 1993 1993 n.a. n.a. Iceland 2008 2011 2007 2010 2008 2011 India 1993 1998 1993 1998 1993 1993 Indonesia 1992 1992 1992 1992 n.a. n.a. Indonesia 1994 1994 1994 1994 n.a. n.a. Indonesia 1997 2002 1997 2002 1997 2001 Ireland 2007 2011 2007 2010 2008 2011 Italy 1990 1995 1990 1995 n.a. n.a. Italy 2008 2011 n.a. n.a. 2008 2011 Japan 1992 2001 1992 2001 1997 2001 Jordan 1989 1991 n.a. n.a. 1989 1991 Kazakhstan 2008 2011 n.a. n.a. 2008 2011 Kenya 1985 1989 1985 1989 1985 1985 Kenya 1992 1995 1992 1995 1992 1994 Korea 1983 1983 1983 1983 n.a. n.a. Korea 1985 1988 1985 1988 n.a. n.a. Korea 1997 2002 1997 2002 1997 1998 Madagascar 1988 1988 n.a. n.a. 1988 1988 Malaysia 1985 1988 1985 1988 n.a. n.a. Malaysia 1997 2001 1997 2001 1997 1999 Mali 1987 1991 n.a. n.a. 1987 1991 Mauritania 1984 1984 n.a. n.a. 1984 1984 Mexico 1981 1985 1981 1982 1981 1985 Mexico 1994 2000 1994 2000 1994 1996 Mongolia 2008 2011 n.a. n.a. 2008 2011 Nepal 1988 1988 n.a. n.a. 1988 1988 Netherlands 2008 2011 2008 2010 2008 2011 New Zealand 1987 1990 1987 1990 n.a. n.a. Nicaragua 1987 1996 1987 1996 n.a. n.a. Nicaragua 1990 1993 n.a. n.a. 1990 1993 Nicaragua 2000 2002 2000 2002 2000 2001 Niger 1983 1985 n.a. n.a. 1983 1985 Nigeria 1992 1995 1992 1995 1991 1995 Nigeria 1997 1997 1997 1997 n.a. n.a. Nigeria 2009 2011 n.a. n.a. 2009 2011 Norway 1987 1993 1987 1993 1991 1993 Panama 1988 1989 1988 1989 1988 1989 Paraguay 1995 1999 1995 1999 1995 1995 Paraguay 2002 2002 2002 2002 n.a. n.a. Peru 1983 1990 1983 1990 1983 1983 Peru 1999 1999 1999 1999 n.a. n.a. Philippines 1981 1987 1981 1987 1983 1986 Philippines 1997 2001 1997 2001 1997 2001 Portugal 2008 2011 2008 2010 2008 2011 Senegal 1988 1991 n.a. n.a. 1988 1991 Sierra Leone 1990 1994 n.a. n.a. 1990 1994 Singapore 1982 1983 1982 1983 n.a. n.a. South Africa 1989 1989 1989 1989 n.a. n.a. Spain 1980 1985 1977 1985 1977 1981 Spain 2008 2011 2008 2010 2008 2011 Sri Lanka 1989 1993 1989 1993 1989 1991 Swaziland 1995 1999 n.a. n.a. 1995 1999 Sweden 1991 1995 1991 1994 1991 1995 Sweden 2008 2011 n.a. n.a. 2008 2011 Switzerland 2008 2011 2008 2009 2008 2011 Thailand 1980 1987 1980 1987 1983 1983 Thailand 1996 2001 1996 2001 1997 2000 Togo 1993 1994 n.a. n.a. 1993 1994 Tunisia 1991 1991 n.a. n.a. 1991 1991 Turkey 1982 1985 1982 1985 1982 1984 Turkey 1991 1991 1991 1991 n.a. n.a. Turkey 1994 1994 1994 1994 n.a. n.a. Turkey 2000 2001 2000 2000 2000 2001 UK 1984 1984 1984 1984 n.a. n.a. UK 1991 1991 1991 1991 n.a. n.a. UK 1995 1995 1995 1995 n.a. n.a. UK 2007 2011 2007 2009 2007 2011 USA 1984 1991 1984 1991 1988 1988 USA 2007 2011 2007 2010 2007 2011 Uruguay 1981 1985 n.a. n.a. 1981 1985 Uruguay 2002 2005 n.a. n.a. 2002 2005 Venezuela 1980 1986 1978 1986 n.a. n.a. Venezuela 1993 1998 1993 1994 1994 1998 1Crisis period used in the paper, equal to R ∪ LV. 2Crisis period according to Reinhart (2010). 3Crisis period according to Laeven and Valencia (2013). View Large Table 2: Systemic Banking Crises, 1980–2010 Country Paper1 R2 LV3 Start End Start End Start End Algeria 1990 1994 1990 1992 1990 1994 Argentina 1980 1982 1980 1982 1980 1982 Argentina 1989 1991 1989 1990 1989 1991 Argentina 1995 1996 1995 1996 1995 1995 Argentina 2001 2003 2001 2003 2001 2003 Australia 1989 1992 1989 1992 n.a. n.a. Austria 2008 2011 2008 2010 2008 2011 Bangladesh 1987 1987 n.a. n.a. 1987 1987 Belgium 2008 2011 2008 2010 2008 2011 Benin 1988 1992 n.a. n.a. 1988 1992 Bolivia 1986 1987 1986 1987 1986 1986 Bolivia 1994 1997 1994 1997 1994 1994 Bolivia 1999 1999 1999 1999 n.a. n.a. Brazil 1985 1985 1985 1985 n.a. n.a. Brazil 1990 1994 1990 1990 1990 1994 Burkina Faso 1990 1994 n.a. n.a. 1990 1994 Burundi 1994 1998 n.a. n.a. 1994 1998 Canada 1983 1985 1983 1985 n.a. n.a. Cape Verde 1993 1993 n.a. n.a. 1993 1993 CAR 1980 1982 1976 1982 1976 1976 CAR 1988 1999 1988 1999 1995 1996 Chad 1983 1983 n.a. n.a. 1983 1983 Chad 1992 1996 n.a. n.a. 1992 1996 Chile 1981 1985 1982 1984 1981 1985 Congo, Dem. Rep. 1983 1983 n.a. n.a. 1983 1983 Congo, Dem. Rep. 1991 1994 n.a. n.a. 1991 1994 Congo, Dem. Rep. 1994 1998 n.a. n.a. 1994 1998 Congo, Rep. 1992 1994 n.a. n.a. 1992 1994 Costa Rica 1987 1991 1987 1987 1987 1991 Costa Rica 1994 1996 1994 1996 1994 1995 Cote dIvoire 1988 1992 1988 1991 1988 1992 Denmark 1987 1992 1987 1992 n.a. n.a. Denmark 2008 2011 2008 2010 2008 2011 Dominican Rep 1996 1996 1996 1996 n.a. n.a. Dominican Rep 2003 2004 2003 2003 2003 2004 Ecuador 1981 1986 1981 1981 1982 1986 Ecuador 1998 2002 1998 2002 1998 2002 Egypt 1980 1983 1981 1983 1980 1980 Egypt 1990 1995 1990 1995 n.a. n.a. El Salvador 1989 1990 1989 1989 1989 1990 Finland 1991 1995 1991 1994 1991 1995 France 1994 1995 1994 1995 n.a. n.a. France 2008 2011 2008 2010 2008 2011 Germany 1991 1994 1991 1994 n.a. n.a. Germany 2008 2011 n.a. n.a. 2008 2011 Ghana 1982 1989 1982 1989 1982 1983 Ghana 1997 1997 1997 1997 n.a. n.a. Greece 1991 1995 1991 1995 n.a. n.a. Greece 2008 2011 2008 2010 2008 2011 Guatemala 1990 1990 1990 1990 n.a. n.a. Guatemala 2001 2001 2001 2001 n.a. n.a. Guatemala 2006 2006 2006 2006 n.a. n.a. Guyana 1993 1993 n.a. n.a. 1993 1993 Honduras 1999 1999 1999 1999 n.a. n.a. Honduras 2001 2002 2001 2002 n.a. n.a. Iceland 1985 1986 1985 1986 n.a. n.a. Iceland 1993 1993 1993 1993 n.a. n.a. Iceland 2008 2011 2007 2010 2008 2011 India 1993 1998 1993 1998 1993 1993 Indonesia 1992 1992 1992 1992 n.a. n.a. Indonesia 1994 1994 1994 1994 n.a. n.a. Indonesia 1997 2002 1997 2002 1997 2001 Ireland 2007 2011 2007 2010 2008 2011 Italy 1990 1995 1990 1995 n.a. n.a. Italy 2008 2011 n.a. n.a. 2008 2011 Japan 1992 2001 1992 2001 1997 2001 Jordan 1989 1991 n.a. n.a. 1989 1991 Kazakhstan 2008 2011 n.a. n.a. 2008 2011 Kenya 1985 1989 1985 1989 1985 1985 Kenya 1992 1995 1992 1995 1992 1994 Korea 1983 1983 1983 1983 n.a. n.a. Korea 1985 1988 1985 1988 n.a. n.a. Korea 1997 2002 1997 2002 1997 1998 Madagascar 1988 1988 n.a. n.a. 1988 1988 Malaysia 1985 1988 1985 1988 n.a. n.a. Malaysia 1997 2001 1997 2001 1997 1999 Mali 1987 1991 n.a. n.a. 1987 1991 Mauritania 1984 1984 n.a. n.a. 1984 1984 Mexico 1981 1985 1981 1982 1981 1985 Mexico 1994 2000 1994 2000 1994 1996 Mongolia 2008 2011 n.a. n.a. 2008 2011 Nepal 1988 1988 n.a. n.a. 1988 1988 Netherlands 2008 2011 2008 2010 2008 2011 New Zealand 1987 1990 1987 1990 n.a. n.a. Nicaragua 1987 1996 1987 1996 n.a. n.a. Nicaragua 1990 1993 n.a. n.a. 1990 1993 Nicaragua 2000 2002 2000 2002 2000 2001 Niger 1983 1985 n.a. n.a. 1983 1985 Nigeria 1992 1995 1992 1995 1991 1995 Nigeria 1997 1997 1997 1997 n.a. n.a. Nigeria 2009 2011 n.a. n.a. 2009 2011 Norway 1987 1993 1987 1993 1991 1993 Panama 1988 1989 1988 1989 1988 1989 Paraguay 1995 1999 1995 1999 1995 1995 Paraguay 2002 2002 2002 2002 n.a. n.a. Peru 1983 1990 1983 1990 1983 1983 Peru 1999 1999 1999 1999 n.a. n.a. Philippines 1981 1987 1981 1987 1983 1986 Philippines 1997 2001 1997 2001 1997 2001 Portugal 2008 2011 2008 2010 2008 2011 Senegal 1988 1991 n.a. n.a. 1988 1991 Sierra Leone 1990 1994 n.a. n.a. 1990 1994 Singapore 1982 1983 1982 1983 n.a. n.a. South Africa 1989 1989 1989 1989 n.a. n.a. Spain 1980 1985 1977 1985 1977 1981 Spain 2008 2011 2008 2010 2008 2011 Sri Lanka 1989 1993 1989 1993 1989 1991 Swaziland 1995 1999 n.a. n.a. 1995 1999 Sweden 1991 1995 1991 1994 1991 1995 Sweden 2008 2011 n.a. n.a. 2008 2011 Switzerland 2008 2011 2008 2009 2008 2011 Thailand 1980 1987 1980 1987 1983 1983 Thailand 1996 2001 1996 2001 1997 2000 Togo 1993 1994 n.a. n.a. 1993 1994 Tunisia 1991 1991 n.a. n.a. 1991 1991 Turkey 1982 1985 1982 1985 1982 1984 Turkey 1991 1991 1991 1991 n.a. n.a. Turkey 1994 1994 1994 1994 n.a. n.a. Turkey 2000 2001 2000 2000 2000 2001 UK 1984 1984 1984 1984 n.a. n.a. UK 1991 1991 1991 1991 n.a. n.a. UK 1995 1995 1995 1995 n.a. n.a. UK 2007 2011 2007 2009 2007 2011 USA 1984 1991 1984 1991 1988 1988 USA 2007 2011 2007 2010 2007 2011 Uruguay 1981 1985 n.a. n.a. 1981 1985 Uruguay 2002 2005 n.a. n.a. 2002 2005 Venezuela 1980 1986 1978 1986 n.a. n.a. Venezuela 1993 1998 1993 1994 1994 1998 Country Paper1 R2 LV3 Start End Start End Start End Algeria 1990 1994 1990 1992 1990 1994 Argentina 1980 1982 1980 1982 1980 1982 Argentina 1989 1991 1989 1990 1989 1991 Argentina 1995 1996 1995 1996 1995 1995 Argentina 2001 2003 2001 2003 2001 2003 Australia 1989 1992 1989 1992 n.a. n.a. Austria 2008 2011 2008 2010 2008 2011 Bangladesh 1987 1987 n.a. n.a. 1987 1987 Belgium 2008 2011 2008 2010 2008 2011 Benin 1988 1992 n.a. n.a. 1988 1992 Bolivia 1986 1987 1986 1987 1986 1986 Bolivia 1994 1997 1994 1997 1994 1994 Bolivia 1999 1999 1999 1999 n.a. n.a. Brazil 1985 1985 1985 1985 n.a. n.a. Brazil 1990 1994 1990 1990 1990 1994 Burkina Faso 1990 1994 n.a. n.a. 1990 1994 Burundi 1994 1998 n.a. n.a. 1994 1998 Canada 1983 1985 1983 1985 n.a. n.a. Cape Verde 1993 1993 n.a. n.a. 1993 1993 CAR 1980 1982 1976 1982 1976 1976 CAR 1988 1999 1988 1999 1995 1996 Chad 1983 1983 n.a. n.a. 1983 1983 Chad 1992 1996 n.a. n.a. 1992 1996 Chile 1981 1985 1982 1984 1981 1985 Congo, Dem. Rep. 1983 1983 n.a. n.a. 1983 1983 Congo, Dem. Rep. 1991 1994 n.a. n.a. 1991 1994 Congo, Dem. Rep. 1994 1998 n.a. n.a. 1994 1998 Congo, Rep. 1992 1994 n.a. n.a. 1992 1994 Costa Rica 1987 1991 1987 1987 1987 1991 Costa Rica 1994 1996 1994 1996 1994 1995 Cote dIvoire 1988 1992 1988 1991 1988 1992 Denmark 1987 1992 1987 1992 n.a. n.a. Denmark 2008 2011 2008 2010 2008 2011 Dominican Rep 1996 1996 1996 1996 n.a. n.a. Dominican Rep 2003 2004 2003 2003 2003 2004 Ecuador 1981 1986 1981 1981 1982 1986 Ecuador 1998 2002 1998 2002 1998 2002 Egypt 1980 1983 1981 1983 1980 1980 Egypt 1990 1995 1990 1995 n.a. n.a. El Salvador 1989 1990 1989 1989 1989 1990 Finland 1991 1995 1991 1994 1991 1995 France 1994 1995 1994 1995 n.a. n.a. France 2008 2011 2008 2010 2008 2011 Germany 1991 1994 1991 1994 n.a. n.a. Germany 2008 2011 n.a. n.a. 2008 2011 Ghana 1982 1989 1982 1989 1982 1983 Ghana 1997 1997 1997 1997 n.a. n.a. Greece 1991 1995 1991 1995 n.a. n.a. Greece 2008 2011 2008 2010 2008 2011 Guatemala 1990 1990 1990 1990 n.a. n.a. Guatemala 2001 2001 2001 2001 n.a. n.a. Guatemala 2006 2006 2006 2006 n.a. n.a. Guyana 1993 1993 n.a. n.a. 1993 1993 Honduras 1999 1999 1999 1999 n.a. n.a. Honduras 2001 2002 2001 2002 n.a. n.a. Iceland 1985 1986 1985 1986 n.a. n.a. Iceland 1993 1993 1993 1993 n.a. n.a. Iceland 2008 2011 2007 2010 2008 2011 India 1993 1998 1993 1998 1993 1993 Indonesia 1992 1992 1992 1992 n.a. n.a. Indonesia 1994 1994 1994 1994 n.a. n.a. Indonesia 1997 2002 1997 2002 1997 2001 Ireland 2007 2011 2007 2010 2008 2011 Italy 1990 1995 1990 1995 n.a. n.a. Italy 2008 2011 n.a. n.a. 2008 2011 Japan 1992 2001 1992 2001 1997 2001 Jordan 1989 1991 n.a. n.a. 1989 1991 Kazakhstan 2008 2011 n.a. n.a. 2008 2011 Kenya 1985 1989 1985 1989 1985 1985 Kenya 1992 1995 1992 1995 1992 1994 Korea 1983 1983 1983 1983 n.a. n.a. Korea 1985 1988 1985 1988 n.a. n.a. Korea 1997 2002 1997 2002 1997 1998 Madagascar 1988 1988 n.a. n.a. 1988 1988 Malaysia 1985 1988 1985 1988 n.a. n.a. Malaysia 1997 2001 1997 2001 1997 1999 Mali 1987 1991 n.a. n.a. 1987 1991 Mauritania 1984 1984 n.a. n.a. 1984 1984 Mexico 1981 1985 1981 1982 1981 1985 Mexico 1994 2000 1994 2000 1994 1996 Mongolia 2008 2011 n.a. n.a. 2008 2011 Nepal 1988 1988 n.a. n.a. 1988 1988 Netherlands 2008 2011 2008 2010 2008 2011 New Zealand 1987 1990 1987 1990 n.a. n.a. Nicaragua 1987 1996 1987 1996 n.a. n.a. Nicaragua 1990 1993 n.a. n.a. 1990 1993 Nicaragua 2000 2002 2000 2002 2000 2001 Niger 1983 1985 n.a. n.a. 1983 1985 Nigeria 1992 1995 1992 1995 1991 1995 Nigeria 1997 1997 1997 1997 n.a. n.a. Nigeria 2009 2011 n.a. n.a. 2009 2011 Norway 1987 1993 1987 1993 1991 1993 Panama 1988 1989 1988 1989 1988 1989 Paraguay 1995 1999 1995 1999 1995 1995 Paraguay 2002 2002 2002 2002 n.a. n.a. Peru 1983 1990 1983 1990 1983 1983 Peru 1999 1999 1999 1999 n.a. n.a. Philippines 1981 1987 1981 1987 1983 1986 Philippines 1997 2001 1997 2001 1997 2001 Portugal 2008 2011 2008 2010 2008 2011 Senegal 1988 1991 n.a. n.a. 1988 1991 Sierra Leone 1990 1994 n.a. n.a. 1990 1994 Singapore 1982 1983 1982 1983 n.a. n.a. South Africa 1989 1989 1989 1989 n.a. n.a. Spain 1980 1985 1977 1985 1977 1981 Spain 2008 2011 2008 2010 2008 2011 Sri Lanka 1989 1993 1989 1993 1989 1991 Swaziland 1995 1999 n.a. n.a. 1995 1999 Sweden 1991 1995 1991 1994 1991 1995 Sweden 2008 2011 n.a. n.a. 2008 2011 Switzerland 2008 2011 2008 2009 2008 2011 Thailand 1980 1987 1980 1987 1983 1983 Thailand 1996 2001 1996 2001 1997 2000 Togo 1993 1994 n.a. n.a. 1993 1994 Tunisia 1991 1991 n.a. n.a. 1991 1991 Turkey 1982 1985 1982 1985 1982 1984 Turkey 1991 1991 1991 1991 n.a. n.a. Turkey 1994 1994 1994 1994 n.a. n.a. Turkey 2000 2001 2000 2000 2000 2001 UK 1984 1984 1984 1984 n.a. n.a. UK 1991 1991 1991 1991 n.a. n.a. UK 1995 1995 1995 1995 n.a. n.a. UK 2007 2011 2007 2009 2007 2011 USA 1984 1991 1984 1991 1988 1988 USA 2007 2011 2007 2010 2007 2011 Uruguay 1981 1985 n.a. n.a. 1981 1985 Uruguay 2002 2005 n.a. n.a. 2002 2005 Venezuela 1980 1986 1978 1986 n.a. n.a. Venezuela 1993 1998 1993 1994 1994 1998 1Crisis period used in the paper, equal to R ∪ LV. 2Crisis period according to Reinhart (2010). 3Crisis period according to Laeven and Valencia (2013). View Large 3.2 Methodology To assess indicators’ capacity to predict banking crises, we use the signal detection approach pioneered by Kaminsky and Reinhart (1999), which is one of the most common methods for the statistical evaluation of early warning indicators (EWIs). According to this methodology, when an indicator takes a value that exceeds a certain threshold, this is a signal that the event of interest (in our case a financial crisis) will materialise within the forecast period. Comparing the signal with the actual realisation of the event allows an assessment of the predictive capacity of the indicator for a given threshold. Forecast errors can be divided in two categories: (a) the indicator has not signalled a crisis that actually occurred in the forecast horizon (missed crisis corresponding to Type I error) and (b) the indicator has incorrectly signalled a crisis that will not materialise in the forecast horizon (false alarm corresponding to Type II error). Then, for the assessment of the performance of the indicator on the full range of possible thresholds we use the area under the receiver operating characteristic (ROC) curve. This is a graphical instrument that permits to visualise an indicator’s signalling quality by using the statistical relation between the ‘true positive rate’ (fraction of crises correctly predicted) and the ‘false positive rate’ (fraction of cases incorrectly classified as crises). The first ratio, also called ‘sensitivity’, is an estimate of the probability of a positive signal conditional on the fact that a crisis will materialise, and relates to the test’s ability to identify a condition correctly. The second ratio, also called ‘specificity’, is an estimate of the probability of incorrectly predicting a crisis and relates to the test’s ability to exclude a condition correctly. These two ratios are not independent and are connected to type I and type II errors: the higher is the sensitivity of the indicator, the lower is its specificity, meaning that the indicator will correctly predict most of the crises but will also send many false alarms. The ROC curve synthesises this trade off by reporting the true positive and the false positive rates for any possible threshold (see Figure 3). For very conservative thresholds, we expect that both the positive rate and the false positive rate are high (i.e., the indicator detects all the crises but sends also many false alarms), for lax thresholds instead we expect that both the true and false positive rates are low (i.e., the indicator misses most of the crisis but does not send many false alarms). Figure 3: View largeDownload slide Comparing ROC Curves. The Figure Illustrates the ROC Curves Associated with Three Indicators with Different Forecasting Capacity. The Indicator That Has a ROC Curve Coincident with the Diagonal, with an Area Under the Curve (AUC) of 0.5, Does not Perform Better than Tossing a Coin Because for any Signal Issued by the Indicator There Is a 50:50 Chance of Predicting Correctly the Event of Interest (e.g., the True Positive Rate Is Equal to the False Positive Rate). The more the AUC Is Distant from 0.5 and the Closer to 1, the Higher Is the Forecasting Performance of the Indicator. Figure 3: View largeDownload slide Comparing ROC Curves. The Figure Illustrates the ROC Curves Associated with Three Indicators with Different Forecasting Capacity. The Indicator That Has a ROC Curve Coincident with the Diagonal, with an Area Under the Curve (AUC) of 0.5, Does not Perform Better than Tossing a Coin Because for any Signal Issued by the Indicator There Is a 50:50 Chance of Predicting Correctly the Event of Interest (e.g., the True Positive Rate Is Equal to the False Positive Rate). The more the AUC Is Distant from 0.5 and the Closer to 1, the Higher Is the Forecasting Performance of the Indicator. The area under the ROC curve (AUC) can be used as a convenient and interpretable summary measure of the discriminatory accuracy of an indicator. This varies between 0 and 1, with 0.5 indicating that the indicator is not informative, as for any positive signal the probability that the event of interest will materialise in the forecast horizon is equal to the probability of a false alarm. Indicators that are expected to increase ahead of the crisis have higher predictive performance the higher is the distance of the AUC from 0.5 and the closer to 1. For these indicators, AUC values below 0.5 are associated with classifiers that perform worse than random guessing as higher values in the prediction are associated with a lower probability that the event of interest will materialise. In our case, this would correspond to a situation where higher values of the gaps are associated with a lower probability of a crisis, which is counterintuitive and misleading. 3.3 Results of the analysis In this section, we contrast the predictive performance of the BIS and frontier gaps at different levels of financial development. The measure of financial development adopted hereafter is the private credit-to-GDP ratio at the time of the crisis,8 which relates the extent of financial intermediation to the size of the economy. As in Drehmann and Juselius (2014), the BIS and frontier gaps are evaluated using both a static and a dynamic approach. The static assessment provides information on the statistical forecasting power of the indicators, while the dynamic analysis allows us to verify whether the indicators satisfy additional important requirements, including timeliness and stability of the signal. 3.3.1 Forecasting performance of BIS and frontier gaps over a fixed forecast period In the static assessment, the predictive performance of the BIS and frontier gaps is tested over a fixed forecast window of 1–3 years (i.e., a signal is considered correct if it forecasts a crisis 1–3 years in advance). The use of a forecast interval, rather than a point-wise forecast horizon, permits to address the uncertainty of dating crisis correctly (Drehmann and Juselius, 2014), which is relevant in our analysis because we use two crisis databases adopting different definitions of systemic banking crisis (Table 2). The choice of the forecast window reflects two opposite considerations. On the one hand, forecast intervals need to be sufficiently long so that policy actions can be taken in time to be effective if systemic risk is detected. On the other hand, forecast windows cannot be too large because this would prevent the use of the whole sample, given that the time series of several countries are too short. Accordingly, the lower bound of the interval is set at 1 year, which is sufficiently distant from the crisis to allow policy makers to implement policies and macro-prudential actions to start impact on the policy objective (reduction in credit growth). The higher bound of the interval, instead, is set at 3 years, which allows the use of the full sample of countries. For this forecast period, we derive, for each country in the sample, the ROC curves corresponding to the two indicators and compute the AUC non-parametrically by trapezoidal integration.9 As mentioned in Section 3.2, we use the AUC to rank the forecasting performance of the BIS and frontier gaps in each country. Accordingly, we define indicator I to outperform indicator J in country h over the forecasting period T when the following criterion is satisfied: AUC(Ih,T)>0.5∩AUC(Ih,T)>AUC(Jh,T). (1) This criterion requires that the best performing indicator is informative and has the highest signalling quality. Figure 4 reports the AUC for the BIS and frontier gaps (vertical axis, left and right panels respectively) against the level of financial development (horizontal axis) for individual countries. In each panel, the observations marked with a diamond correspond to countries where the respective indicator outperforms the other based on criterion (1), while the observations marked with a star correspond to countries for which both indicators are uninformative (AUC below or equal to 0.5). Figure 4: View largeDownload slide Predictive Performance of BIS and Frontier Gaps (AUC Metric). The Figure Illustrates the Predictive Performance of the BIS and Frontier Gaps (Left and Right Charts, Respectively), as Measured by the AUC, Against the Level of Financial Development for Individual Countries. In Each Country, the Indicator Is Considered Informative if Its AUC Is Superior to 0.5. The Figure Shows that the Relative Performance of the BIS (frontier) Gap Is Superior to the Performance of the Frontier (BIS) Gap in Countries at an Intermediate (Low and High) Level of Financial Development. Source: IMF IFS, Finstats and Authors’ Calculations. Figure 4: View largeDownload slide Predictive Performance of BIS and Frontier Gaps (AUC Metric). The Figure Illustrates the Predictive Performance of the BIS and Frontier Gaps (Left and Right Charts, Respectively), as Measured by the AUC, Against the Level of Financial Development for Individual Countries. In Each Country, the Indicator Is Considered Informative if Its AUC Is Superior to 0.5. The Figure Shows that the Relative Performance of the BIS (frontier) Gap Is Superior to the Performance of the Frontier (BIS) Gap in Countries at an Intermediate (Low and High) Level of Financial Development. Source: IMF IFS, Finstats and Authors’ Calculations. Figure 5: View largeDownload slide Distributions of the Difference between Median AUC for Frontier and BIS Gaps at different levels of financial development. The Figure Shows that, Based on the Bootstrapped Distributions, the Probability that the Difference between the Median AUC of the BIS and Frontier Gaps Is Null Is Inferior to 10% in all Country Groups, Which Confirms that the Difference in the Performance of the Two Indicators Is Statistically Significant. Source: IMF IFS, Finstats and Authors’ Calculations. Figure 5: View largeDownload slide Distributions of the Difference between Median AUC for Frontier and BIS Gaps at different levels of financial development. The Figure Shows that, Based on the Bootstrapped Distributions, the Probability that the Difference between the Median AUC of the BIS and Frontier Gaps Is Null Is Inferior to 10% in all Country Groups, Which Confirms that the Difference in the Performance of the Two Indicators Is Statistically Significant. Source: IMF IFS, Finstats and Authors’ Calculations. This preliminary graphical analysis permits to identify three main findings: The BIS gap has an inverted U-shape performance across levels of financial development (polynomial fitting in Figure 4, left panel). The indicator shows an average weak forecasting performance in countries at low levels of financial development and a high performance in countries with intermediate and large financial sectors. In countries with an over-sized financial sector, the BIS gap performs well but is dominated by the frontier gap. The frontier gap has a U-shape performance (polynomial fitting in Figure 4, right panel). The indicator has an average strong performance in countries at a low level of financial development and with an over-sized financial sector but it has a relatively weak performance in countries with an intermediate or large financial sector. In a number of countries, both measures of excess credit are a poor predictor of financial crises (observations marked with a star in Figure 4). This implies that, in these countries, financial stress originated from sources different from fast credit growth. It is worth noticing that these observations are concentrated in countries with a low level of financial development. Based on these preliminary findings, we split the countries in our sample into three groups: (a) group 1, which includes countries with a credit-to-GDP ratio up to 60% at the time of the crisis (low level of financial development); (b) group 2, comprising countries with a credit-to-GDP ratio higher than 60% and up to 160% at the time of the crisis (intermediate and high level of financial development); and (iii) group 3, including countries with a credit-to-GDP ratio above 160% at the time of the crisis (over-sized financial sector). The cutoff levels of the credit to GDP ratio for each group are consistent with stylised facts (e.g. low and middle income countries have an average level of credit not superior to 60% of GDP, see Figure 1) and empirical literature (for over-sized financial sectors, see Cecchetti and Kharroubi, 2012; European Systemic Risk Board, 2014). In Table 3, we restrict the focus on the 55 observations for which at least one of the two indicators is informative (i.e., the AUC is superior to 0.5) and we synthesise the predictive performance of the two gaps in each of the three groups with the AUC recorded on median by the countries belonging to the corresponding subsample. In Group 1, the frontier gap outperforms the BIS gap in two thirds of the observations, recording a median area under the curve of 0.76 against a median AUC of 0.63 for the BIS gap. In Group 2, instead, the BIS gap performs better than the frontier gap in 86% of the cases and registers a median AUC of 0.83 against a median value of 0.60 for the frontier gap. Finally, in Group 3, the frontier gap has the best predictive performance, even if both the BIS and frontier gaps have high median AUCs (0.88 against 0.96). Table 3: Forecasting Performance of the BIS and Frontier Gaps over a Fixed Forecast Period across Levels of Financial Development Indicator Median(AUC)1 Standard deviation median (AUC) C.I.2 median(AUC) p-Value3 median(AUC) Best indicator (percent of observations) Sample size Low level of financial development (Group 1) BIS gap 0.63 0.10 [0.39,0.76] 0.16 35.1 37 Frontier Gap 0.76*** 0.06 [0.62,0.85] 0.00 64.9 Intermediate and high level of financial development (Group 2) BIS gap 0.83*** 0.04 [0.74,0.91] 0.00 85.7 14 Frontier Gap 0.60* 0.10 [0.33,0.78] 0.10 14.3 Oversized financial sector (Group 3) BIS gap 0.88** 0.17 [0.26,0.95] 0.05 0.00 4 Frontier Gap 0.96*** 0.02 [0.93,1.00] 0.00 100.0 Indicator Median(AUC)1 Standard deviation median (AUC) C.I.2 median(AUC) p-Value3 median(AUC) Best indicator (percent of observations) Sample size Low level of financial development (Group 1) BIS gap 0.63 0.10 [0.39,0.76] 0.16 35.1 37 Frontier Gap 0.76*** 0.06 [0.62,0.85] 0.00 64.9 Intermediate and high level of financial development (Group 2) BIS gap 0.83*** 0.04 [0.74,0.91] 0.00 85.7 14 Frontier Gap 0.60* 0.10 [0.33,0.78] 0.10 14.3 Oversized financial sector (Group 3) BIS gap 0.88** 0.17 [0.26,0.95] 0.05 0.00 4 Frontier Gap 0.96*** 0.02 [0.93,1.00] 0.00 100.0 The table compares the median AUC of the BIS and frontier gaps recorded in each of the three country group subsamples. Results confirm that the BIS (frontier) gap outperforms in Group 2 (Groups 1 and 3) both in terms of predictive power and statistical significance. 1 *, ** and *** denote statistical significance at the 10, 5 and 1% level, respectively. 2The 95% bootstrap confidence intervals are estimated by the cutoff values for the middle 95% of each bootstrap distribution. 3The bootstrap p-values are computed as the proportion of bootstrap samples where the simulated median AUCs are equal or inferior to 0.5. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5. Source: IMF IFS, Finstats and Authors’ Calculations. View Large Table 3: Forecasting Performance of the BIS and Frontier Gaps over a Fixed Forecast Period across Levels of Financial Development Indicator Median(AUC)1 Standard deviation median (AUC) C.I.2 median(AUC) p-Value3 median(AUC) Best indicator (percent of observations) Sample size Low level of financial development (Group 1) BIS gap 0.63 0.10 [0.39,0.76] 0.16 35.1 37 Frontier Gap 0.76*** 0.06 [0.62,0.85] 0.00 64.9 Intermediate and high level of financial development (Group 2) BIS gap 0.83*** 0.04 [0.74,0.91] 0.00 85.7 14 Frontier Gap 0.60* 0.10 [0.33,0.78] 0.10 14.3 Oversized financial sector (Group 3) BIS gap 0.88** 0.17 [0.26,0.95] 0.05 0.00 4 Frontier Gap 0.96*** 0.02 [0.93,1.00] 0.00 100.0 Indicator Median(AUC)1 Standard deviation median (AUC) C.I.2 median(AUC) p-Value3 median(AUC) Best indicator (percent of observations) Sample size Low level of financial development (Group 1) BIS gap 0.63 0.10 [0.39,0.76] 0.16 35.1 37 Frontier Gap 0.76*** 0.06 [0.62,0.85] 0.00 64.9 Intermediate and high level of financial development (Group 2) BIS gap 0.83*** 0.04 [0.74,0.91] 0.00 85.7 14 Frontier Gap 0.60* 0.10 [0.33,0.78] 0.10 14.3 Oversized financial sector (Group 3) BIS gap 0.88** 0.17 [0.26,0.95] 0.05 0.00 4 Frontier Gap 0.96*** 0.02 [0.93,1.00] 0.00 100.0 The table compares the median AUC of the BIS and frontier gaps recorded in each of the three country group subsamples. Results confirm that the BIS (frontier) gap outperforms in Group 2 (Groups 1 and 3) both in terms of predictive power and statistical significance. 1 *, ** and *** denote statistical significance at the 10, 5 and 1% level, respectively. 2The 95% bootstrap confidence intervals are estimated by the cutoff values for the middle 95% of each bootstrap distribution. 3The bootstrap p-values are computed as the proportion of bootstrap samples where the simulated median AUCs are equal or inferior to 0.5. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5. Source: IMF IFS, Finstats and Authors’ Calculations. View Large To test whether these results are statistically significant, we perform statistical inference on the basis of non-parametric estimation rather than asymptotic normal approximation due to the small sample size.10 Accordingly, we use bootstrap resampling to calculate the standard errors, point-wise confidence intervals (C.I.), and p-values of the median AUC of the BIS and frontier gaps in each country group. The bootstrap algorithm uses 1,000 independent replications, each reflecting the original sample partition among the three country groups. Bootstrap test statistics (C.I. and p-values) confirm that the best indicator in each country group has a statistically significant median AUC (Table 3),11 although results for the third group should be taken with caution as the original sample is too small to ensure enough variation in the bootstrap distributions. Up to this point, the results of our analysis have found that, for the fix forecast window under observation, the frontier gap has superior statistical forecasting power in Groups 1 and 3, while the BIS gap has stronger predictive performance in group 2. However, given that the AUCs’ confidence intervals of the two indicators overlap in each group,12 it is not granted that their performance is statistically different, implying that there might not be a statistically dominant EWI in each group. To address this remaining uncertainty, we complete the assessment in the static approach by testing whether the performance of the two gaps is statistically different in each group. For this purpose, we use the 1,000 bootstrap samples to derive the distribution of the difference between the median AUC of the frontier and BIS gaps (Figure 5) and we test the null hypothesis that this difference is equal to zero in each group. Table 4 reports the results of this analysis. Based on the bootstrap test statistics (C.I. and p-values), the null hypothesis is rejected for all groups (at the 10% significance level for Group 1). Again, results for Group 3 should be taken with caution due to the paucity of observations. Table 4: Difference between Bootstrap Median AUCs across Levels of Financial Development Indicator C.I.1 p-Value2 Difference in median AUC (Group 1) [−0.04,0.37] 0.06* Difference in median AUC (Group 2) [−0.46,−0.06] 0.01*** Difference in median AUC (Group 3) [−0.01,0.74] 0.02** Indicator C.I.1 p-Value2 Difference in median AUC (Group 1) [−0.04,0.37] 0.06* Difference in median AUC (Group 2) [−0.46,−0.06] 0.01*** Difference in median AUC (Group 3) [−0.01,0.74] 0.02** In this table, the null hypothesis that the difference between the median AUC of the frontier and BIS gaps is null is tested. Results confirm that the difference is significant in each country group at a minimum significance level of 10%. 1The 95% bootstrap confidence intervals are estimated by the cutoff values for the middle 95% of each bootstrap distribution. 2The bootstrap p-values for Groups 1 and 3 (Group 2) are computed as the proportion of bootstrap samples where the simulated median AUCs are equal or inferior (superior) to 0. *, ** and *** denote statistical significance at the 10, 5 and 1% level, respectively. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5 for both the frontier and BIS gaps. Source: IMF IFS, Finstats and Authors’ Calculations. View Large Table 4: Difference between Bootstrap Median AUCs across Levels of Financial Development Indicator C.I.1 p-Value2 Difference in median AUC (Group 1) [−0.04,0.37] 0.06* Difference in median AUC (Group 2) [−0.46,−0.06] 0.01*** Difference in median AUC (Group 3) [−0.01,0.74] 0.02** Indicator C.I.1 p-Value2 Difference in median AUC (Group 1) [−0.04,0.37] 0.06* Difference in median AUC (Group 2) [−0.46,−0.06] 0.01*** Difference in median AUC (Group 3) [−0.01,0.74] 0.02** In this table, the null hypothesis that the difference between the median AUC of the frontier and BIS gaps is null is tested. Results confirm that the difference is significant in each country group at a minimum significance level of 10%. 1The 95% bootstrap confidence intervals are estimated by the cutoff values for the middle 95% of each bootstrap distribution. 2The bootstrap p-values for Groups 1 and 3 (Group 2) are computed as the proportion of bootstrap samples where the simulated median AUCs are equal or inferior (superior) to 0. *, ** and *** denote statistical significance at the 10, 5 and 1% level, respectively. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5 for both the frontier and BIS gaps. Source: IMF IFS, Finstats and Authors’ Calculations. View Large For robustness, we conclude the static assessment by verifying whether the results described above (and synthesised in the first column of Table 5) are sensitive to: the cutoff level of the credit to GDP ratio for each group; the choice to assign countries to the three groups based on a specific year and the identification and timing of banking crises. Table 5: Median AUCs across Levels of Financial Development—Robustness Checks Indicator First crisis1 First crisis1 First crisis1 Last crisis2 First crisis1 First crisis1 LV-R3 LV-R3 LV-R3 LV-R3 R4 LV5 Cutoff = 606 Cutoff = 507 Cutoff = 708 Cutoff = 606 Cutoff = 606 Cutoff = 606 Group 1 BIS gap 0.63 0.66 0.63 0.59 0.73* 0.60* Frontier gap 0.76*** 0.78*** 0.76*** 0.78*** 0.83*** 0.74*** Sample size 37 32 39 31 20 30 Group 2 BIS gap 0.83*** 0.78*** 0.82*** 0.80*** 0.82*** 0.79*** Frontier gap 0.60* 0.61 0.60* 0.63* 0.56 0.68** Sample size 14 19 12 16 16 12 Group 3 BIS gap 0.88** 0.88** 0.88** 0.82* 0.88** 0.89 Frontier gap 0.96*** 0.96*** 0.96*** 0.90*** 0.96*** 0.98*** Sample size 4 4 4 8 4 7 Indicator First crisis1 First crisis1 First crisis1 Last crisis2 First crisis1 First crisis1 LV-R3 LV-R3 LV-R3 LV-R3 R4 LV5 Cutoff = 606 Cutoff = 507 Cutoff = 708 Cutoff = 606 Cutoff = 606 Cutoff = 606 Group 1 BIS gap 0.63 0.66 0.63 0.59 0.73* 0.60* Frontier gap 0.76*** 0.78*** 0.76*** 0.78*** 0.83*** 0.74*** Sample size 37 32 39 31 20 30 Group 2 BIS gap 0.83*** 0.78*** 0.82*** 0.80*** 0.82*** 0.79*** Frontier gap 0.60* 0.61 0.60* 0.63* 0.56 0.68** Sample size 14 19 12 16 16 12 Group 3 BIS gap 0.88** 0.88** 0.88** 0.82* 0.88** 0.89 Frontier gap 0.96*** 0.96*** 0.96*** 0.90*** 0.96*** 0.98*** Sample size 4 4 4 8 4 7 In this table, the estimates of the static assessment (first column) are compared with those obtained using different assumptions. Results confirm that the overall thrust of results remain robust. *, ** and *** denote statistical significance at the 10, 5 and 1% level, respectively. 1Countries are classified in the three groups based on their credit-to-GDP ratio at the time of the first crisis. 2Countries are classified in the three groups based on their credit-to-GDP ratio at the time of the last crisis. 3The identification and timing of banking crises are based on the database that merges Reinhart (2010) and Laeven and Valencia (2013). 4The identification and timing of banking crises are based on Reinhart (2010) database. 5The identification and timing of banking crises are based on Laeven and Valencia (2013) database. 6Countries are classified in Group 1 if their credit-to-GDP ratio is up to 60%. 7Countries are classified in Group 1 if their credit-to-GDP ratio is up to 50%. 8Countries are classified in Group 1 if their credit-to-GDP ratio is up to 70%. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5 for both the frontier and BIS gaps. Source: IMF IFS, Finstats and Authors’ Calculations. View Large Table 5: Median AUCs across Levels of Financial Development—Robustness Checks Indicator First crisis1 First crisis1 First crisis1 Last crisis2 First crisis1 First crisis1 LV-R3 LV-R3 LV-R3 LV-R3 R4 LV5 Cutoff = 606 Cutoff = 507 Cutoff = 708 Cutoff = 606 Cutoff = 606 Cutoff = 606 Group 1 BIS gap 0.63 0.66 0.63 0.59 0.73* 0.60* Frontier gap 0.76*** 0.78*** 0.76*** 0.78*** 0.83*** 0.74*** Sample size 37 32 39 31 20 30 Group 2 BIS gap 0.83*** 0.78*** 0.82*** 0.80*** 0.82*** 0.79*** Frontier gap 0.60* 0.61 0.60* 0.63* 0.56 0.68** Sample size 14 19 12 16 16 12 Group 3 BIS gap 0.88** 0.88** 0.88** 0.82* 0.88** 0.89 Frontier gap 0.96*** 0.96*** 0.96*** 0.90*** 0.96*** 0.98*** Sample size 4 4 4 8 4 7 Indicator First crisis1 First crisis1 First crisis1 Last crisis2 First crisis1 First crisis1 LV-R3 LV-R3 LV-R3 LV-R3 R4 LV5 Cutoff = 606 Cutoff = 507 Cutoff = 708 Cutoff = 606 Cutoff = 606 Cutoff = 606 Group 1 BIS gap 0.63 0.66 0.63 0.59 0.73* 0.60* Frontier gap 0.76*** 0.78*** 0.76*** 0.78*** 0.83*** 0.74*** Sample size 37 32 39 31 20 30 Group 2 BIS gap 0.83*** 0.78*** 0.82*** 0.80*** 0.82*** 0.79*** Frontier gap 0.60* 0.61 0.60* 0.63* 0.56 0.68** Sample size 14 19 12 16 16 12 Group 3 BIS gap 0.88** 0.88** 0.88** 0.82* 0.88** 0.89 Frontier gap 0.96*** 0.96*** 0.96*** 0.90*** 0.96*** 0.98*** Sample size 4 4 4 8 4 7 In this table, the estimates of the static assessment (first column) are compared with those obtained using different assumptions. Results confirm that the overall thrust of results remain robust. *, ** and *** denote statistical significance at the 10, 5 and 1% level, respectively. 1Countries are classified in the three groups based on their credit-to-GDP ratio at the time of the first crisis. 2Countries are classified in the three groups based on their credit-to-GDP ratio at the time of the last crisis. 3The identification and timing of banking crises are based on the database that merges Reinhart (2010) and Laeven and Valencia (2013). 4The identification and timing of banking crises are based on Reinhart (2010) database. 5The identification and timing of banking crises are based on Laeven and Valencia (2013) database. 6Countries are classified in Group 1 if their credit-to-GDP ratio is up to 60%. 7Countries are classified in Group 1 if their credit-to-GDP ratio is up to 50%. 8Countries are classified in Group 1 if their credit-to-GDP ratio is up to 70%. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5 for both the frontier and BIS gaps. Source: IMF IFS, Finstats and Authors’ Calculations. View Large Relative to the first point, Table 5 shows that the thrust of the results for Groups 1 and 2 would not change using alternative cutoff levels that are in the proximity of the proposed thresholds (Table 5 illustrates results with cutoffs 50 and 70 in the second and third column, respectively). This result reflects the fact that, while point estimates differ (Table 6), the frontier gap has the highest median performance at any level of the credit-to-GDP ratio up to 60%, while the BIS gap has the best median performance at any level of the credit-to-GDP ratio above 60% and up to 160%. Checking robustness for Group 3 is more difficult given the small sample size and the fact that there are no observations with a credit-to-GDP ratio in the interval]160, 200[%. Table 6: Median AUC across Levels of Financial Development Private credit-to-GDP ratio (%) 0–20 20–40 40–60 60–80 80–100 100–120 120–140 140–160 160–180 180–200 200–220 220–240 Indicator BIS Gap 0.56 0.58 0.72 0.76 0.92 0.78 0.85 0.87 0.86 0.26 Frontier Gap 0.74 0.88 0.87 0.74 0.71 0.40 0.76 0.58 0.98 0.99 Sample Size 11 15 11 4 3 2 2 3 0 0 3 1 Private credit-to-GDP ratio (%) 0–20 20–40 40–60 60–80 80–100 100–120 120–140 140–160 160–180 180–200 200–220 220–240 Indicator BIS Gap 0.56 0.58 0.72 0.76 0.92 0.78 0.85 0.87 0.86 0.26 Frontier Gap 0.74 0.88 0.87 0.74 0.71 0.40 0.76 0.58 0.98 0.99 Sample Size 11 15 11 4 3 2 2 3 0 0 3 1 This table shows the median AUC of the BIS and frontier gaps in countries with different levels of financial development. Results confirm that the frontier gap has the highest median performance at any level of the credit-to-GDP ratio up to 60%, while the BIS gap has the highest performance at any level above 60 and up to 160%. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5 for both the frontier and BIS gaps. Source: IMF IFS, Finstats and Authors’ Calculations. View Large Table 6: Median AUC across Levels of Financial Development Private credit-to-GDP ratio (%) 0–20 20–40 40–60 60–80 80–100 100–120 120–140 140–160 160–180 180–200 200–220 220–240 Indicator BIS Gap 0.56 0.58 0.72 0.76 0.92 0.78 0.85 0.87 0.86 0.26 Frontier Gap 0.74 0.88 0.87 0.74 0.71 0.40 0.76 0.58 0.98 0.99 Sample Size 11 15 11 4 3 2 2 3 0 0 3 1 Private credit-to-GDP ratio (%) 0–20 20–40 40–60 60–80 80–100 100–120 120–140 140–160 160–180 180–200 200–220 220–240 Indicator BIS Gap 0.56 0.58 0.72 0.76 0.92 0.78 0.85 0.87 0.86 0.26 Frontier Gap 0.74 0.88 0.87 0.74 0.71 0.40 0.76 0.58 0.98 0.99 Sample Size 11 15 11 4 3 2 2 3 0 0 3 1 This table shows the median AUC of the BIS and frontier gaps in countries with different levels of financial development. Results confirm that the frontier gap has the highest median performance at any level of the credit-to-GDP ratio up to 60%, while the BIS gap has the highest performance at any level above 60 and up to 160%. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5 for both the frontier and BIS gaps. Source: IMF IFS, Finstats and Authors’ Calculations. View Large Results are also insensitive to the choice to assign countries to the three groups based on a specific year. For instance, Table 5 shows that classifying countries in the three groups based on the level of the credit-to-GDP ratio at the time of the most recent crisis (fourth column) would produce results very similar to those obtained classifying countries based on the value of the ratio at the time of the first crisis (first column). In this case, the robustness of results reflects the fact that, while the credit-to-GDP ratio of individual countries varies over time, only in few cases this variation is large enough to determine a change in the classification of countries from the first crisis to the most recent one. Specifically, in our sample only five countries move from Group 1 to Group 2, three countries move from Group 2 to Group 3 and one country from Group 1 to Group 3. These changes have limited impact on the median performance of the BIS and frontier gaps. Finally, the issue pertaining the uncertainty associated with the identification and timing of banking crises is already partially addressed by using a forecast interval, rather than a point-wise forecast horizon, as discussed at the beginning of this section. As additional robustness check, Table 5 reports the results obtained by using the crisis databases from Reinhart (2010) and from Laeven and Valencia (2013) separately (fifth and sixth column, respectively), rather than merging them in a unified database (first column). Also in this case, the overall thrust of results remains robust: both in terms of predictive power and statistical significance, the frontier gap is the dominant EWI for Group 1 and 3, while the BIS gap outperforms in Group 2. Overall, the outcome of the static assessment provides strong evidence that, for the fix forecast period under consideration, the frontier gap is a statistically superior EWI of financial crises in countries at an early stage of financial development, while the BIS gap outperforms in countries with a medium and high level of financial development. In countries with an over-sized financial sector, the frontier gap seems to be the dominant EWI, even if the size of the sample under observation is too small to perform proper statistical inference. 3.3.2 Dynamic forecasting performance of the BIS and frontier gaps The results in Section 3.3.1 are only indicative of the forecasting performance of the BIS and frontier gaps over a fix forecast window of 3 years but do not provide information on the performance at specific (point-wise) forecast horizons, which is relevant to test the forecasting power over time and assess the timing of the signal and its persistence across horizons. The timing of a signal is relevant because data are reported with lags, particularly in developing countries, and policy makers tend to adjust policies gradually to changes in macroeconomic conditions. In addition, changes in macro-prudential tools may take time to impact on their policy objective (Drehmann and Juselius, 2014; CGFS, 2012). It follows that ideal EWIs should start issuing signals well before a crisis. Also the persistence (or stability) of the signal is an important requirement because EWIs that issue stable signals reduce uncertainty regarding risk buildup, thus allowing for more decisive policy actions. Based on these considerations, in this section we assess the dynamic predictive performance of the BIS and frontier gaps in individual countries based on three criteria. To test the predictive power across time, we use again criterion (1), although applied to individual (point-wise) forecast horizons rather than to a forecast interval. The second criterion pertains to the timeliness of the signal and requires that the indicator starts issuing informative signals at least 3 years in advance: AUC(Ih,τ)>0.5withτ≥3. (2) Finally, the third criterion refers to the stability of the signal and requires that, after issuing the first informative signal, the indicator stays informative over the remaining forecast horizon. For instance, assuming that the first informative signal is issued 3 years before the crisis, the criterion requires that: AUC(Ih,3−j)>0.5∀j∈{1,2}. (3) To test criteria (1) to (3) on our sample, we derive the AUC for the BIS and frontier gaps for five consecutive point-wise horizons for individual countries.13 To have a consistent sample across forecast horizons, we drop six countries because their series of the credit-to-GDP ratio are not sufficiently long to assess the predictive performance of the gaps five years before a crisis. This reduces the sample size from 79 to 73.14 As in the static assessment, Figure 6 and Table 7 restrict the focus on the observations for which at least one of the two indicators is informative (i.e. the AUC is superior to 0.5) and synthesise the results for all forecast horizons and country groups. Figure 6: View largeDownload slide Signalling Quality, as Measured by the AUC, of BIS (Left) and Frontier Gaps (Right) at Different Levels of Financial Development and Different Forecast Periods. The Figure Shows that the Best Indicator in Each Country Group (i.e., the BIS Gap for Group 2 and the Frontier Gap for Groups 1 and 3) Has an Informative and Significant AUC across Projection Horizons. This Confirms that, in Each Group, the Best Indicator Issues Timely and Persistent Signals. Source: IMF IFS, Finstats and Authors’ Calculations. Figure 6: View largeDownload slide Signalling Quality, as Measured by the AUC, of BIS (Left) and Frontier Gaps (Right) at Different Levels of Financial Development and Different Forecast Periods. The Figure Shows that the Best Indicator in Each Country Group (i.e., the BIS Gap for Group 2 and the Frontier Gap for Groups 1 and 3) Has an Informative and Significant AUC across Projection Horizons. This Confirms that, in Each Group, the Best Indicator Issues Timely and Persistent Signals. Source: IMF IFS, Finstats and Authors’ Calculations. Table 7: Dynamic Predictive Performance of BIS and Frontier Gaps across Levels of Financial Development t−1 t−2 t−3 t−4 t−5 Group 1 Bis gap Median (AUC) 0.62 0.64 0.67 0.65 0.66 C.I. high1 0.74 0.77 0.79 0.78 0.76 C.I. low1 0.37 0.38 0.37 0.39 0.43 p-Value2 0.04 0.05 0.08 0.11 0.11 Frontier gap Median (AUC) 0.70*** 0.76*** 0.82*** 0.79*** 0.82*** C.I. high1 0.83 0.84 0.93 0.88 0.88 C.I. low1 0.56 0.54 0.71 0.63 0.63 p-Value2 0.01 0.01 0.00 0.00 0.01 Group 2 Bis gap Median (AUC) 0.85*** 0.82*** 0.81*** 0.75*** 0.72*** C.I. high1 0.95 0.92 0.90 0.84 0.80 C.I. low1 0.69 0.73 0.71 0.65 0.63 p-Value2 0.00 0.00 0.00 0.00 0.00 Frontier gap Median (AUC) 0.61 0.64 0.60 0.52 0.54 C.I. high1 0.77 0.79 0.82 0.73 0.76 C.I. low1 0.40 0.42 0.10 0.34 0.30 p-Value2 0.13 0.11 0.25 0.37 0.11 Group 3 Bis gap Median (AUC) 0.75 0.82* 0.88** 0.89** 0.91** C.I. high1 1.00 0.99 0.95 0.94 0.97 C.I. low1 0.06 0.15 0.26 0.35 0.44 p-Value2 0.20 0.06 0.05 0.04 0.05 Frontier gap Median (AUC) 0.92*** 0.94*** 0.96*** 0.98*** 1.00*** C.I. high1 1.00 1.00 1.00 1.00 1.00 C.I. low1 0.81 0.88 0.90 0.90 0.91 p-Value2 0.00 0.00 0.00 0.00 0.00 t−1 t−2 t−3 t−4 t−5 Group 1 Bis gap Median (AUC) 0.62 0.64 0.67 0.65 0.66 C.I. high1 0.74 0.77 0.79 0.78 0.76 C.I. low1 0.37 0.38 0.37 0.39 0.43 p-Value2 0.04 0.05 0.08 0.11 0.11 Frontier gap Median (AUC) 0.70*** 0.76*** 0.82*** 0.79*** 0.82*** C.I. high1 0.83 0.84 0.93 0.88 0.88 C.I. low1 0.56 0.54 0.71 0.63 0.63 p-Value2 0.01 0.01 0.00 0.00 0.01 Group 2 Bis gap Median (AUC) 0.85*** 0.82*** 0.81*** 0.75*** 0.72*** C.I. high1 0.95 0.92 0.90 0.84 0.80 C.I. low1 0.69 0.73 0.71 0.65 0.63 p-Value2 0.00 0.00 0.00 0.00 0.00 Frontier gap Median (AUC) 0.61 0.64 0.60 0.52 0.54 C.I. high1 0.77 0.79 0.82 0.73 0.76 C.I. low1 0.40 0.42 0.10 0.34 0.30 p-Value2 0.13 0.11 0.25 0.37 0.11 Group 3 Bis gap Median (AUC) 0.75 0.82* 0.88** 0.89** 0.91** C.I. high1 1.00 0.99 0.95 0.94 0.97 C.I. low1 0.06 0.15 0.26 0.35 0.44 p-Value2 0.20 0.06 0.05 0.04 0.05 Frontier gap Median (AUC) 0.92*** 0.94*** 0.96*** 0.98*** 1.00*** C.I. high1 1.00 1.00 1.00 1.00 1.00 C.I. low1 0.81 0.88 0.90 0.90 0.91 p-Value2 0.00 0.00 0.00 0.00 0.00 This table shows the median AUC of the BIS and frontier gaps at different levels of financial development across five forecast horizons. Result confirms that the BIS (frontier) gap issues timely and persistent signals in Group 2 (Groups 1 and 3). 1The 95% bootstrap confidence intervals are estimated by the cutoff values for the middle 95% of each bootstrap distribution. 2The bootstrap p-Values are computed as the proportion of bootstrap samples where the simulated median AUCs are equal or inferior to 0.5. *, ** and *** denote statistical significance at the 10, 5 and 1% level, respectively. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5 for both the frontier and BIS gaps. Source: IMF IFS, Finstats and Authors’ Calculations. View Large Table 7: Dynamic Predictive Performance of BIS and Frontier Gaps across Levels of Financial Development t−1 t−2 t−3 t−4 t−5 Group 1 Bis gap Median (AUC) 0.62 0.64 0.67 0.65 0.66 C.I. high1 0.74 0.77 0.79 0.78 0.76 C.I. low1 0.37 0.38 0.37 0.39 0.43 p-Value2 0.04 0.05 0.08 0.11 0.11 Frontier gap Median (AUC) 0.70*** 0.76*** 0.82*** 0.79*** 0.82*** C.I. high1 0.83 0.84 0.93 0.88 0.88 C.I. low1 0.56 0.54 0.71 0.63 0.63 p-Value2 0.01 0.01 0.00 0.00 0.01 Group 2 Bis gap Median (AUC) 0.85*** 0.82*** 0.81*** 0.75*** 0.72*** C.I. high1 0.95 0.92 0.90 0.84 0.80 C.I. low1 0.69 0.73 0.71 0.65 0.63 p-Value2 0.00 0.00 0.00 0.00 0.00 Frontier gap Median (AUC) 0.61 0.64 0.60 0.52 0.54 C.I. high1 0.77 0.79 0.82 0.73 0.76 C.I. low1 0.40 0.42 0.10 0.34 0.30 p-Value2 0.13 0.11 0.25 0.37 0.11 Group 3 Bis gap Median (AUC) 0.75 0.82* 0.88** 0.89** 0.91** C.I. high1 1.00 0.99 0.95 0.94 0.97 C.I. low1 0.06 0.15 0.26 0.35 0.44 p-Value2 0.20 0.06 0.05 0.04 0.05 Frontier gap Median (AUC) 0.92*** 0.94*** 0.96*** 0.98*** 1.00*** C.I. high1 1.00 1.00 1.00 1.00 1.00 C.I. low1 0.81 0.88 0.90 0.90 0.91 p-Value2 0.00 0.00 0.00 0.00 0.00 t−1 t−2 t−3 t−4 t−5 Group 1 Bis gap Median (AUC) 0.62 0.64 0.67 0.65 0.66 C.I. high1 0.74 0.77 0.79 0.78 0.76 C.I. low1 0.37 0.38 0.37 0.39 0.43 p-Value2 0.04 0.05 0.08 0.11 0.11 Frontier gap Median (AUC) 0.70*** 0.76*** 0.82*** 0.79*** 0.82*** C.I. high1 0.83 0.84 0.93 0.88 0.88 C.I. low1 0.56 0.54 0.71 0.63 0.63 p-Value2 0.01 0.01 0.00 0.00 0.01 Group 2 Bis gap Median (AUC) 0.85*** 0.82*** 0.81*** 0.75*** 0.72*** C.I. high1 0.95 0.92 0.90 0.84 0.80 C.I. low1 0.69 0.73 0.71 0.65 0.63 p-Value2 0.00 0.00 0.00 0.00 0.00 Frontier gap Median (AUC) 0.61 0.64 0.60 0.52 0.54 C.I. high1 0.77 0.79 0.82 0.73 0.76 C.I. low1 0.40 0.42 0.10 0.34 0.30 p-Value2 0.13 0.11 0.25 0.37 0.11 Group 3 Bis gap Median (AUC) 0.75 0.82* 0.88** 0.89** 0.91** C.I. high1 1.00 0.99 0.95 0.94 0.97 C.I. low1 0.06 0.15 0.26 0.35 0.44 p-Value2 0.20 0.06 0.05 0.04 0.05 Frontier gap Median (AUC) 0.92*** 0.94*** 0.96*** 0.98*** 1.00*** C.I. high1 1.00 1.00 1.00 1.00 1.00 C.I. low1 0.81 0.88 0.90 0.90 0.91 p-Value2 0.00 0.00 0.00 0.00 0.00 This table shows the median AUC of the BIS and frontier gaps at different levels of financial development across five forecast horizons. Result confirms that the BIS (frontier) gap issues timely and persistent signals in Group 2 (Groups 1 and 3). 1The 95% bootstrap confidence intervals are estimated by the cutoff values for the middle 95% of each bootstrap distribution. 2The bootstrap p-Values are computed as the proportion of bootstrap samples where the simulated median AUCs are equal or inferior to 0.5. *, ** and *** denote statistical significance at the 10, 5 and 1% level, respectively. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5 for both the frontier and BIS gaps. Source: IMF IFS, Finstats and Authors’ Calculations. View Large These results permit to derive three main findings pertaining the dynamic performance of the BIS and frontier gaps: In terms of predictive power, as defined by criterion (1), the BIS gap dominates the frontier gap at all horizons in countries with an intermediate and high level of financial development (Group 2), while the frontier gap outperforms in countries at an early stage of financial development (Group 1) and with an over-sized financial sector (Group 3). In terms of timing, the BIS and frontier gaps satisfy criterion (2) in all groups in the sample under observation. However, considering only results that are statistically significant, the BIS gap satisfies criterion (2) only in Groups 2 and 3, while the frontier gap in Groups 1 and 3. Pertaining to stability, the BIS and frontier gaps satisfy criterion (3) in all groups in the sample under observation. However, if we limit the analysis to statistically significant results, the BIS gap issues stable signals only in Group 2 and the frontier gap in Groups 1 and 3. In summary, the results in this section show that BIS (frontier) gap has the highest dynamic forecasting performance in Group 2 (Group 1) based on all criteria: predictive power, timeliness, and stability. Results for Group 3 suggest that the frontier gap has the best dynamic performance but the size of the sample under observation is too small to perform proper statistical inference and draw firm conclusions. 3.3.3 Overall assessment The evidence in Sections 3.3.1 and 3.3.2 provides strong support to the idea that the assessment of systemic risk buildup associated with excess credit should be performed with different EWIs at different levels of financial development. In countries at an early stage of financial development, the frontier gap outperforms the BIS gap in both the static and the dynamic assessments. In this country group, the frontier gap starts issuing informative signals 5 years ahead of a crisis and records a consistently high predictive power across forecasting horizons (with the AUC fluctuating between 0.70 and 0.82 over the 5 years). In addition, results are strongly significant for all forecasting horizons. The predictive power of the BIS gap is remarkably lower and median AUC estimates are not significant. This result is likely to reflect higher capacity of the frontier gap to discriminate financial deepening from systemic risk buildup compared to the BIS gap, which is relevant in developing countries. A healthy process of financial deepening, indeed, is associated with a situation where the frontier gap is negative and the acceleration of credit reflects a process of catch up with respect to the financial possibility frontier. This process would be incorrectly signalled as systemic risk buildup by the BIS gap but not by the frontier gap. In countries at an intermediate and high level of financial development, the results are inverted: the BIS gap has the highest static and dynamic predictive performances. These results are in line with previous findings in the literature according which the BIS gap outperforms other indicators as predictor of banking crises in advanced economies and emerging markets (Drehmann and Juselius, 2014; Drehmann and Tsatsaronis, 2014; Detken et al., 2014). This might reflect the fact that in countries with a developed financial sector the risks of additional financial deepening exceed the benefits. Finally, in countries with over-sized financial sectors (Group 3), the frontier gap seems to be the most powerful predictor but results should be taken with caution as they are based on a very small sample. A few caveats are worth mentioning when assessing the results for Group 1. First, the relation between credit and GDP should be expected to be less stable in developing countries than in other economies due to large output volatility. Drops in GDP, which are not mirrored by a fall in credit to the private sector of the same size, would result in an increase of both the BIS and frontier gaps that might, therefore, erroneously signal overheating.15 In addition, data limitations are extensive in developing countries. Quarterly series of credit and GDP, commonly used to test the performance of the BIS gap, are available only for a limited number of developing countries. Also, in those countries the series of credit and GDP may be marred by data gaps and structural breaks (due to civil wars, changes in the compilation of statistics and large swings in exchange rates) that can significantly impact the measurement of the credit-to-GDP ratio (Drehmann and Tsatsaronis, 2014). It must also be noted that the benchmark estimates used in this paper, and derived from FinStats, are based on a limited number of structural variables, while a vast array of other relevant factors, that might be critical to the development of the financial system, are excluded. 4. Conclusion We have provided evidence that systemic risk assessment based on measures of excess credit should be performed with different indicators at different stages of financial development. Measures that compare the level of credit with its historical trend are good proxies for systemic risk buildup in countries at an advanced stage of financial development. Instead, indicators that contrast actual credit to a benchmark level based on a country’s structural characteristics work better for countries with an underdeveloped financial sector. This fills a void in the EWS literature, as the existing analytical work on this subject has mainly tested the predictive performance of leading indicators of financial crises in advanced economies, where crises are expected to present common patterns and similarities. The findings in this paper also shed some light on the relationship between credit, financial stability, and deepening. In particular, In countries at an early stage of financial development, credit expansion is mainly positively associated with economic growth (financial deepening) (Rajan and Zingales, 1998; Levine et al., 2000; Levine, 2005). Systemic risk may, however, arise when credit expansion exceeds the absorption capacity of the economy, as determined by its structural characteristics. In countries with a developed financial system, on the other hand, rapid credit growth is frequently associated with excessive risk taking, weakening of lending standards, and the likelihood of a systemic banking crisis. In these countries, an acceleration of credit growth above its secular trend is likely to reflect systemic risk buildup rather than additional financial deepening. This explains the higher predictive performance of the BIS gap. This result is also in line with the empirical finding showing that booms that start at a higher level of financial depth are more likely to turn into a financial crisis (Dell’Ariccia et al., 2012; Meng and Gonzalez, 2017). While results for countries with an over-sized financial sector are not conclusive, the size of the financial sector seems to be a source of systemic risk per se, independent from acceleration in credit growth. This may reflect the fact that large financial systems are likely to be accompanied by riskier business models, larger interconnectedness, and higher vulnerability to global shocks. In addition, larger financial systems tend to have more systemically important banks that, on average, create more individual and systemic risk than smaller banks (Laeven et al., 2014). We also found that for some developing countries systemic risk buildup stems from sources other than excess credit. This paper showed, indeed, that in these countries banking crises may occur even if credit is not excessive by any measure. This may reflect the fact that banking sectors in developing countries are frequently more fragile than in advanced economies. Thus, there remains a need for the development of additional leading indicators and early warning models that are specific for those countries and focus on sources of vulnerability that are different from excess credit. The identification of these risk factors is important also in the case of crises preceded by excessive credit expansion, as they could heighten the severity of financial disruption connected to the bursting of the credit bubble. We leave a systematic examination of this issue for future research. In future work, it might also be interesting to explore whether, in developing countries, the assessment of systemic risk buildup based on the frontier gap should be complemented with signals from the BIS gap. In fact, while in our analysis the frontier gap showed a higher median predictive performance in countries at an early stage of financial development, the BIS gap outperformed the frontier gap in several cases (35% in the static assessment, Table 3). This may reflect the fact that despite financial development being below capacity (negative frontier gap), the acceleration of credit was too fast (strongly positive BIS gap) and led to the buildup of vulnerabilities, particularly in the absence of strong supervision. Finally, this paper has not addressed the issue of determining the threshold levels of the BIS and frontier gaps that may pose significant risks to financial stability in developing countries. Previous empirical research on this subject is based on the evidence of advanced economies and large emerging markets, or on large samples including countries at different levels of financial development. Two factors, pulling in opposite directions, should be taken into consideration when extending the analysis to developing countries. On the one side, financial deepening could justify higher BIS gap thresholds than in advanced economies. On the other, financial systems and supervisory and regulatory frameworks are frequently more weaker in developing countries. In these circumstances, even small deviations of the credit-to-GDP ratio from its long-term trend, or from the frontier, can be sufficient to spur a banking crisis. This issue also remains open for future research. References Al-Hussainy E. , Coppola A. , Feyen E. , Ize A. , Kibuuka K. , Ren H. ( 2011 ) FinStats2011: A Ready-to-Use Tool to Benchmark Financial Sectors Across Countries and Over Time . Washington : World Bank . Arcand J. , Berkes E. , Panizza U. ( 2015 ) ‘ Too Much Finance? ’, Journal of Economic Growth , 20 ( 2 ): 105 – 48 . Google Scholar CrossRef Search ADS Arena M. , Bouza S. , Dabla-Norris E. , Gerling K. , Njie L. ( 2015 ) Credit Booms andMacroeconomic Dynamics: Stylized Facts and Lessons for Low-Income Countries, IMF Working Papers, 15/11. Barajas A. , Beck T. , Dabla-Norris E. , Yousefi S. R. ( 2013 ) Too Cold, Too Hot, Or Just Right? Assessing Financial Sector Development Across theGlobe, IMF Working Papers, 13/81. Barajas A. , Dell’Ariccia G. , Levchenko A. ( 2007 ) Credit Booms: The Good, the Bad, and the Ugly. Unpublished manuscript, International Monetary Fund. Basel Committee on Banking Supervision ( 2010 ) Guidance for National Authorities Operating the Countercyclical Capital Buffer . Basel : Basel Committee on Banking Supervision . Beck T. , Feyen E. , Ize A. , Moizeszowicz F. ( 2008 ). Benchmarking Financial Development, World Bank Policy Research Working Paper, No. 4638. Bernanke B. , Gertler M. , Gilchrist S. ( 1999 ) ‘The financial accelerator in a quantitative business cycle framework’, in Taylor J. , Woodford M. (eds) , Handbook of Macroeconomics , Vol. 1C . Amsterdam : Elsevier , pp. 1531 – 1614 . Borio C. ( 2012 ) The financial cycle and macroeconomics: What have we learnt?, BIS Working Papers, No 395. Borio C. , Drehmann M. ( 2009 ) ‘ Assessing the Risk of Banking Crises—Revisited ’, BIS Quarterly Review , March : 29 – 46 . Borio C. , Lowe P. ( 2002 ) ‘ Assessing the Risk of Banking Crises ’, BIS Quarterly Review , December : 43 – 54 . Cecchetti S. , Kharroubi E. ( 2012 ) Reassessing the impact of finance on growth, BIS Working Papers, No. 381. Committee on the Global Financial System (CGFS) ( 2012 ) Operationalising the selection and application of macroprudential instruments, CGFS Publications, No 48. Cottarelli C. , Dell’Ariccia G. , Vladkova-Hollar I. ( 2005 ) ‘ Early Birds, Late Risers, and Sleeping Beauties: Bank Credit Growth to the Private Sector in Central and Eastern Europe and in the Balkans ’, Journal of Banking & Finance , 29 ( 1 ): 83 – 104 . Google Scholar CrossRef Search ADS Crowe C. , Dell’Ariccia G. , Igan D. , Rabanal P. ( 2013 ) ‘ How to Deal with Real Estate Booms: Lessons from Country Experiences ’, Journal of Financial Stability , 9 : 300 – 19 . Google Scholar CrossRef Search ADS Dembiermont C. , Drehmann M. , Muksakunratana S. ( 2013 ) ‘ How Much Does the Private Sector Really Borrow—A New Database for Total Credit to the Private Non-Financial Sector ’, BIS Quarterly Review , March : 65 – 81 . Dell’Ariccia G. , Igan D. , Laeven L. , Tong H. ( 2012 ). Policies for Macrofinancial Stability: Options to Deal with Credit Booms, IMF Staff Discussion Note, 12/06. Demirgüç-Kunt A. , Detragiache E. , Gupta P. ( 2006 ) ‘ Inside the Crisis: An Empirical Analysis of Banking Systems in Distress ’, Journal of International Money and Finance , 25 ( No 5 ): 702 – 18 . Google Scholar CrossRef Search ADS Detken C. , Weeken O. , Alessi L. , Bonfim D. , Boucinha M. M. , Castro C. , Frontczak S. , Giordana G. , Giese J. , Jahn N. , Kakes J. , Klaus B. , Lang J. H. , Puzanova N. , Welz P. ( 2014 ). Operationalising the countercyclical capital buffer: indicator selection, threshold identification and calibration options, European Systemic Risk Board Occasional Paper Series, No 5, June 2014. Drehmann M. , Borio C. , Gambacorta L. , Jimenez G. , Trucharte C. ( 2010 ). Countercyclical capital buffers: Exploring options, BIS Working Paper Series, No 317. Drehmann M. , Borio C. , Tsatsaronis K. ( 2011 ). Anchoring countercyclical capital buffers: the role of credit aggregates, BIS Working Paper Series, No 355. Drehmann M. , Juselius M. ( 2014 ) ‘ Evaluating Early Warning Indicators of Banking Crises: Satisfying Policy Requirements ’, International Journal of Forecasting , Elsevier, 30 ( 3 ): 759 – 80 . Google Scholar CrossRef Search ADS Drehmann M. , Tsatsaronis K. ( 2014 ) ‘ The Credit-to-GDP Gap and Countercyclical Capital Buffers: Question and Answers ’, BIS Quarterly Review , March : 55 – 73 . Edge R. , Meisenzahl R. ( 2011 ) ‘ The Unreliability of Credit-to-GDP Ratio Gaps in Real-Time: Implications for Countercyclical Capital Buffers ’, International Journal of Central Banking , December : 261 – 98 . European Systemic Risk Board . 2014 . IsEurope Overbanked?, Reports of the Advisory Scientific Committee, No 4, June. Feyen E. , Kibuuka K. , Sourrouille D. ( 2015 ) FinStats 2015: A Ready-to-Use Tool to Benchmark Financial Sectors across Countries and over Time User Guide and Benchmarking Methodology . Washington : World Bank . April. Gambacorta L. ( 2009 ) ‘ Monetary Policy and the Risk-Taking Channel ’, BIS Qarterly Review , December : 43 – 53 . Geršl A. , Seidler J. ( 2012 ) ‘ Excessive Credit Growth and Countercyclical Capital Buffers in Basel III, an Empirical Evidence from Central and East European Countries ’, ACTA VSFS Economic Studies and Analyses , University of Finance and Administration, 6 ( 2 ): 91 – 107 . Gilchrist S. , Ortiz A. , Zakrajsek E. ( 2009 ) Credit Risk and the Macroeconomy: Evidence from an Estimated DSGE. Unpublished manuscript. Boston University. Gourinchas P. , Obstfeld M. ( 2012 ) ‘ Stories of the Twentieth Century for the Twenty First ’, American Economic Journal: Macroeconomics , Volume 4 ( No.1 ): 226 – 65 . Google Scholar CrossRef Search ADS Gourinchas P. , Valdes R. , Landerretche O. ( 2001 ) ‘ Lending Booms: Latin America and the World ’, Economia , 1 ( 2 ): 47 – 99 . Hodrick R. , Prescott E. C. ( 1997 ) ‘ Postwar U.S. Business Cycles: An Empirical Investigation ’, Journal of Money, Credit, and Banking , 29 ( 1 ): 1 – 16 . Google Scholar CrossRef Search ADS International Monetary Fund ( 2011 ) Towards Operationalizing Macroprudential policy: When to Act?, Global Financial Stability Report, September, Washington: IMF. International Monetary Fund ( 2014 ) Staff Guidance Note on Macroprudential Policy—Considerations for Low Income Countries . Washington : IMF . Janes H. , Longton G. , Pepe M. ( 2009 ) ‘ Accommodating Covariates in ROC analysis ’, Stata Journal , 9 ( 1 ): 17 – 39 . Google Scholar PubMed Jorda O. , Schularick M. , Taylor A. ( 2011 ) ‘ Financial Crises, Credit Booms, and External Imbalances: 140 Years of Lessons ’, IMF Economic Review , 59 : 340 – 78 . Google Scholar CrossRef Search ADS Kaminsky G. , Reinhart C. ( 1999 ) ‘ The Twin Crises: The Causes of Banking and Balance-of-Payments Problems ’, American Economic Review , 89 : 473 – 500 . Google Scholar CrossRef Search ADS Kiyotaki N. , Moore J. ( 1997 ) ‘ Credit Cycles ’, Journal of Political Economy , 105 ( 2 ): 211 – 48 . Google Scholar CrossRef Search ADS Laeven L. , Ratnovski L. , Tong H. 2014 . Bank Size and Systemic Risk. IMF Staff Discussion Notes, 14/04. Laeven L. , Valencia F. ( 2013 ) ‘ Systemic Banking Crises Database ’, IMF Economic Review , 61 : 225 – 70 . Google Scholar CrossRef Search ADS Levine R. ( 2005 ) ‘Finance and growth: theory and evidence’, in Aghion P. , Durlauf S. N. (eds) , Handbook of Economic Growth . Amsterdam : Elsevier , pp. 865 – 934 . Levine R. E. , Loayza N. V. , Beck T. ( 2000 ) ‘ Financial Intermediation and Growth: Causality and Causes ’, Journal of Monetary Economics , 16 ( 1 ): 31 – 77 . Google Scholar CrossRef Search ADS Mendoza E. G. , Terrones M. E. 2008 . An Anatomy of Credit Booms: Evidence from Macro Aggregates and Micro Data, NBER Working Papers, No 14049. Meng C. , Gonzalez R. L. ( 2017 ) ‘ Credit Booms in Developing Countries: Are they Different from Those in Advanced Economies and Emerging Market Countries? ’, Open Economies Review , No 28 : 547 – 79 . Google Scholar CrossRef Search ADS Rajan R. G. , Zingales L. ( 1998 ) ‘ Financial Dependence and Growth ’, The American Economic Review , 88 ( 3 ): 559 – 86 . Ravn M. O. , Uhlig H. ( 2002 ) ‘ On Adjusting the Hodrick-Prescott Filter for the Frequency of Observations ’, Review of Economics and Statistics , 84 ( 2 ): 371 – 5 . Google Scholar CrossRef Search ADS Reinhart C. 2010 . This time is different chartbook: country histories on debt, default, and financial crises, NBER Working Papers, No 15815. Reinhart C. , Rogoff K. ( 2009 ) This Time is Different: Eight Centuries of Financial Folly . Princeton : Princeton University Press . Sanusi L. S. 2010 . Global Financial Meltdown and the Reforms in the Nigerian Banking Sector, Speech at the Public Lecture delivered at the Convocation Square, Abubakar Tafawa Balewa University , Bauchi . Schularick M. , Taylor A. M. ( 2012 ) ‘ Credit Booms Gone Bust: Monetary Policy, Leverage Cycles, and Financial Crises, 1870–2008 ’, American Economic Review , 102 ( 2 ): 1029 – 61 . Google Scholar CrossRef Search ADS World Bank . 2010 . Comments on the consultative document countercyclical capital buffer proposal. Footnotes 1 The consolidation was forced by the introduction of stricter capital requirements. The Central Bank of Nigeria announced on 2 July 2004 that banks would be required to achieve a minimum capital level of Naira 25 billion up from Naira 2 billion, by 31 December 2005. 2 Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Greece, Iceland, India, Indonesia, Ireland, Italy, Japan, Korea, Mexico, Netherlands, Portugal, Singapore, South Africa, Spain, Sweden, Switzerland, Thailand, UK and the USA. 3 BIS data cover total credit, including also cross-border finance to the private non-financial sector (Dembiermont et al., 2013). 4 The indicator covers only domestic credit, thus omitting credit booms driven by cross-border finance. According to Drehmann and Tsatsaronis (2014), credit gaps based on total credit outperform those based on narrower measures of credit. The discrepancy is, however, expected to be larger in countries with market-based financial systems and open capital accounts. For those financial sectors, we mainly use the BIS series of total credit. In developing economies, credit to the private non-financial sector is mainly extended by domestic banks and total domestic credit may thus be considered a reliable proxy for total credit in those countries. 5 This implies that to estimate the value of the long-term trend in a certain year the filter uses only the information up to that year, even if information for the following years is available. This allows reprising the operative environment of supervisors and policy makers, who must decide when to activate the countercyclical capital buffer without knowing the future values of the credit-to-GDP ratio. Edge and Meisenzahl (2011) observed that the real-time estimates of the gap are not reliable because they suffer of the end-of-sample problem, meaning that each point estimate might be revised when new information becomes available. Drehmann and Tsatsaronis (2014), however, showed that for policy-relevant horizons the gap computed using a trend derived from a double-sided HP filter performs much worse than the gap obtained from a backward-looking one-sided HP filter. 6 For business cycle analysis, the standard smoothing parameter for quarterly data is 1,600 (Hodrick and Prescott, 1997). To adjust λ when the frequency of the data is different from quarterly, Ravn and Uhlig (2002) multiplied λ with the fourth power of the frequency ratio. Thus, in case of annual data the parameter λ should be set at (1/4)4∗1600=6.25. This value, however, would be inappropriate for the analysis of the financial cycle that is on average much longer than the business cycle (Drehmann et al., 2010). Assuming that credit cycles are four times as long as business cycles, we adopt a value of λ equal to 44∗6.25=1600. 7 In Laeven and Valencia (2013), policy interventions in the banking sector are considered significant if at least three out of the following six measures have been used: (a) deposit freezes and/or bank holidays, (b) significant bank nationalisations, (c) bank restructuring gross costs corresponding at least to 3% of GDP, (d) extensive liquidity support (5% of deposits and liabilities to nonresidents), (e) significant guarantees put in place and (f) significant asset purchases (at least 5% of GDP). 8 For countries with multiple occurrences of banking crisis, the credit-to-GDP ratio used corresponds to the first crisis. 9 For each series, we do not consider the signals issued during crisis periods because there is evidence (Demirgüç-Kunt et al., 2006; Laeven and Valencia, 2013) that credit-to-GDP ratios do not always fall during crises episodes. This might happen when the decline in GDP growth due to the crisis is higher than the slowdown in lending because of preexisting credit lines that are drawn upon during the crisis. This statistical increase in the credit-to-GDP during crisis episodes, due to a contraction in GDP rather than to an expansion in credit, may induce noise in the credit-to-GDP gap indicator that might incorrectly signal another future crisis (false alarm). 10 Non-parametric estimation allows also to address the complication arising from the fact that the data are likely to be correlated over time within individual countries (Janes et al., 2009). 11 In the paper, a result deemed significant indicates at least 5% significance if not otherwise indicated. 12 For instance, in Group 1, the C.I. of the BIS gap's AUC is [0.39, 0.76], while the C.I. of the frontier gap's AUC is [0.62, 0.85]. 13 The computation of the AUC differs from the static procedure in that the signal is calculated for any of the five forecasting horizons separately rather than over an interval. This entails that for any point-wise forecast horizon τ, a signal is considered correct if it forecasts a crisis τ years in advance. It is worth noticing that, while the point-wise forecast analysis allows assessing the timing and stability of the signals issued by EWIs, it may lead to inaccurate AUC estimates when there is uncertainty on crisis dates. 14 The six countries are Brazil, Cape Verde, Kazakhstan, Mauritania, Niger and Singapore. 15 World Bank (2010) notices that in low- and middle-income countries, the application of the rule for the countercyclical buffer based on the BIS gap may result in erratic activation and deactivation of the buffer due to higher volatility of credit and GDP growth in those countries. © The Author(s) 2018. Published by Oxford University Press on behalf of the Centre for the Study of African Economies, all rights reserved. For Permissions, please email: journals.permissions@oup.com This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices) http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of African Economies Oxford University Press

Make It or Break It? Assessing Credit Booms in Developing Countries

Loading next page...
 
/lp/ou_press/make-it-or-break-it-assessing-credit-booms-in-developing-countries-kklhkI0huv
Publisher
Oxford University Press
Copyright
© The Author(s) 2018. Published by Oxford University Press on behalf of the Centre for the Study of African Economies, all rights reserved. For Permissions, please email: journals.permissions@oup.com
ISSN
0963-8024
eISSN
1464-3723
D.O.I.
10.1093/jae/ejy004
Publisher site
See Article on Publisher Site

Abstract

Abstract In this paper, we argue that the assessment of systemic risk associated with rapid credit growth should be conducted with different tools at different stages of financial development. In particular, using a signal detection framework, we show that, when the level of financial development is low, the most used leading indicator of banking crisis, namely the credit-to-GDP gap, has poor predictive performance, as it is unable to disentangle episodes of financial deepening from bubble-like credit booms. On the contrary, a new class of indicators, comparing credit levels to structural benchmarks, has good capacity to correctly predict financial crises in economies in the early stages of financial development. 1. Background The past 10 years witnessed a global trend towards increased financial deepening in developing countries (Figure 1). According to an extensive theoretical and empirical literature on the finance-growth nexus, this development is expected to support economic growth, by promoting a higher mobilisation of savings, improving the allocation of capital, enhancing risk management, and facilitating transactions (Rajan and Zingales, 1998; Levine et al., 2000; Levine, 2005). However, there is also a vast literature on financial crises and early warning systems (EWS), mainly based on the experience of advanced economies, which warns that fast acceleration in credit expansion may entail tradeoffs in terms of financial stability (Gourinchas et al., 2001; Cottarelli et al., 2005; Mendoza and Terrones, 2008; Dell’Ariccia et al., 2012; Crowe et al., 2013). Figure 1: View largeDownload slide Rapid Credit Growth in Developing Countries. Source: World Bank World Development Indicators Database. Figure 1: View largeDownload slide Rapid Credit Growth in Developing Countries. Source: World Bank World Development Indicators Database. Evidence from recent empirical literature seems to suggest that the negative effects of credit growth are less likely to manifest in developing countries. Several studies show in fact that credit booms are less frequently associated with systemic banking crises in those countries (Barajas et al., 2007; Dell’Ariccia et al., 2012; Arena et al., 2015; Meng and Gonzalez, 2017). This finding is likely to reflect the fact that, in economies in the early stages of financial development, rapid credit expansions are mostly connected to a healthy process of financial development rather than to the buildup of financial vulnerabilities, but the more the financial sector develops in size and sophistication, the more the risks of additional financial deepening exceed the benefits. Nonetheless, excessive credit growth can be a source of risk also in countries at an early stage of financial development, as there are limits to a country’s capacity to absorb financial deepening at each point in time. The systemic financial crisis experienced in Nigeria in 2009 offers a recent lesson on the dangers of credit booms in developing economies. The crisis followed a consolidation of the banking sector in the years 2005–06,1 which spurred a large credit expansion (Figure 2)—further fuelled by large oil-related inflows and a loose monetary policy stance. The acceleration in credit was too rapid to be absorbed in productive sectors of the economy and significant flows were channelled to non-priority sectors and to the capital markets. When the stress generated by the global financial crisis burst the equity bubble and oil prices collapsed, the credit boom ended in a systemic banking crisis (Sanusi, 2010). Figure 2: View largeDownload slide Nigeria: A Boom Bust Episode. Source: World Bank World Development Indicators Database. Figure 2: View largeDownload slide Nigeria: A Boom Bust Episode. Source: World Bank World Development Indicators Database. The issue is then understanding whether the fast pace of credit growth reflects a process of financial deepening that is beneficial for the real economy (make it), or if instead, it is indicative of systemic risk buildup, detrimental to loan quality and banking system stability (break it). In this paper, we aim to investigate this issue more in detail. We will show that the choice of the metric used to measure excess credit is critical to disentangle episodes of financial deepening from bubble-like credit booms. In particular, our results suggest that traditional leading indicators of systemic risk buildup, which compare credit levels to their long-term trends, are unable to tell the difference between good and bad booms in countries at a low level of financial development. On the contrary, indicators of excess credit comparing credit levels to a structural benchmark have a median good predictive performance in those countries. The findings in this paper are important from a policy perspective. First, assessing when credit growth can pose risks to financial stability is relevant for the activation and calibration of macro-prudential tools. An incorrect assessment of the risks associated to a boom might lead to an erroneous activation or calibration of macro-prudential tools, with the risk of either hindering financial development or, on the opposite side of the spectrum, leaving systemic risk buildup undetected. Second, as the analysis suggests that factors other than excess credit are frequently the source of financial stress in countries at a low level of financial development, alternative early warning indicators must be sought for these countries. The risk is that, by using a leading indicator of crises based on excess credit, sources of systemic risk peculiar to developing countries would pass unnoticed if they do not manifest themselves as credit booms. This paper relates to two strands of literature. First, it is associated with the literature on the nexus between boom-bust episodes and financial crises (Mendoza and Terrones, 2008; Reinhart and Rogoff, 2009; Jorda et al., 2011; Dell’Ariccia et al., 2012; Gourinchas and Obstfeld, 2012; Laeven and Valencia, 2013), and to the empirical research on EWS and indicators of financial crises (Kaminsky and Reinhart, 1999; Borio and Drehmann, 2009). Our work, however, differs from this empirical research in the fact that it focuses on the relationship between financial stability and financial development and proposes different leading indicators of systemic risk buildup depending on the stage of financial development. The paper is also connected to the literature on benchmarking financial systems (Beck et al., 2008; Al-Hussainy et al., 2011), as it uses the financial possibility frontier concept to distinguish between episodes of financial deepening and occurrences of systemic risk buildup in countries at an early stage of financial development. The rest of the paper is organised as follows. Section 2 classifies measures of excess credit into two major categories: cyclical and structural indicators. Section 3 tests the predictive performance of financial crises of the two classes of indicators across income levels and regions. Finally, Section 4 summarises the findings and suggests future research. 2. Metrics of Excess Credit How much credit is too much? There are two possible ways of approaching this question. The first entails assessing credit levels along the time dimension. Based on this approach, credit would be considered excessive if it is significantly above historical values. The second option consists in evaluating credit levels along the cross-sectional dimension. In this case, credit would be regarded as excessive if it is higher than the level recorded in economies with similar structural characteristics. These two approaches permit to define two classes of metrics of excessive credit, which we will refer hereafter as cyclical and structural indicators, respectively. 2.1 Cyclical indicators Based on the evidence of advanced economies and large emerging markets, the early warning literature relates the concept of excess credit to the notion of financial cycles and suggests that ‘peaks in the financial cycle (i.e. booms) are closely associated with systemic banking crises’ (Borio, 2012). A number of empirical and theoretical papers provide explanations for why lending booms may lead to financial stress. One chain of causation links credit booms and banking crises to excessive risk taking during the upswing of the financial cycle. This, in turn, may be stimulated by accommodative monetary policies, especially those in place for extended periods (Gambacorta et al., 2009). These dynamics tend to be amplified by a financial accelerator mechanism, where the supply of credit increases and credit standards are loosened pari passu with an improvement in collateral values (Kiyotaki and Moore, 1997; Bernanke et al., 1999; Gilchrist et al., 2009; Schularick et al., 2012). In general, the boom phase is captured by deviations of a credit measure from its historical trend, thereby defining a gap. Methodologies, however, differ substantially in the choice of the credit measure and in the computation of the trend. Among others, the early warning literature assigns a prominent role to the gap computed as the difference between the credit-to-GDP ratio and its long-term trend, calculated with a backward-looking Hodrick–Prescott (HP) filter with a high smoothing parameter. Empirical research from the Basel Committee on Banking Supervision has shown that, indeed, this gap (BIS gap hereafter) is a valuable leading indicator of systemic banking crises in advanced economies (Borio and Lowe, 2002; Drehmann et al., 2010, 2011), and, as such, Basel III has endorsed it as a guide to set the countercyclical capital buffer (BCBS, 2010). To date, the predictive performance of the BIS gap has been mainly tested on advanced economies. When the analysis has been extended to developing countries and emerging markets, results are mixed. A study by the IMF (2011), which assessed the performance of the BIS gap on a very large sample including advanced economies, emerging markets and low-income countries (169 countries in total), found that, for both emerging markets and low-income countries, the BIS gap did not perform well as a signalling variable. Drehmann and Tsatsaronis (2014), instead, using a sample including both advanced economies and emerging markets, found evidence that in emerging markets the BIS gap remains a good indicator of financial stress, albeit the performance is not as strong as in advanced economies. Emerging markets included in this sample, however, have mainly a developed financial sector. As highlighted in the background section, and argued by a number of commentators (Gerš and Seidler, 2012; IMF, 2014), rapid credit growth in low-income countries and emerging markets may reflect improved economic fundamentals and financial deepening. In this case, the signalling power of the BIS gap would not be compromised only if financial deepening occurs at a steady pace, as this would be embedded in the long-term trend and would not impact on the gap (Drehmann and Tsatsaronis, 2014). If, instead, financial deepening takes the form of sudden and rapid increases in credit growth, these would not be captured in the trend and might be signalled by the gap as buildup of financial vulnerability. 2.2 Structural indicators The literature on benchmarking financial systems offers an alternative approach for measuring excess credit, based on the concept of financial possibility frontier. According to this research strand, the development of a country’s financial system is critically influenced by structural factors that are invariant in the short term and often lie outside the purview of policy makers (Beck et al., 2008; Al-Hussainy et al., 2011). Those factors impose an upper limit to financial deepening in an economy at a given point in time, represented by the financial possibility frontier. This is the efficient (and safe) level of financial development, typically derived from a panel regression that estimates the relationship between a measure of financial development Y and a set of structural indicators X: Yi,t=f(Xi,t)+ϵi,t, where i and t are the country and time indexes, and ϵ is an error term. The difference between actual and predicted levels of financial development provides important information on the status of a country’s financial sector. A negative gap would signal an inefficient financial sector that does not operate at capacity. Instead, overshooting the predicted level of financial development would be associated with overheating pressures and financial stress (Barajas et al., 2013). The gap with respect to the frontier is, thus, a structural indicator of financial performance, as opposed to the gap with respect to the long-term trend, which is a cyclical measure. To date the capacity of structural indicators to predict financial crises has not been investigated. Barajas et al. (2013), however, found evidence that periods with a positive gap (i.e. levels of private credit to GDP above the benchmark) are more likely to be associated with booms that end in low growth episodes or even banking crises (bad booms). Instead, zero gaps (when a country’s private credit to GDP is close to its structural benchmark) have the lowest incidence of bad booms. In addition, they found that large changes in the gap, especially positive changes, are associated with a higher likelihood that the boom will be sub-par or end in a crisis. These findings are promising and suggest that the class of structural indicators could represent an alternative to cyclical indicators as predictors of financial crisis. In addition, these indicators might provide a better measure of excess credit in developing countries, as they account for financial deepening while flagging whether financial development is consistent with a country’s structural characteristics. 3. Systemic Risk Assessment across Levels of Financial Development In this section, we compare the predictive performance of the two classes of indicators across levels of financial development. The goal is to assess whether excess credit should be measured with different metrics at different stages of financial development. 3.1 Data Our analysis is conducted on a sample that includes 79 countries belonging to all income levels and regions (Table 1). Table 1: Sample Composition Country Income Region Data Source First Observation Last Observation Algeria UMI MENA IFS 1970 2011 Argentina UMI LAC IFS 1970 2011 Australia HI EAP BIS 1970 2011 Austria HI ECA BIS 1970 2011 Bangladesh LI EAP IFS 1974 2011 Belgium HI ECA BIS 1971 2011 Benin LI SSA IFS 1970 2011 Bolivia LMI LAC IFS 1970 2011 Brazil UMI LAC IFS 1970 1985 Burkina Faso LI SSA IFS 1970 2011 Burundi LI SSA IFS 1970 2011 Cameroon HI MNA IFS 1970 2011 Canada LMI SSA BIS 1970 2011 Cape Verde UMI SSA IFS 1980 2011 Centr. Afric. Rep. LI SSA IFS 1970 2011 Chad LI SSA IFS 1970 2011 Chile HI LAC IFS 1970 2011 Congo, Dem. Rep. LI SSA IFS 1970 1995 Congo, Rep. LMI SSA IFS 1970 2011 Costa Rica UMI LAC IFS 1970 2011 Cote d’Ivoire LMI SSA IFS 1970 2011 Denmark HI ECA BIS 1970 2011 Dominican Republic UMI LAC IFS 1970 2011 Ecuador UMI LAC IFS 1970 2011 Egypt, Arab Rep. LMI MENA IFS 1970 2011 El Salvador LMI LAC IFS 1970 2011 Finland HI ECA BIS 1971 2011 France HI ECA BIS 1970 2011 Germany HI ECA BIS 1970 2011 Ghana LMI SSA IFS 1970 2011 Greece HI ECA BIS 1971 2011 Guatemala LMI LAC IFS 1970 2011 Guyana LMI LAC IFS 1970 2011 Honduras LMI LAC IFS 1970 2011 Iceland HI ECA IFS 1970 2011 India LMI EAP BIS 1970 2011 Indonesia LMI EAP BIS 1976 2011 Ireland HI ECA BIS 1972 2011 Italy HI ECA BIS 1970 2011 Japan HI EAP BIS 1970 2011 Jordan UMI MENA IFS 1970 2011 Kazakhstan LMI ECA IFS 1993 2011 Kenya LI SSA IFS 1970 2011 Korea, Rep. HI EAP BIS 1970 2011 Madagascar LI SSA IFS 1970 2011 Malaysia UMI EAP IFS 1970 2011 Mali LI SSA IFS 1970 2011 Mauritania LI MENA IFS 1970 2011 Mexico UMI LAC BIS 1991 2011 Mongolia UMI LAC IFS 1991 2011 Nepal LI EAP IFS 1970 2011 Netherlands HI ECA BIS 1970 2011 New Zealand HI EAP IFS 1970 2010 Nicaragua LMI LAC IFS 1970 2011 Niger LI SSA IFS 1970 2011 Nigeria LMI SSA IFS 1970 2011 Norway HI ECA BIS 1970 2011 Panama LMI LAC IFS 1970 2011 Paraguay LMI LAC IFS 1970 2011 Peru UMI LAC IFS 1970 2011 Philippines LMI EAP IFS 1970 2011 Portugal HI ECA BIS 1970 2011 Senegal LMI SSA IFS 1970 2011 Sierra Leone LI SSA IFS 1970 2011 Singapore HI EAP BIS 1971 2011 South Africa LMI SSA BIS 1970 2011 Spain HI ECA BIS 1972 2011 Sri Lanka LMI EAP IFS 1970 2011 Swaziland LMI SSA IFS 1973 2011 Sweden HI ECA BIS 1970 2011 Switzerland HI ECA BIS 1970 2011 Thailand UMI EAP BIS 1970 2011 Togo LI SSA IFS 1970 2011 Tunisia UMI MENA IFS 1970 2011 Turkey UMI ECA IFS 1970 2011 UK HI ECA BIS 1970 2011 USA HI MNA BIS 1970 2011 Uruguay HI LAC IFS 1970 2011 Venezuela, RB UMI LAC IFS 1970 2011 Country Income Region Data Source First Observation Last Observation Algeria UMI MENA IFS 1970 2011 Argentina UMI LAC IFS 1970 2011 Australia HI EAP BIS 1970 2011 Austria HI ECA BIS 1970 2011 Bangladesh LI EAP IFS 1974 2011 Belgium HI ECA BIS 1971 2011 Benin LI SSA IFS 1970 2011 Bolivia LMI LAC IFS 1970 2011 Brazil UMI LAC IFS 1970 1985 Burkina Faso LI SSA IFS 1970 2011 Burundi LI SSA IFS 1970 2011 Cameroon HI MNA IFS 1970 2011 Canada LMI SSA BIS 1970 2011 Cape Verde UMI SSA IFS 1980 2011 Centr. Afric. Rep. LI SSA IFS 1970 2011 Chad LI SSA IFS 1970 2011 Chile HI LAC IFS 1970 2011 Congo, Dem. Rep. LI SSA IFS 1970 1995 Congo, Rep. LMI SSA IFS 1970 2011 Costa Rica UMI LAC IFS 1970 2011 Cote d’Ivoire LMI SSA IFS 1970 2011 Denmark HI ECA BIS 1970 2011 Dominican Republic UMI LAC IFS 1970 2011 Ecuador UMI LAC IFS 1970 2011 Egypt, Arab Rep. LMI MENA IFS 1970 2011 El Salvador LMI LAC IFS 1970 2011 Finland HI ECA BIS 1971 2011 France HI ECA BIS 1970 2011 Germany HI ECA BIS 1970 2011 Ghana LMI SSA IFS 1970 2011 Greece HI ECA BIS 1971 2011 Guatemala LMI LAC IFS 1970 2011 Guyana LMI LAC IFS 1970 2011 Honduras LMI LAC IFS 1970 2011 Iceland HI ECA IFS 1970 2011 India LMI EAP BIS 1970 2011 Indonesia LMI EAP BIS 1976 2011 Ireland HI ECA BIS 1972 2011 Italy HI ECA BIS 1970 2011 Japan HI EAP BIS 1970 2011 Jordan UMI MENA IFS 1970 2011 Kazakhstan LMI ECA IFS 1993 2011 Kenya LI SSA IFS 1970 2011 Korea, Rep. HI EAP BIS 1970 2011 Madagascar LI SSA IFS 1970 2011 Malaysia UMI EAP IFS 1970 2011 Mali LI SSA IFS 1970 2011 Mauritania LI MENA IFS 1970 2011 Mexico UMI LAC BIS 1991 2011 Mongolia UMI LAC IFS 1991 2011 Nepal LI EAP IFS 1970 2011 Netherlands HI ECA BIS 1970 2011 New Zealand HI EAP IFS 1970 2010 Nicaragua LMI LAC IFS 1970 2011 Niger LI SSA IFS 1970 2011 Nigeria LMI SSA IFS 1970 2011 Norway HI ECA BIS 1970 2011 Panama LMI LAC IFS 1970 2011 Paraguay LMI LAC IFS 1970 2011 Peru UMI LAC IFS 1970 2011 Philippines LMI EAP IFS 1970 2011 Portugal HI ECA BIS 1970 2011 Senegal LMI SSA IFS 1970 2011 Sierra Leone LI SSA IFS 1970 2011 Singapore HI EAP BIS 1971 2011 South Africa LMI SSA BIS 1970 2011 Spain HI ECA BIS 1972 2011 Sri Lanka LMI EAP IFS 1970 2011 Swaziland LMI SSA IFS 1973 2011 Sweden HI ECA BIS 1970 2011 Switzerland HI ECA BIS 1970 2011 Thailand UMI EAP BIS 1970 2011 Togo LI SSA IFS 1970 2011 Tunisia UMI MENA IFS 1970 2011 Turkey UMI ECA IFS 1970 2011 UK HI ECA BIS 1970 2011 USA HI MNA BIS 1970 2011 Uruguay HI LAC IFS 1970 2011 Venezuela, RB UMI LAC IFS 1970 2011 Note: LI, LMI, UMI, HI stand for low income, lower middle income, upper middle income and high income, respectively. SSA, MENA, LAC, EAP, SA, NA and ECA stand for sub-Saharan Africa, Middle-East and North Africa, Latina America and Caribbean, East Asia and Pacific, South Asia, North America, and Europe and Central Asia, respectively. Income level classification is from the World Bank and refers to 2013. View Large Table 1: Sample Composition Country Income Region Data Source First Observation Last Observation Algeria UMI MENA IFS 1970 2011 Argentina UMI LAC IFS 1970 2011 Australia HI EAP BIS 1970 2011 Austria HI ECA BIS 1970 2011 Bangladesh LI EAP IFS 1974 2011 Belgium HI ECA BIS 1971 2011 Benin LI SSA IFS 1970 2011 Bolivia LMI LAC IFS 1970 2011 Brazil UMI LAC IFS 1970 1985 Burkina Faso LI SSA IFS 1970 2011 Burundi LI SSA IFS 1970 2011 Cameroon HI MNA IFS 1970 2011 Canada LMI SSA BIS 1970 2011 Cape Verde UMI SSA IFS 1980 2011 Centr. Afric. Rep. LI SSA IFS 1970 2011 Chad LI SSA IFS 1970 2011 Chile HI LAC IFS 1970 2011 Congo, Dem. Rep. LI SSA IFS 1970 1995 Congo, Rep. LMI SSA IFS 1970 2011 Costa Rica UMI LAC IFS 1970 2011 Cote d’Ivoire LMI SSA IFS 1970 2011 Denmark HI ECA BIS 1970 2011 Dominican Republic UMI LAC IFS 1970 2011 Ecuador UMI LAC IFS 1970 2011 Egypt, Arab Rep. LMI MENA IFS 1970 2011 El Salvador LMI LAC IFS 1970 2011 Finland HI ECA BIS 1971 2011 France HI ECA BIS 1970 2011 Germany HI ECA BIS 1970 2011 Ghana LMI SSA IFS 1970 2011 Greece HI ECA BIS 1971 2011 Guatemala LMI LAC IFS 1970 2011 Guyana LMI LAC IFS 1970 2011 Honduras LMI LAC IFS 1970 2011 Iceland HI ECA IFS 1970 2011 India LMI EAP BIS 1970 2011 Indonesia LMI EAP BIS 1976 2011 Ireland HI ECA BIS 1972 2011 Italy HI ECA BIS 1970 2011 Japan HI EAP BIS 1970 2011 Jordan UMI MENA IFS 1970 2011 Kazakhstan LMI ECA IFS 1993 2011 Kenya LI SSA IFS 1970 2011 Korea, Rep. HI EAP BIS 1970 2011 Madagascar LI SSA IFS 1970 2011 Malaysia UMI EAP IFS 1970 2011 Mali LI SSA IFS 1970 2011 Mauritania LI MENA IFS 1970 2011 Mexico UMI LAC BIS 1991 2011 Mongolia UMI LAC IFS 1991 2011 Nepal LI EAP IFS 1970 2011 Netherlands HI ECA BIS 1970 2011 New Zealand HI EAP IFS 1970 2010 Nicaragua LMI LAC IFS 1970 2011 Niger LI SSA IFS 1970 2011 Nigeria LMI SSA IFS 1970 2011 Norway HI ECA BIS 1970 2011 Panama LMI LAC IFS 1970 2011 Paraguay LMI LAC IFS 1970 2011 Peru UMI LAC IFS 1970 2011 Philippines LMI EAP IFS 1970 2011 Portugal HI ECA BIS 1970 2011 Senegal LMI SSA IFS 1970 2011 Sierra Leone LI SSA IFS 1970 2011 Singapore HI EAP BIS 1971 2011 South Africa LMI SSA BIS 1970 2011 Spain HI ECA BIS 1972 2011 Sri Lanka LMI EAP IFS 1970 2011 Swaziland LMI SSA IFS 1973 2011 Sweden HI ECA BIS 1970 2011 Switzerland HI ECA BIS 1970 2011 Thailand UMI EAP BIS 1970 2011 Togo LI SSA IFS 1970 2011 Tunisia UMI MENA IFS 1970 2011 Turkey UMI ECA IFS 1970 2011 UK HI ECA BIS 1970 2011 USA HI MNA BIS 1970 2011 Uruguay HI LAC IFS 1970 2011 Venezuela, RB UMI LAC IFS 1970 2011 Country Income Region Data Source First Observation Last Observation Algeria UMI MENA IFS 1970 2011 Argentina UMI LAC IFS 1970 2011 Australia HI EAP BIS 1970 2011 Austria HI ECA BIS 1970 2011 Bangladesh LI EAP IFS 1974 2011 Belgium HI ECA BIS 1971 2011 Benin LI SSA IFS 1970 2011 Bolivia LMI LAC IFS 1970 2011 Brazil UMI LAC IFS 1970 1985 Burkina Faso LI SSA IFS 1970 2011 Burundi LI SSA IFS 1970 2011 Cameroon HI MNA IFS 1970 2011 Canada LMI SSA BIS 1970 2011 Cape Verde UMI SSA IFS 1980 2011 Centr. Afric. Rep. LI SSA IFS 1970 2011 Chad LI SSA IFS 1970 2011 Chile HI LAC IFS 1970 2011 Congo, Dem. Rep. LI SSA IFS 1970 1995 Congo, Rep. LMI SSA IFS 1970 2011 Costa Rica UMI LAC IFS 1970 2011 Cote d’Ivoire LMI SSA IFS 1970 2011 Denmark HI ECA BIS 1970 2011 Dominican Republic UMI LAC IFS 1970 2011 Ecuador UMI LAC IFS 1970 2011 Egypt, Arab Rep. LMI MENA IFS 1970 2011 El Salvador LMI LAC IFS 1970 2011 Finland HI ECA BIS 1971 2011 France HI ECA BIS 1970 2011 Germany HI ECA BIS 1970 2011 Ghana LMI SSA IFS 1970 2011 Greece HI ECA BIS 1971 2011 Guatemala LMI LAC IFS 1970 2011 Guyana LMI LAC IFS 1970 2011 Honduras LMI LAC IFS 1970 2011 Iceland HI ECA IFS 1970 2011 India LMI EAP BIS 1970 2011 Indonesia LMI EAP BIS 1976 2011 Ireland HI ECA BIS 1972 2011 Italy HI ECA BIS 1970 2011 Japan HI EAP BIS 1970 2011 Jordan UMI MENA IFS 1970 2011 Kazakhstan LMI ECA IFS 1993 2011 Kenya LI SSA IFS 1970 2011 Korea, Rep. HI EAP BIS 1970 2011 Madagascar LI SSA IFS 1970 2011 Malaysia UMI EAP IFS 1970 2011 Mali LI SSA IFS 1970 2011 Mauritania LI MENA IFS 1970 2011 Mexico UMI LAC BIS 1991 2011 Mongolia UMI LAC IFS 1991 2011 Nepal LI EAP IFS 1970 2011 Netherlands HI ECA BIS 1970 2011 New Zealand HI EAP IFS 1970 2010 Nicaragua LMI LAC IFS 1970 2011 Niger LI SSA IFS 1970 2011 Nigeria LMI SSA IFS 1970 2011 Norway HI ECA BIS 1970 2011 Panama LMI LAC IFS 1970 2011 Paraguay LMI LAC IFS 1970 2011 Peru UMI LAC IFS 1970 2011 Philippines LMI EAP IFS 1970 2011 Portugal HI ECA BIS 1970 2011 Senegal LMI SSA IFS 1970 2011 Sierra Leone LI SSA IFS 1970 2011 Singapore HI EAP BIS 1971 2011 South Africa LMI SSA BIS 1970 2011 Spain HI ECA BIS 1972 2011 Sri Lanka LMI EAP IFS 1970 2011 Swaziland LMI SSA IFS 1973 2011 Sweden HI ECA BIS 1970 2011 Switzerland HI ECA BIS 1970 2011 Thailand UMI EAP BIS 1970 2011 Togo LI SSA IFS 1970 2011 Tunisia UMI MENA IFS 1970 2011 Turkey UMI ECA IFS 1970 2011 UK HI ECA BIS 1970 2011 USA HI MNA BIS 1970 2011 Uruguay HI LAC IFS 1970 2011 Venezuela, RB UMI LAC IFS 1970 2011 Note: LI, LMI, UMI, HI stand for low income, lower middle income, upper middle income and high income, respectively. SSA, MENA, LAC, EAP, SA, NA and ECA stand for sub-Saharan Africa, Middle-East and North Africa, Latina America and Caribbean, East Asia and Pacific, South Asia, North America, and Europe and Central Asia, respectively. Income level classification is from the World Bank and refers to 2013. View Large The cyclical measure of excess credit tested in this section is the BIS gap, the most widely used leading indicator of financial crisis in the EWS literature. The gap for each country is computed as the percentage deviation of the credit-to-GDP ratio from its long-term trend. The series of private credit-to-GDP is derived from two alternative sources, depending on data availability. For 27 countries,2 mainly advanced economies and large emerging markets, we use the new BIS series of total credit to the private sector, adjusted for structural breaks.3 Quarterly data have been averaged to form an annual series. For the remaining countries, we use the annual series of domestic credit to the private non-financial sector from the World Bank Global Financial Development Database.4 We require that the series of each country does not include data gaps and that data are available for at least 10 years prior to a crisis. The estimates of the long-term trend are obtained by using a one-sided (backward-looking) HP filter, which allows us to obtain real-time estimates of the gap.5 The smoothing parameter λ is set equal to 1600.6 Finally, to obtain more robust estimates of the trend, we start the computation 10 years after the beginning of each series so that the HP filter can use a minimum of 10 observations for each data point estimate of the trend. As far as regards structural indicators, we use a gap (frontier gap hereafter) computed as the percentage deviation of a country’s credit-to-GDP ratio from its financial possibility frontier (or benchmark), as retrieved from FinStats. This is a tool developed by the World Bank and updated every year, that implements the methodology in Beck et al. (2008) and estimates frontiers for the quasi-totality of countries in the world (177 countries) through a pooled quantile (median) regression (Feyen et al., 2015). Frontiers for each country are obtained by regressing private credit to GDP on a set of structural characteristics, including GDP per capita and its square (to account for potential non-linearities linking economic and financial development, see e.g., Arcand et al., 2015), population size and density, the age dependency ratio and year-fixed effects. Regressors also include a number of dummies to control for additional structural factors, including a dummy for natural resource exporters, as worldwide evidence shows that resource rich countries tend to have comparatively smaller financial sectors than other countries at similar levels of income, reflecting the fact that oil revenues can boost GDP out of proportion with the country’s overall level of economic and financial development. The identification and timing of banking crises are based on the systemic banking crises database by Laeven and Valencia (2013). The authors define a banking crisis as an event that satisfies two conditions: (a) significant signs of financial distress in the banking system (as indicated by significant bank runs, losses in the banking system, and/or bank liquidations) and (b) significant policy intervention measures in response to large losses in the banking sector.7 Given the paucity of banking crises based on this definition, we extend the Laeven and Valencia (2013) database based on Reinhart (2010) that uses an alternative definition of banking crisis, as an event that satisfies one of the following two conditions: (a) bank runs that lead to the closure, merging, or takeover by the public sector of one or more financial institutions or (b) if there are no runs, the closure, merger, takeover, or large-scale government support of the banking sector (Reinhart and Rogoff, 2009). Table 2 lists the crisis episodes used in this paper (first column), satisfying either Reinhart (2010) (second column of the Table) or Laeven and Valencia (2013) (third column) definitions. Table 2: Systemic Banking Crises, 1980–2010 Country Paper1 R2 LV3 Start End Start End Start End Algeria 1990 1994 1990 1992 1990 1994 Argentina 1980 1982 1980 1982 1980 1982 Argentina 1989 1991 1989 1990 1989 1991 Argentina 1995 1996 1995 1996 1995 1995 Argentina 2001 2003 2001 2003 2001 2003 Australia 1989 1992 1989 1992 n.a. n.a. Austria 2008 2011 2008 2010 2008 2011 Bangladesh 1987 1987 n.a. n.a. 1987 1987 Belgium 2008 2011 2008 2010 2008 2011 Benin 1988 1992 n.a. n.a. 1988 1992 Bolivia 1986 1987 1986 1987 1986 1986 Bolivia 1994 1997 1994 1997 1994 1994 Bolivia 1999 1999 1999 1999 n.a. n.a. Brazil 1985 1985 1985 1985 n.a. n.a. Brazil 1990 1994 1990 1990 1990 1994 Burkina Faso 1990 1994 n.a. n.a. 1990 1994 Burundi 1994 1998 n.a. n.a. 1994 1998 Canada 1983 1985 1983 1985 n.a. n.a. Cape Verde 1993 1993 n.a. n.a. 1993 1993 CAR 1980 1982 1976 1982 1976 1976 CAR 1988 1999 1988 1999 1995 1996 Chad 1983 1983 n.a. n.a. 1983 1983 Chad 1992 1996 n.a. n.a. 1992 1996 Chile 1981 1985 1982 1984 1981 1985 Congo, Dem. Rep. 1983 1983 n.a. n.a. 1983 1983 Congo, Dem. Rep. 1991 1994 n.a. n.a. 1991 1994 Congo, Dem. Rep. 1994 1998 n.a. n.a. 1994 1998 Congo, Rep. 1992 1994 n.a. n.a. 1992 1994 Costa Rica 1987 1991 1987 1987 1987 1991 Costa Rica 1994 1996 1994 1996 1994 1995 Cote dIvoire 1988 1992 1988 1991 1988 1992 Denmark 1987 1992 1987 1992 n.a. n.a. Denmark 2008 2011 2008 2010 2008 2011 Dominican Rep 1996 1996 1996 1996 n.a. n.a. Dominican Rep 2003 2004 2003 2003 2003 2004 Ecuador 1981 1986 1981 1981 1982 1986 Ecuador 1998 2002 1998 2002 1998 2002 Egypt 1980 1983 1981 1983 1980 1980 Egypt 1990 1995 1990 1995 n.a. n.a. El Salvador 1989 1990 1989 1989 1989 1990 Finland 1991 1995 1991 1994 1991 1995 France 1994 1995 1994 1995 n.a. n.a. France 2008 2011 2008 2010 2008 2011 Germany 1991 1994 1991 1994 n.a. n.a. Germany 2008 2011 n.a. n.a. 2008 2011 Ghana 1982 1989 1982 1989 1982 1983 Ghana 1997 1997 1997 1997 n.a. n.a. Greece 1991 1995 1991 1995 n.a. n.a. Greece 2008 2011 2008 2010 2008 2011 Guatemala 1990 1990 1990 1990 n.a. n.a. Guatemala 2001 2001 2001 2001 n.a. n.a. Guatemala 2006 2006 2006 2006 n.a. n.a. Guyana 1993 1993 n.a. n.a. 1993 1993 Honduras 1999 1999 1999 1999 n.a. n.a. Honduras 2001 2002 2001 2002 n.a. n.a. Iceland 1985 1986 1985 1986 n.a. n.a. Iceland 1993 1993 1993 1993 n.a. n.a. Iceland 2008 2011 2007 2010 2008 2011 India 1993 1998 1993 1998 1993 1993 Indonesia 1992 1992 1992 1992 n.a. n.a. Indonesia 1994 1994 1994 1994 n.a. n.a. Indonesia 1997 2002 1997 2002 1997 2001 Ireland 2007 2011 2007 2010 2008 2011 Italy 1990 1995 1990 1995 n.a. n.a. Italy 2008 2011 n.a. n.a. 2008 2011 Japan 1992 2001 1992 2001 1997 2001 Jordan 1989 1991 n.a. n.a. 1989 1991 Kazakhstan 2008 2011 n.a. n.a. 2008 2011 Kenya 1985 1989 1985 1989 1985 1985 Kenya 1992 1995 1992 1995 1992 1994 Korea 1983 1983 1983 1983 n.a. n.a. Korea 1985 1988 1985 1988 n.a. n.a. Korea 1997 2002 1997 2002 1997 1998 Madagascar 1988 1988 n.a. n.a. 1988 1988 Malaysia 1985 1988 1985 1988 n.a. n.a. Malaysia 1997 2001 1997 2001 1997 1999 Mali 1987 1991 n.a. n.a. 1987 1991 Mauritania 1984 1984 n.a. n.a. 1984 1984 Mexico 1981 1985 1981 1982 1981 1985 Mexico 1994 2000 1994 2000 1994 1996 Mongolia 2008 2011 n.a. n.a. 2008 2011 Nepal 1988 1988 n.a. n.a. 1988 1988 Netherlands 2008 2011 2008 2010 2008 2011 New Zealand 1987 1990 1987 1990 n.a. n.a. Nicaragua 1987 1996 1987 1996 n.a. n.a. Nicaragua 1990 1993 n.a. n.a. 1990 1993 Nicaragua 2000 2002 2000 2002 2000 2001 Niger 1983 1985 n.a. n.a. 1983 1985 Nigeria 1992 1995 1992 1995 1991 1995 Nigeria 1997 1997 1997 1997 n.a. n.a. Nigeria 2009 2011 n.a. n.a. 2009 2011 Norway 1987 1993 1987 1993 1991 1993 Panama 1988 1989 1988 1989 1988 1989 Paraguay 1995 1999 1995 1999 1995 1995 Paraguay 2002 2002 2002 2002 n.a. n.a. Peru 1983 1990 1983 1990 1983 1983 Peru 1999 1999 1999 1999 n.a. n.a. Philippines 1981 1987 1981 1987 1983 1986 Philippines 1997 2001 1997 2001 1997 2001 Portugal 2008 2011 2008 2010 2008 2011 Senegal 1988 1991 n.a. n.a. 1988 1991 Sierra Leone 1990 1994 n.a. n.a. 1990 1994 Singapore 1982 1983 1982 1983 n.a. n.a. South Africa 1989 1989 1989 1989 n.a. n.a. Spain 1980 1985 1977 1985 1977 1981 Spain 2008 2011 2008 2010 2008 2011 Sri Lanka 1989 1993 1989 1993 1989 1991 Swaziland 1995 1999 n.a. n.a. 1995 1999 Sweden 1991 1995 1991 1994 1991 1995 Sweden 2008 2011 n.a. n.a. 2008 2011 Switzerland 2008 2011 2008 2009 2008 2011 Thailand 1980 1987 1980 1987 1983 1983 Thailand 1996 2001 1996 2001 1997 2000 Togo 1993 1994 n.a. n.a. 1993 1994 Tunisia 1991 1991 n.a. n.a. 1991 1991 Turkey 1982 1985 1982 1985 1982 1984 Turkey 1991 1991 1991 1991 n.a. n.a. Turkey 1994 1994 1994 1994 n.a. n.a. Turkey 2000 2001 2000 2000 2000 2001 UK 1984 1984 1984 1984 n.a. n.a. UK 1991 1991 1991 1991 n.a. n.a. UK 1995 1995 1995 1995 n.a. n.a. UK 2007 2011 2007 2009 2007 2011 USA 1984 1991 1984 1991 1988 1988 USA 2007 2011 2007 2010 2007 2011 Uruguay 1981 1985 n.a. n.a. 1981 1985 Uruguay 2002 2005 n.a. n.a. 2002 2005 Venezuela 1980 1986 1978 1986 n.a. n.a. Venezuela 1993 1998 1993 1994 1994 1998 Country Paper1 R2 LV3 Start End Start End Start End Algeria 1990 1994 1990 1992 1990 1994 Argentina 1980 1982 1980 1982 1980 1982 Argentina 1989 1991 1989 1990 1989 1991 Argentina 1995 1996 1995 1996 1995 1995 Argentina 2001 2003 2001 2003 2001 2003 Australia 1989 1992 1989 1992 n.a. n.a. Austria 2008 2011 2008 2010 2008 2011 Bangladesh 1987 1987 n.a. n.a. 1987 1987 Belgium 2008 2011 2008 2010 2008 2011 Benin 1988 1992 n.a. n.a. 1988 1992 Bolivia 1986 1987 1986 1987 1986 1986 Bolivia 1994 1997 1994 1997 1994 1994 Bolivia 1999 1999 1999 1999 n.a. n.a. Brazil 1985 1985 1985 1985 n.a. n.a. Brazil 1990 1994 1990 1990 1990 1994 Burkina Faso 1990 1994 n.a. n.a. 1990 1994 Burundi 1994 1998 n.a. n.a. 1994 1998 Canada 1983 1985 1983 1985 n.a. n.a. Cape Verde 1993 1993 n.a. n.a. 1993 1993 CAR 1980 1982 1976 1982 1976 1976 CAR 1988 1999 1988 1999 1995 1996 Chad 1983 1983 n.a. n.a. 1983 1983 Chad 1992 1996 n.a. n.a. 1992 1996 Chile 1981 1985 1982 1984 1981 1985 Congo, Dem. Rep. 1983 1983 n.a. n.a. 1983 1983 Congo, Dem. Rep. 1991 1994 n.a. n.a. 1991 1994 Congo, Dem. Rep. 1994 1998 n.a. n.a. 1994 1998 Congo, Rep. 1992 1994 n.a. n.a. 1992 1994 Costa Rica 1987 1991 1987 1987 1987 1991 Costa Rica 1994 1996 1994 1996 1994 1995 Cote dIvoire 1988 1992 1988 1991 1988 1992 Denmark 1987 1992 1987 1992 n.a. n.a. Denmark 2008 2011 2008 2010 2008 2011 Dominican Rep 1996 1996 1996 1996 n.a. n.a. Dominican Rep 2003 2004 2003 2003 2003 2004 Ecuador 1981 1986 1981 1981 1982 1986 Ecuador 1998 2002 1998 2002 1998 2002 Egypt 1980 1983 1981 1983 1980 1980 Egypt 1990 1995 1990 1995 n.a. n.a. El Salvador 1989 1990 1989 1989 1989 1990 Finland 1991 1995 1991 1994 1991 1995 France 1994 1995 1994 1995 n.a. n.a. France 2008 2011 2008 2010 2008 2011 Germany 1991 1994 1991 1994 n.a. n.a. Germany 2008 2011 n.a. n.a. 2008 2011 Ghana 1982 1989 1982 1989 1982 1983 Ghana 1997 1997 1997 1997 n.a. n.a. Greece 1991 1995 1991 1995 n.a. n.a. Greece 2008 2011 2008 2010 2008 2011 Guatemala 1990 1990 1990 1990 n.a. n.a. Guatemala 2001 2001 2001 2001 n.a. n.a. Guatemala 2006 2006 2006 2006 n.a. n.a. Guyana 1993 1993 n.a. n.a. 1993 1993 Honduras 1999 1999 1999 1999 n.a. n.a. Honduras 2001 2002 2001 2002 n.a. n.a. Iceland 1985 1986 1985 1986 n.a. n.a. Iceland 1993 1993 1993 1993 n.a. n.a. Iceland 2008 2011 2007 2010 2008 2011 India 1993 1998 1993 1998 1993 1993 Indonesia 1992 1992 1992 1992 n.a. n.a. Indonesia 1994 1994 1994 1994 n.a. n.a. Indonesia 1997 2002 1997 2002 1997 2001 Ireland 2007 2011 2007 2010 2008 2011 Italy 1990 1995 1990 1995 n.a. n.a. Italy 2008 2011 n.a. n.a. 2008 2011 Japan 1992 2001 1992 2001 1997 2001 Jordan 1989 1991 n.a. n.a. 1989 1991 Kazakhstan 2008 2011 n.a. n.a. 2008 2011 Kenya 1985 1989 1985 1989 1985 1985 Kenya 1992 1995 1992 1995 1992 1994 Korea 1983 1983 1983 1983 n.a. n.a. Korea 1985 1988 1985 1988 n.a. n.a. Korea 1997 2002 1997 2002 1997 1998 Madagascar 1988 1988 n.a. n.a. 1988 1988 Malaysia 1985 1988 1985 1988 n.a. n.a. Malaysia 1997 2001 1997 2001 1997 1999 Mali 1987 1991 n.a. n.a. 1987 1991 Mauritania 1984 1984 n.a. n.a. 1984 1984 Mexico 1981 1985 1981 1982 1981 1985 Mexico 1994 2000 1994 2000 1994 1996 Mongolia 2008 2011 n.a. n.a. 2008 2011 Nepal 1988 1988 n.a. n.a. 1988 1988 Netherlands 2008 2011 2008 2010 2008 2011 New Zealand 1987 1990 1987 1990 n.a. n.a. Nicaragua 1987 1996 1987 1996 n.a. n.a. Nicaragua 1990 1993 n.a. n.a. 1990 1993 Nicaragua 2000 2002 2000 2002 2000 2001 Niger 1983 1985 n.a. n.a. 1983 1985 Nigeria 1992 1995 1992 1995 1991 1995 Nigeria 1997 1997 1997 1997 n.a. n.a. Nigeria 2009 2011 n.a. n.a. 2009 2011 Norway 1987 1993 1987 1993 1991 1993 Panama 1988 1989 1988 1989 1988 1989 Paraguay 1995 1999 1995 1999 1995 1995 Paraguay 2002 2002 2002 2002 n.a. n.a. Peru 1983 1990 1983 1990 1983 1983 Peru 1999 1999 1999 1999 n.a. n.a. Philippines 1981 1987 1981 1987 1983 1986 Philippines 1997 2001 1997 2001 1997 2001 Portugal 2008 2011 2008 2010 2008 2011 Senegal 1988 1991 n.a. n.a. 1988 1991 Sierra Leone 1990 1994 n.a. n.a. 1990 1994 Singapore 1982 1983 1982 1983 n.a. n.a. South Africa 1989 1989 1989 1989 n.a. n.a. Spain 1980 1985 1977 1985 1977 1981 Spain 2008 2011 2008 2010 2008 2011 Sri Lanka 1989 1993 1989 1993 1989 1991 Swaziland 1995 1999 n.a. n.a. 1995 1999 Sweden 1991 1995 1991 1994 1991 1995 Sweden 2008 2011 n.a. n.a. 2008 2011 Switzerland 2008 2011 2008 2009 2008 2011 Thailand 1980 1987 1980 1987 1983 1983 Thailand 1996 2001 1996 2001 1997 2000 Togo 1993 1994 n.a. n.a. 1993 1994 Tunisia 1991 1991 n.a. n.a. 1991 1991 Turkey 1982 1985 1982 1985 1982 1984 Turkey 1991 1991 1991 1991 n.a. n.a. Turkey 1994 1994 1994 1994 n.a. n.a. Turkey 2000 2001 2000 2000 2000 2001 UK 1984 1984 1984 1984 n.a. n.a. UK 1991 1991 1991 1991 n.a. n.a. UK 1995 1995 1995 1995 n.a. n.a. UK 2007 2011 2007 2009 2007 2011 USA 1984 1991 1984 1991 1988 1988 USA 2007 2011 2007 2010 2007 2011 Uruguay 1981 1985 n.a. n.a. 1981 1985 Uruguay 2002 2005 n.a. n.a. 2002 2005 Venezuela 1980 1986 1978 1986 n.a. n.a. Venezuela 1993 1998 1993 1994 1994 1998 1Crisis period used in the paper, equal to R ∪ LV. 2Crisis period according to Reinhart (2010). 3Crisis period according to Laeven and Valencia (2013). View Large Table 2: Systemic Banking Crises, 1980–2010 Country Paper1 R2 LV3 Start End Start End Start End Algeria 1990 1994 1990 1992 1990 1994 Argentina 1980 1982 1980 1982 1980 1982 Argentina 1989 1991 1989 1990 1989 1991 Argentina 1995 1996 1995 1996 1995 1995 Argentina 2001 2003 2001 2003 2001 2003 Australia 1989 1992 1989 1992 n.a. n.a. Austria 2008 2011 2008 2010 2008 2011 Bangladesh 1987 1987 n.a. n.a. 1987 1987 Belgium 2008 2011 2008 2010 2008 2011 Benin 1988 1992 n.a. n.a. 1988 1992 Bolivia 1986 1987 1986 1987 1986 1986 Bolivia 1994 1997 1994 1997 1994 1994 Bolivia 1999 1999 1999 1999 n.a. n.a. Brazil 1985 1985 1985 1985 n.a. n.a. Brazil 1990 1994 1990 1990 1990 1994 Burkina Faso 1990 1994 n.a. n.a. 1990 1994 Burundi 1994 1998 n.a. n.a. 1994 1998 Canada 1983 1985 1983 1985 n.a. n.a. Cape Verde 1993 1993 n.a. n.a. 1993 1993 CAR 1980 1982 1976 1982 1976 1976 CAR 1988 1999 1988 1999 1995 1996 Chad 1983 1983 n.a. n.a. 1983 1983 Chad 1992 1996 n.a. n.a. 1992 1996 Chile 1981 1985 1982 1984 1981 1985 Congo, Dem. Rep. 1983 1983 n.a. n.a. 1983 1983 Congo, Dem. Rep. 1991 1994 n.a. n.a. 1991 1994 Congo, Dem. Rep. 1994 1998 n.a. n.a. 1994 1998 Congo, Rep. 1992 1994 n.a. n.a. 1992 1994 Costa Rica 1987 1991 1987 1987 1987 1991 Costa Rica 1994 1996 1994 1996 1994 1995 Cote dIvoire 1988 1992 1988 1991 1988 1992 Denmark 1987 1992 1987 1992 n.a. n.a. Denmark 2008 2011 2008 2010 2008 2011 Dominican Rep 1996 1996 1996 1996 n.a. n.a. Dominican Rep 2003 2004 2003 2003 2003 2004 Ecuador 1981 1986 1981 1981 1982 1986 Ecuador 1998 2002 1998 2002 1998 2002 Egypt 1980 1983 1981 1983 1980 1980 Egypt 1990 1995 1990 1995 n.a. n.a. El Salvador 1989 1990 1989 1989 1989 1990 Finland 1991 1995 1991 1994 1991 1995 France 1994 1995 1994 1995 n.a. n.a. France 2008 2011 2008 2010 2008 2011 Germany 1991 1994 1991 1994 n.a. n.a. Germany 2008 2011 n.a. n.a. 2008 2011 Ghana 1982 1989 1982 1989 1982 1983 Ghana 1997 1997 1997 1997 n.a. n.a. Greece 1991 1995 1991 1995 n.a. n.a. Greece 2008 2011 2008 2010 2008 2011 Guatemala 1990 1990 1990 1990 n.a. n.a. Guatemala 2001 2001 2001 2001 n.a. n.a. Guatemala 2006 2006 2006 2006 n.a. n.a. Guyana 1993 1993 n.a. n.a. 1993 1993 Honduras 1999 1999 1999 1999 n.a. n.a. Honduras 2001 2002 2001 2002 n.a. n.a. Iceland 1985 1986 1985 1986 n.a. n.a. Iceland 1993 1993 1993 1993 n.a. n.a. Iceland 2008 2011 2007 2010 2008 2011 India 1993 1998 1993 1998 1993 1993 Indonesia 1992 1992 1992 1992 n.a. n.a. Indonesia 1994 1994 1994 1994 n.a. n.a. Indonesia 1997 2002 1997 2002 1997 2001 Ireland 2007 2011 2007 2010 2008 2011 Italy 1990 1995 1990 1995 n.a. n.a. Italy 2008 2011 n.a. n.a. 2008 2011 Japan 1992 2001 1992 2001 1997 2001 Jordan 1989 1991 n.a. n.a. 1989 1991 Kazakhstan 2008 2011 n.a. n.a. 2008 2011 Kenya 1985 1989 1985 1989 1985 1985 Kenya 1992 1995 1992 1995 1992 1994 Korea 1983 1983 1983 1983 n.a. n.a. Korea 1985 1988 1985 1988 n.a. n.a. Korea 1997 2002 1997 2002 1997 1998 Madagascar 1988 1988 n.a. n.a. 1988 1988 Malaysia 1985 1988 1985 1988 n.a. n.a. Malaysia 1997 2001 1997 2001 1997 1999 Mali 1987 1991 n.a. n.a. 1987 1991 Mauritania 1984 1984 n.a. n.a. 1984 1984 Mexico 1981 1985 1981 1982 1981 1985 Mexico 1994 2000 1994 2000 1994 1996 Mongolia 2008 2011 n.a. n.a. 2008 2011 Nepal 1988 1988 n.a. n.a. 1988 1988 Netherlands 2008 2011 2008 2010 2008 2011 New Zealand 1987 1990 1987 1990 n.a. n.a. Nicaragua 1987 1996 1987 1996 n.a. n.a. Nicaragua 1990 1993 n.a. n.a. 1990 1993 Nicaragua 2000 2002 2000 2002 2000 2001 Niger 1983 1985 n.a. n.a. 1983 1985 Nigeria 1992 1995 1992 1995 1991 1995 Nigeria 1997 1997 1997 1997 n.a. n.a. Nigeria 2009 2011 n.a. n.a. 2009 2011 Norway 1987 1993 1987 1993 1991 1993 Panama 1988 1989 1988 1989 1988 1989 Paraguay 1995 1999 1995 1999 1995 1995 Paraguay 2002 2002 2002 2002 n.a. n.a. Peru 1983 1990 1983 1990 1983 1983 Peru 1999 1999 1999 1999 n.a. n.a. Philippines 1981 1987 1981 1987 1983 1986 Philippines 1997 2001 1997 2001 1997 2001 Portugal 2008 2011 2008 2010 2008 2011 Senegal 1988 1991 n.a. n.a. 1988 1991 Sierra Leone 1990 1994 n.a. n.a. 1990 1994 Singapore 1982 1983 1982 1983 n.a. n.a. South Africa 1989 1989 1989 1989 n.a. n.a. Spain 1980 1985 1977 1985 1977 1981 Spain 2008 2011 2008 2010 2008 2011 Sri Lanka 1989 1993 1989 1993 1989 1991 Swaziland 1995 1999 n.a. n.a. 1995 1999 Sweden 1991 1995 1991 1994 1991 1995 Sweden 2008 2011 n.a. n.a. 2008 2011 Switzerland 2008 2011 2008 2009 2008 2011 Thailand 1980 1987 1980 1987 1983 1983 Thailand 1996 2001 1996 2001 1997 2000 Togo 1993 1994 n.a. n.a. 1993 1994 Tunisia 1991 1991 n.a. n.a. 1991 1991 Turkey 1982 1985 1982 1985 1982 1984 Turkey 1991 1991 1991 1991 n.a. n.a. Turkey 1994 1994 1994 1994 n.a. n.a. Turkey 2000 2001 2000 2000 2000 2001 UK 1984 1984 1984 1984 n.a. n.a. UK 1991 1991 1991 1991 n.a. n.a. UK 1995 1995 1995 1995 n.a. n.a. UK 2007 2011 2007 2009 2007 2011 USA 1984 1991 1984 1991 1988 1988 USA 2007 2011 2007 2010 2007 2011 Uruguay 1981 1985 n.a. n.a. 1981 1985 Uruguay 2002 2005 n.a. n.a. 2002 2005 Venezuela 1980 1986 1978 1986 n.a. n.a. Venezuela 1993 1998 1993 1994 1994 1998 Country Paper1 R2 LV3 Start End Start End Start End Algeria 1990 1994 1990 1992 1990 1994 Argentina 1980 1982 1980 1982 1980 1982 Argentina 1989 1991 1989 1990 1989 1991 Argentina 1995 1996 1995 1996 1995 1995 Argentina 2001 2003 2001 2003 2001 2003 Australia 1989 1992 1989 1992 n.a. n.a. Austria 2008 2011 2008 2010 2008 2011 Bangladesh 1987 1987 n.a. n.a. 1987 1987 Belgium 2008 2011 2008 2010 2008 2011 Benin 1988 1992 n.a. n.a. 1988 1992 Bolivia 1986 1987 1986 1987 1986 1986 Bolivia 1994 1997 1994 1997 1994 1994 Bolivia 1999 1999 1999 1999 n.a. n.a. Brazil 1985 1985 1985 1985 n.a. n.a. Brazil 1990 1994 1990 1990 1990 1994 Burkina Faso 1990 1994 n.a. n.a. 1990 1994 Burundi 1994 1998 n.a. n.a. 1994 1998 Canada 1983 1985 1983 1985 n.a. n.a. Cape Verde 1993 1993 n.a. n.a. 1993 1993 CAR 1980 1982 1976 1982 1976 1976 CAR 1988 1999 1988 1999 1995 1996 Chad 1983 1983 n.a. n.a. 1983 1983 Chad 1992 1996 n.a. n.a. 1992 1996 Chile 1981 1985 1982 1984 1981 1985 Congo, Dem. Rep. 1983 1983 n.a. n.a. 1983 1983 Congo, Dem. Rep. 1991 1994 n.a. n.a. 1991 1994 Congo, Dem. Rep. 1994 1998 n.a. n.a. 1994 1998 Congo, Rep. 1992 1994 n.a. n.a. 1992 1994 Costa Rica 1987 1991 1987 1987 1987 1991 Costa Rica 1994 1996 1994 1996 1994 1995 Cote dIvoire 1988 1992 1988 1991 1988 1992 Denmark 1987 1992 1987 1992 n.a. n.a. Denmark 2008 2011 2008 2010 2008 2011 Dominican Rep 1996 1996 1996 1996 n.a. n.a. Dominican Rep 2003 2004 2003 2003 2003 2004 Ecuador 1981 1986 1981 1981 1982 1986 Ecuador 1998 2002 1998 2002 1998 2002 Egypt 1980 1983 1981 1983 1980 1980 Egypt 1990 1995 1990 1995 n.a. n.a. El Salvador 1989 1990 1989 1989 1989 1990 Finland 1991 1995 1991 1994 1991 1995 France 1994 1995 1994 1995 n.a. n.a. France 2008 2011 2008 2010 2008 2011 Germany 1991 1994 1991 1994 n.a. n.a. Germany 2008 2011 n.a. n.a. 2008 2011 Ghana 1982 1989 1982 1989 1982 1983 Ghana 1997 1997 1997 1997 n.a. n.a. Greece 1991 1995 1991 1995 n.a. n.a. Greece 2008 2011 2008 2010 2008 2011 Guatemala 1990 1990 1990 1990 n.a. n.a. Guatemala 2001 2001 2001 2001 n.a. n.a. Guatemala 2006 2006 2006 2006 n.a. n.a. Guyana 1993 1993 n.a. n.a. 1993 1993 Honduras 1999 1999 1999 1999 n.a. n.a. Honduras 2001 2002 2001 2002 n.a. n.a. Iceland 1985 1986 1985 1986 n.a. n.a. Iceland 1993 1993 1993 1993 n.a. n.a. Iceland 2008 2011 2007 2010 2008 2011 India 1993 1998 1993 1998 1993 1993 Indonesia 1992 1992 1992 1992 n.a. n.a. Indonesia 1994 1994 1994 1994 n.a. n.a. Indonesia 1997 2002 1997 2002 1997 2001 Ireland 2007 2011 2007 2010 2008 2011 Italy 1990 1995 1990 1995 n.a. n.a. Italy 2008 2011 n.a. n.a. 2008 2011 Japan 1992 2001 1992 2001 1997 2001 Jordan 1989 1991 n.a. n.a. 1989 1991 Kazakhstan 2008 2011 n.a. n.a. 2008 2011 Kenya 1985 1989 1985 1989 1985 1985 Kenya 1992 1995 1992 1995 1992 1994 Korea 1983 1983 1983 1983 n.a. n.a. Korea 1985 1988 1985 1988 n.a. n.a. Korea 1997 2002 1997 2002 1997 1998 Madagascar 1988 1988 n.a. n.a. 1988 1988 Malaysia 1985 1988 1985 1988 n.a. n.a. Malaysia 1997 2001 1997 2001 1997 1999 Mali 1987 1991 n.a. n.a. 1987 1991 Mauritania 1984 1984 n.a. n.a. 1984 1984 Mexico 1981 1985 1981 1982 1981 1985 Mexico 1994 2000 1994 2000 1994 1996 Mongolia 2008 2011 n.a. n.a. 2008 2011 Nepal 1988 1988 n.a. n.a. 1988 1988 Netherlands 2008 2011 2008 2010 2008 2011 New Zealand 1987 1990 1987 1990 n.a. n.a. Nicaragua 1987 1996 1987 1996 n.a. n.a. Nicaragua 1990 1993 n.a. n.a. 1990 1993 Nicaragua 2000 2002 2000 2002 2000 2001 Niger 1983 1985 n.a. n.a. 1983 1985 Nigeria 1992 1995 1992 1995 1991 1995 Nigeria 1997 1997 1997 1997 n.a. n.a. Nigeria 2009 2011 n.a. n.a. 2009 2011 Norway 1987 1993 1987 1993 1991 1993 Panama 1988 1989 1988 1989 1988 1989 Paraguay 1995 1999 1995 1999 1995 1995 Paraguay 2002 2002 2002 2002 n.a. n.a. Peru 1983 1990 1983 1990 1983 1983 Peru 1999 1999 1999 1999 n.a. n.a. Philippines 1981 1987 1981 1987 1983 1986 Philippines 1997 2001 1997 2001 1997 2001 Portugal 2008 2011 2008 2010 2008 2011 Senegal 1988 1991 n.a. n.a. 1988 1991 Sierra Leone 1990 1994 n.a. n.a. 1990 1994 Singapore 1982 1983 1982 1983 n.a. n.a. South Africa 1989 1989 1989 1989 n.a. n.a. Spain 1980 1985 1977 1985 1977 1981 Spain 2008 2011 2008 2010 2008 2011 Sri Lanka 1989 1993 1989 1993 1989 1991 Swaziland 1995 1999 n.a. n.a. 1995 1999 Sweden 1991 1995 1991 1994 1991 1995 Sweden 2008 2011 n.a. n.a. 2008 2011 Switzerland 2008 2011 2008 2009 2008 2011 Thailand 1980 1987 1980 1987 1983 1983 Thailand 1996 2001 1996 2001 1997 2000 Togo 1993 1994 n.a. n.a. 1993 1994 Tunisia 1991 1991 n.a. n.a. 1991 1991 Turkey 1982 1985 1982 1985 1982 1984 Turkey 1991 1991 1991 1991 n.a. n.a. Turkey 1994 1994 1994 1994 n.a. n.a. Turkey 2000 2001 2000 2000 2000 2001 UK 1984 1984 1984 1984 n.a. n.a. UK 1991 1991 1991 1991 n.a. n.a. UK 1995 1995 1995 1995 n.a. n.a. UK 2007 2011 2007 2009 2007 2011 USA 1984 1991 1984 1991 1988 1988 USA 2007 2011 2007 2010 2007 2011 Uruguay 1981 1985 n.a. n.a. 1981 1985 Uruguay 2002 2005 n.a. n.a. 2002 2005 Venezuela 1980 1986 1978 1986 n.a. n.a. Venezuela 1993 1998 1993 1994 1994 1998 1Crisis period used in the paper, equal to R ∪ LV. 2Crisis period according to Reinhart (2010). 3Crisis period according to Laeven and Valencia (2013). View Large 3.2 Methodology To assess indicators’ capacity to predict banking crises, we use the signal detection approach pioneered by Kaminsky and Reinhart (1999), which is one of the most common methods for the statistical evaluation of early warning indicators (EWIs). According to this methodology, when an indicator takes a value that exceeds a certain threshold, this is a signal that the event of interest (in our case a financial crisis) will materialise within the forecast period. Comparing the signal with the actual realisation of the event allows an assessment of the predictive capacity of the indicator for a given threshold. Forecast errors can be divided in two categories: (a) the indicator has not signalled a crisis that actually occurred in the forecast horizon (missed crisis corresponding to Type I error) and (b) the indicator has incorrectly signalled a crisis that will not materialise in the forecast horizon (false alarm corresponding to Type II error). Then, for the assessment of the performance of the indicator on the full range of possible thresholds we use the area under the receiver operating characteristic (ROC) curve. This is a graphical instrument that permits to visualise an indicator’s signalling quality by using the statistical relation between the ‘true positive rate’ (fraction of crises correctly predicted) and the ‘false positive rate’ (fraction of cases incorrectly classified as crises). The first ratio, also called ‘sensitivity’, is an estimate of the probability of a positive signal conditional on the fact that a crisis will materialise, and relates to the test’s ability to identify a condition correctly. The second ratio, also called ‘specificity’, is an estimate of the probability of incorrectly predicting a crisis and relates to the test’s ability to exclude a condition correctly. These two ratios are not independent and are connected to type I and type II errors: the higher is the sensitivity of the indicator, the lower is its specificity, meaning that the indicator will correctly predict most of the crises but will also send many false alarms. The ROC curve synthesises this trade off by reporting the true positive and the false positive rates for any possible threshold (see Figure 3). For very conservative thresholds, we expect that both the positive rate and the false positive rate are high (i.e., the indicator detects all the crises but sends also many false alarms), for lax thresholds instead we expect that both the true and false positive rates are low (i.e., the indicator misses most of the crisis but does not send many false alarms). Figure 3: View largeDownload slide Comparing ROC Curves. The Figure Illustrates the ROC Curves Associated with Three Indicators with Different Forecasting Capacity. The Indicator That Has a ROC Curve Coincident with the Diagonal, with an Area Under the Curve (AUC) of 0.5, Does not Perform Better than Tossing a Coin Because for any Signal Issued by the Indicator There Is a 50:50 Chance of Predicting Correctly the Event of Interest (e.g., the True Positive Rate Is Equal to the False Positive Rate). The more the AUC Is Distant from 0.5 and the Closer to 1, the Higher Is the Forecasting Performance of the Indicator. Figure 3: View largeDownload slide Comparing ROC Curves. The Figure Illustrates the ROC Curves Associated with Three Indicators with Different Forecasting Capacity. The Indicator That Has a ROC Curve Coincident with the Diagonal, with an Area Under the Curve (AUC) of 0.5, Does not Perform Better than Tossing a Coin Because for any Signal Issued by the Indicator There Is a 50:50 Chance of Predicting Correctly the Event of Interest (e.g., the True Positive Rate Is Equal to the False Positive Rate). The more the AUC Is Distant from 0.5 and the Closer to 1, the Higher Is the Forecasting Performance of the Indicator. The area under the ROC curve (AUC) can be used as a convenient and interpretable summary measure of the discriminatory accuracy of an indicator. This varies between 0 and 1, with 0.5 indicating that the indicator is not informative, as for any positive signal the probability that the event of interest will materialise in the forecast horizon is equal to the probability of a false alarm. Indicators that are expected to increase ahead of the crisis have higher predictive performance the higher is the distance of the AUC from 0.5 and the closer to 1. For these indicators, AUC values below 0.5 are associated with classifiers that perform worse than random guessing as higher values in the prediction are associated with a lower probability that the event of interest will materialise. In our case, this would correspond to a situation where higher values of the gaps are associated with a lower probability of a crisis, which is counterintuitive and misleading. 3.3 Results of the analysis In this section, we contrast the predictive performance of the BIS and frontier gaps at different levels of financial development. The measure of financial development adopted hereafter is the private credit-to-GDP ratio at the time of the crisis,8 which relates the extent of financial intermediation to the size of the economy. As in Drehmann and Juselius (2014), the BIS and frontier gaps are evaluated using both a static and a dynamic approach. The static assessment provides information on the statistical forecasting power of the indicators, while the dynamic analysis allows us to verify whether the indicators satisfy additional important requirements, including timeliness and stability of the signal. 3.3.1 Forecasting performance of BIS and frontier gaps over a fixed forecast period In the static assessment, the predictive performance of the BIS and frontier gaps is tested over a fixed forecast window of 1–3 years (i.e., a signal is considered correct if it forecasts a crisis 1–3 years in advance). The use of a forecast interval, rather than a point-wise forecast horizon, permits to address the uncertainty of dating crisis correctly (Drehmann and Juselius, 2014), which is relevant in our analysis because we use two crisis databases adopting different definitions of systemic banking crisis (Table 2). The choice of the forecast window reflects two opposite considerations. On the one hand, forecast intervals need to be sufficiently long so that policy actions can be taken in time to be effective if systemic risk is detected. On the other hand, forecast windows cannot be too large because this would prevent the use of the whole sample, given that the time series of several countries are too short. Accordingly, the lower bound of the interval is set at 1 year, which is sufficiently distant from the crisis to allow policy makers to implement policies and macro-prudential actions to start impact on the policy objective (reduction in credit growth). The higher bound of the interval, instead, is set at 3 years, which allows the use of the full sample of countries. For this forecast period, we derive, for each country in the sample, the ROC curves corresponding to the two indicators and compute the AUC non-parametrically by trapezoidal integration.9 As mentioned in Section 3.2, we use the AUC to rank the forecasting performance of the BIS and frontier gaps in each country. Accordingly, we define indicator I to outperform indicator J in country h over the forecasting period T when the following criterion is satisfied: AUC(Ih,T)>0.5∩AUC(Ih,T)>AUC(Jh,T). (1) This criterion requires that the best performing indicator is informative and has the highest signalling quality. Figure 4 reports the AUC for the BIS and frontier gaps (vertical axis, left and right panels respectively) against the level of financial development (horizontal axis) for individual countries. In each panel, the observations marked with a diamond correspond to countries where the respective indicator outperforms the other based on criterion (1), while the observations marked with a star correspond to countries for which both indicators are uninformative (AUC below or equal to 0.5). Figure 4: View largeDownload slide Predictive Performance of BIS and Frontier Gaps (AUC Metric). The Figure Illustrates the Predictive Performance of the BIS and Frontier Gaps (Left and Right Charts, Respectively), as Measured by the AUC, Against the Level of Financial Development for Individual Countries. In Each Country, the Indicator Is Considered Informative if Its AUC Is Superior to 0.5. The Figure Shows that the Relative Performance of the BIS (frontier) Gap Is Superior to the Performance of the Frontier (BIS) Gap in Countries at an Intermediate (Low and High) Level of Financial Development. Source: IMF IFS, Finstats and Authors’ Calculations. Figure 4: View largeDownload slide Predictive Performance of BIS and Frontier Gaps (AUC Metric). The Figure Illustrates the Predictive Performance of the BIS and Frontier Gaps (Left and Right Charts, Respectively), as Measured by the AUC, Against the Level of Financial Development for Individual Countries. In Each Country, the Indicator Is Considered Informative if Its AUC Is Superior to 0.5. The Figure Shows that the Relative Performance of the BIS (frontier) Gap Is Superior to the Performance of the Frontier (BIS) Gap in Countries at an Intermediate (Low and High) Level of Financial Development. Source: IMF IFS, Finstats and Authors’ Calculations. Figure 5: View largeDownload slide Distributions of the Difference between Median AUC for Frontier and BIS Gaps at different levels of financial development. The Figure Shows that, Based on the Bootstrapped Distributions, the Probability that the Difference between the Median AUC of the BIS and Frontier Gaps Is Null Is Inferior to 10% in all Country Groups, Which Confirms that the Difference in the Performance of the Two Indicators Is Statistically Significant. Source: IMF IFS, Finstats and Authors’ Calculations. Figure 5: View largeDownload slide Distributions of the Difference between Median AUC for Frontier and BIS Gaps at different levels of financial development. The Figure Shows that, Based on the Bootstrapped Distributions, the Probability that the Difference between the Median AUC of the BIS and Frontier Gaps Is Null Is Inferior to 10% in all Country Groups, Which Confirms that the Difference in the Performance of the Two Indicators Is Statistically Significant. Source: IMF IFS, Finstats and Authors’ Calculations. This preliminary graphical analysis permits to identify three main findings: The BIS gap has an inverted U-shape performance across levels of financial development (polynomial fitting in Figure 4, left panel). The indicator shows an average weak forecasting performance in countries at low levels of financial development and a high performance in countries with intermediate and large financial sectors. In countries with an over-sized financial sector, the BIS gap performs well but is dominated by the frontier gap. The frontier gap has a U-shape performance (polynomial fitting in Figure 4, right panel). The indicator has an average strong performance in countries at a low level of financial development and with an over-sized financial sector but it has a relatively weak performance in countries with an intermediate or large financial sector. In a number of countries, both measures of excess credit are a poor predictor of financial crises (observations marked with a star in Figure 4). This implies that, in these countries, financial stress originated from sources different from fast credit growth. It is worth noticing that these observations are concentrated in countries with a low level of financial development. Based on these preliminary findings, we split the countries in our sample into three groups: (a) group 1, which includes countries with a credit-to-GDP ratio up to 60% at the time of the crisis (low level of financial development); (b) group 2, comprising countries with a credit-to-GDP ratio higher than 60% and up to 160% at the time of the crisis (intermediate and high level of financial development); and (iii) group 3, including countries with a credit-to-GDP ratio above 160% at the time of the crisis (over-sized financial sector). The cutoff levels of the credit to GDP ratio for each group are consistent with stylised facts (e.g. low and middle income countries have an average level of credit not superior to 60% of GDP, see Figure 1) and empirical literature (for over-sized financial sectors, see Cecchetti and Kharroubi, 2012; European Systemic Risk Board, 2014). In Table 3, we restrict the focus on the 55 observations for which at least one of the two indicators is informative (i.e., the AUC is superior to 0.5) and we synthesise the predictive performance of the two gaps in each of the three groups with the AUC recorded on median by the countries belonging to the corresponding subsample. In Group 1, the frontier gap outperforms the BIS gap in two thirds of the observations, recording a median area under the curve of 0.76 against a median AUC of 0.63 for the BIS gap. In Group 2, instead, the BIS gap performs better than the frontier gap in 86% of the cases and registers a median AUC of 0.83 against a median value of 0.60 for the frontier gap. Finally, in Group 3, the frontier gap has the best predictive performance, even if both the BIS and frontier gaps have high median AUCs (0.88 against 0.96). Table 3: Forecasting Performance of the BIS and Frontier Gaps over a Fixed Forecast Period across Levels of Financial Development Indicator Median(AUC)1 Standard deviation median (AUC) C.I.2 median(AUC) p-Value3 median(AUC) Best indicator (percent of observations) Sample size Low level of financial development (Group 1) BIS gap 0.63 0.10 [0.39,0.76] 0.16 35.1 37 Frontier Gap 0.76*** 0.06 [0.62,0.85] 0.00 64.9 Intermediate and high level of financial development (Group 2) BIS gap 0.83*** 0.04 [0.74,0.91] 0.00 85.7 14 Frontier Gap 0.60* 0.10 [0.33,0.78] 0.10 14.3 Oversized financial sector (Group 3) BIS gap 0.88** 0.17 [0.26,0.95] 0.05 0.00 4 Frontier Gap 0.96*** 0.02 [0.93,1.00] 0.00 100.0 Indicator Median(AUC)1 Standard deviation median (AUC) C.I.2 median(AUC) p-Value3 median(AUC) Best indicator (percent of observations) Sample size Low level of financial development (Group 1) BIS gap 0.63 0.10 [0.39,0.76] 0.16 35.1 37 Frontier Gap 0.76*** 0.06 [0.62,0.85] 0.00 64.9 Intermediate and high level of financial development (Group 2) BIS gap 0.83*** 0.04 [0.74,0.91] 0.00 85.7 14 Frontier Gap 0.60* 0.10 [0.33,0.78] 0.10 14.3 Oversized financial sector (Group 3) BIS gap 0.88** 0.17 [0.26,0.95] 0.05 0.00 4 Frontier Gap 0.96*** 0.02 [0.93,1.00] 0.00 100.0 The table compares the median AUC of the BIS and frontier gaps recorded in each of the three country group subsamples. Results confirm that the BIS (frontier) gap outperforms in Group 2 (Groups 1 and 3) both in terms of predictive power and statistical significance. 1 *, ** and *** denote statistical significance at the 10, 5 and 1% level, respectively. 2The 95% bootstrap confidence intervals are estimated by the cutoff values for the middle 95% of each bootstrap distribution. 3The bootstrap p-values are computed as the proportion of bootstrap samples where the simulated median AUCs are equal or inferior to 0.5. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5. Source: IMF IFS, Finstats and Authors’ Calculations. View Large Table 3: Forecasting Performance of the BIS and Frontier Gaps over a Fixed Forecast Period across Levels of Financial Development Indicator Median(AUC)1 Standard deviation median (AUC) C.I.2 median(AUC) p-Value3 median(AUC) Best indicator (percent of observations) Sample size Low level of financial development (Group 1) BIS gap 0.63 0.10 [0.39,0.76] 0.16 35.1 37 Frontier Gap 0.76*** 0.06 [0.62,0.85] 0.00 64.9 Intermediate and high level of financial development (Group 2) BIS gap 0.83*** 0.04 [0.74,0.91] 0.00 85.7 14 Frontier Gap 0.60* 0.10 [0.33,0.78] 0.10 14.3 Oversized financial sector (Group 3) BIS gap 0.88** 0.17 [0.26,0.95] 0.05 0.00 4 Frontier Gap 0.96*** 0.02 [0.93,1.00] 0.00 100.0 Indicator Median(AUC)1 Standard deviation median (AUC) C.I.2 median(AUC) p-Value3 median(AUC) Best indicator (percent of observations) Sample size Low level of financial development (Group 1) BIS gap 0.63 0.10 [0.39,0.76] 0.16 35.1 37 Frontier Gap 0.76*** 0.06 [0.62,0.85] 0.00 64.9 Intermediate and high level of financial development (Group 2) BIS gap 0.83*** 0.04 [0.74,0.91] 0.00 85.7 14 Frontier Gap 0.60* 0.10 [0.33,0.78] 0.10 14.3 Oversized financial sector (Group 3) BIS gap 0.88** 0.17 [0.26,0.95] 0.05 0.00 4 Frontier Gap 0.96*** 0.02 [0.93,1.00] 0.00 100.0 The table compares the median AUC of the BIS and frontier gaps recorded in each of the three country group subsamples. Results confirm that the BIS (frontier) gap outperforms in Group 2 (Groups 1 and 3) both in terms of predictive power and statistical significance. 1 *, ** and *** denote statistical significance at the 10, 5 and 1% level, respectively. 2The 95% bootstrap confidence intervals are estimated by the cutoff values for the middle 95% of each bootstrap distribution. 3The bootstrap p-values are computed as the proportion of bootstrap samples where the simulated median AUCs are equal or inferior to 0.5. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5. Source: IMF IFS, Finstats and Authors’ Calculations. View Large To test whether these results are statistically significant, we perform statistical inference on the basis of non-parametric estimation rather than asymptotic normal approximation due to the small sample size.10 Accordingly, we use bootstrap resampling to calculate the standard errors, point-wise confidence intervals (C.I.), and p-values of the median AUC of the BIS and frontier gaps in each country group. The bootstrap algorithm uses 1,000 independent replications, each reflecting the original sample partition among the three country groups. Bootstrap test statistics (C.I. and p-values) confirm that the best indicator in each country group has a statistically significant median AUC (Table 3),11 although results for the third group should be taken with caution as the original sample is too small to ensure enough variation in the bootstrap distributions. Up to this point, the results of our analysis have found that, for the fix forecast window under observation, the frontier gap has superior statistical forecasting power in Groups 1 and 3, while the BIS gap has stronger predictive performance in group 2. However, given that the AUCs’ confidence intervals of the two indicators overlap in each group,12 it is not granted that their performance is statistically different, implying that there might not be a statistically dominant EWI in each group. To address this remaining uncertainty, we complete the assessment in the static approach by testing whether the performance of the two gaps is statistically different in each group. For this purpose, we use the 1,000 bootstrap samples to derive the distribution of the difference between the median AUC of the frontier and BIS gaps (Figure 5) and we test the null hypothesis that this difference is equal to zero in each group. Table 4 reports the results of this analysis. Based on the bootstrap test statistics (C.I. and p-values), the null hypothesis is rejected for all groups (at the 10% significance level for Group 1). Again, results for Group 3 should be taken with caution due to the paucity of observations. Table 4: Difference between Bootstrap Median AUCs across Levels of Financial Development Indicator C.I.1 p-Value2 Difference in median AUC (Group 1) [−0.04,0.37] 0.06* Difference in median AUC (Group 2) [−0.46,−0.06] 0.01*** Difference in median AUC (Group 3) [−0.01,0.74] 0.02** Indicator C.I.1 p-Value2 Difference in median AUC (Group 1) [−0.04,0.37] 0.06* Difference in median AUC (Group 2) [−0.46,−0.06] 0.01*** Difference in median AUC (Group 3) [−0.01,0.74] 0.02** In this table, the null hypothesis that the difference between the median AUC of the frontier and BIS gaps is null is tested. Results confirm that the difference is significant in each country group at a minimum significance level of 10%. 1The 95% bootstrap confidence intervals are estimated by the cutoff values for the middle 95% of each bootstrap distribution. 2The bootstrap p-values for Groups 1 and 3 (Group 2) are computed as the proportion of bootstrap samples where the simulated median AUCs are equal or inferior (superior) to 0. *, ** and *** denote statistical significance at the 10, 5 and 1% level, respectively. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5 for both the frontier and BIS gaps. Source: IMF IFS, Finstats and Authors’ Calculations. View Large Table 4: Difference between Bootstrap Median AUCs across Levels of Financial Development Indicator C.I.1 p-Value2 Difference in median AUC (Group 1) [−0.04,0.37] 0.06* Difference in median AUC (Group 2) [−0.46,−0.06] 0.01*** Difference in median AUC (Group 3) [−0.01,0.74] 0.02** Indicator C.I.1 p-Value2 Difference in median AUC (Group 1) [−0.04,0.37] 0.06* Difference in median AUC (Group 2) [−0.46,−0.06] 0.01*** Difference in median AUC (Group 3) [−0.01,0.74] 0.02** In this table, the null hypothesis that the difference between the median AUC of the frontier and BIS gaps is null is tested. Results confirm that the difference is significant in each country group at a minimum significance level of 10%. 1The 95% bootstrap confidence intervals are estimated by the cutoff values for the middle 95% of each bootstrap distribution. 2The bootstrap p-values for Groups 1 and 3 (Group 2) are computed as the proportion of bootstrap samples where the simulated median AUCs are equal or inferior (superior) to 0. *, ** and *** denote statistical significance at the 10, 5 and 1% level, respectively. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5 for both the frontier and BIS gaps. Source: IMF IFS, Finstats and Authors’ Calculations. View Large For robustness, we conclude the static assessment by verifying whether the results described above (and synthesised in the first column of Table 5) are sensitive to: the cutoff level of the credit to GDP ratio for each group; the choice to assign countries to the three groups based on a specific year and the identification and timing of banking crises. Table 5: Median AUCs across Levels of Financial Development—Robustness Checks Indicator First crisis1 First crisis1 First crisis1 Last crisis2 First crisis1 First crisis1 LV-R3 LV-R3 LV-R3 LV-R3 R4 LV5 Cutoff = 606 Cutoff = 507 Cutoff = 708 Cutoff = 606 Cutoff = 606 Cutoff = 606 Group 1 BIS gap 0.63 0.66 0.63 0.59 0.73* 0.60* Frontier gap 0.76*** 0.78*** 0.76*** 0.78*** 0.83*** 0.74*** Sample size 37 32 39 31 20 30 Group 2 BIS gap 0.83*** 0.78*** 0.82*** 0.80*** 0.82*** 0.79*** Frontier gap 0.60* 0.61 0.60* 0.63* 0.56 0.68** Sample size 14 19 12 16 16 12 Group 3 BIS gap 0.88** 0.88** 0.88** 0.82* 0.88** 0.89 Frontier gap 0.96*** 0.96*** 0.96*** 0.90*** 0.96*** 0.98*** Sample size 4 4 4 8 4 7 Indicator First crisis1 First crisis1 First crisis1 Last crisis2 First crisis1 First crisis1 LV-R3 LV-R3 LV-R3 LV-R3 R4 LV5 Cutoff = 606 Cutoff = 507 Cutoff = 708 Cutoff = 606 Cutoff = 606 Cutoff = 606 Group 1 BIS gap 0.63 0.66 0.63 0.59 0.73* 0.60* Frontier gap 0.76*** 0.78*** 0.76*** 0.78*** 0.83*** 0.74*** Sample size 37 32 39 31 20 30 Group 2 BIS gap 0.83*** 0.78*** 0.82*** 0.80*** 0.82*** 0.79*** Frontier gap 0.60* 0.61 0.60* 0.63* 0.56 0.68** Sample size 14 19 12 16 16 12 Group 3 BIS gap 0.88** 0.88** 0.88** 0.82* 0.88** 0.89 Frontier gap 0.96*** 0.96*** 0.96*** 0.90*** 0.96*** 0.98*** Sample size 4 4 4 8 4 7 In this table, the estimates of the static assessment (first column) are compared with those obtained using different assumptions. Results confirm that the overall thrust of results remain robust. *, ** and *** denote statistical significance at the 10, 5 and 1% level, respectively. 1Countries are classified in the three groups based on their credit-to-GDP ratio at the time of the first crisis. 2Countries are classified in the three groups based on their credit-to-GDP ratio at the time of the last crisis. 3The identification and timing of banking crises are based on the database that merges Reinhart (2010) and Laeven and Valencia (2013). 4The identification and timing of banking crises are based on Reinhart (2010) database. 5The identification and timing of banking crises are based on Laeven and Valencia (2013) database. 6Countries are classified in Group 1 if their credit-to-GDP ratio is up to 60%. 7Countries are classified in Group 1 if their credit-to-GDP ratio is up to 50%. 8Countries are classified in Group 1 if their credit-to-GDP ratio is up to 70%. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5 for both the frontier and BIS gaps. Source: IMF IFS, Finstats and Authors’ Calculations. View Large Table 5: Median AUCs across Levels of Financial Development—Robustness Checks Indicator First crisis1 First crisis1 First crisis1 Last crisis2 First crisis1 First crisis1 LV-R3 LV-R3 LV-R3 LV-R3 R4 LV5 Cutoff = 606 Cutoff = 507 Cutoff = 708 Cutoff = 606 Cutoff = 606 Cutoff = 606 Group 1 BIS gap 0.63 0.66 0.63 0.59 0.73* 0.60* Frontier gap 0.76*** 0.78*** 0.76*** 0.78*** 0.83*** 0.74*** Sample size 37 32 39 31 20 30 Group 2 BIS gap 0.83*** 0.78*** 0.82*** 0.80*** 0.82*** 0.79*** Frontier gap 0.60* 0.61 0.60* 0.63* 0.56 0.68** Sample size 14 19 12 16 16 12 Group 3 BIS gap 0.88** 0.88** 0.88** 0.82* 0.88** 0.89 Frontier gap 0.96*** 0.96*** 0.96*** 0.90*** 0.96*** 0.98*** Sample size 4 4 4 8 4 7 Indicator First crisis1 First crisis1 First crisis1 Last crisis2 First crisis1 First crisis1 LV-R3 LV-R3 LV-R3 LV-R3 R4 LV5 Cutoff = 606 Cutoff = 507 Cutoff = 708 Cutoff = 606 Cutoff = 606 Cutoff = 606 Group 1 BIS gap 0.63 0.66 0.63 0.59 0.73* 0.60* Frontier gap 0.76*** 0.78*** 0.76*** 0.78*** 0.83*** 0.74*** Sample size 37 32 39 31 20 30 Group 2 BIS gap 0.83*** 0.78*** 0.82*** 0.80*** 0.82*** 0.79*** Frontier gap 0.60* 0.61 0.60* 0.63* 0.56 0.68** Sample size 14 19 12 16 16 12 Group 3 BIS gap 0.88** 0.88** 0.88** 0.82* 0.88** 0.89 Frontier gap 0.96*** 0.96*** 0.96*** 0.90*** 0.96*** 0.98*** Sample size 4 4 4 8 4 7 In this table, the estimates of the static assessment (first column) are compared with those obtained using different assumptions. Results confirm that the overall thrust of results remain robust. *, ** and *** denote statistical significance at the 10, 5 and 1% level, respectively. 1Countries are classified in the three groups based on their credit-to-GDP ratio at the time of the first crisis. 2Countries are classified in the three groups based on their credit-to-GDP ratio at the time of the last crisis. 3The identification and timing of banking crises are based on the database that merges Reinhart (2010) and Laeven and Valencia (2013). 4The identification and timing of banking crises are based on Reinhart (2010) database. 5The identification and timing of banking crises are based on Laeven and Valencia (2013) database. 6Countries are classified in Group 1 if their credit-to-GDP ratio is up to 60%. 7Countries are classified in Group 1 if their credit-to-GDP ratio is up to 50%. 8Countries are classified in Group 1 if their credit-to-GDP ratio is up to 70%. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5 for both the frontier and BIS gaps. Source: IMF IFS, Finstats and Authors’ Calculations. View Large Relative to the first point, Table 5 shows that the thrust of the results for Groups 1 and 2 would not change using alternative cutoff levels that are in the proximity of the proposed thresholds (Table 5 illustrates results with cutoffs 50 and 70 in the second and third column, respectively). This result reflects the fact that, while point estimates differ (Table 6), the frontier gap has the highest median performance at any level of the credit-to-GDP ratio up to 60%, while the BIS gap has the best median performance at any level of the credit-to-GDP ratio above 60% and up to 160%. Checking robustness for Group 3 is more difficult given the small sample size and the fact that there are no observations with a credit-to-GDP ratio in the interval]160, 200[%. Table 6: Median AUC across Levels of Financial Development Private credit-to-GDP ratio (%) 0–20 20–40 40–60 60–80 80–100 100–120 120–140 140–160 160–180 180–200 200–220 220–240 Indicator BIS Gap 0.56 0.58 0.72 0.76 0.92 0.78 0.85 0.87 0.86 0.26 Frontier Gap 0.74 0.88 0.87 0.74 0.71 0.40 0.76 0.58 0.98 0.99 Sample Size 11 15 11 4 3 2 2 3 0 0 3 1 Private credit-to-GDP ratio (%) 0–20 20–40 40–60 60–80 80–100 100–120 120–140 140–160 160–180 180–200 200–220 220–240 Indicator BIS Gap 0.56 0.58 0.72 0.76 0.92 0.78 0.85 0.87 0.86 0.26 Frontier Gap 0.74 0.88 0.87 0.74 0.71 0.40 0.76 0.58 0.98 0.99 Sample Size 11 15 11 4 3 2 2 3 0 0 3 1 This table shows the median AUC of the BIS and frontier gaps in countries with different levels of financial development. Results confirm that the frontier gap has the highest median performance at any level of the credit-to-GDP ratio up to 60%, while the BIS gap has the highest performance at any level above 60 and up to 160%. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5 for both the frontier and BIS gaps. Source: IMF IFS, Finstats and Authors’ Calculations. View Large Table 6: Median AUC across Levels of Financial Development Private credit-to-GDP ratio (%) 0–20 20–40 40–60 60–80 80–100 100–120 120–140 140–160 160–180 180–200 200–220 220–240 Indicator BIS Gap 0.56 0.58 0.72 0.76 0.92 0.78 0.85 0.87 0.86 0.26 Frontier Gap 0.74 0.88 0.87 0.74 0.71 0.40 0.76 0.58 0.98 0.99 Sample Size 11 15 11 4 3 2 2 3 0 0 3 1 Private credit-to-GDP ratio (%) 0–20 20–40 40–60 60–80 80–100 100–120 120–140 140–160 160–180 180–200 200–220 220–240 Indicator BIS Gap 0.56 0.58 0.72 0.76 0.92 0.78 0.85 0.87 0.86 0.26 Frontier Gap 0.74 0.88 0.87 0.74 0.71 0.40 0.76 0.58 0.98 0.99 Sample Size 11 15 11 4 3 2 2 3 0 0 3 1 This table shows the median AUC of the BIS and frontier gaps in countries with different levels of financial development. Results confirm that the frontier gap has the highest median performance at any level of the credit-to-GDP ratio up to 60%, while the BIS gap has the highest performance at any level above 60 and up to 160%. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5 for both the frontier and BIS gaps. Source: IMF IFS, Finstats and Authors’ Calculations. View Large Results are also insensitive to the choice to assign countries to the three groups based on a specific year. For instance, Table 5 shows that classifying countries in the three groups based on the level of the credit-to-GDP ratio at the time of the most recent crisis (fourth column) would produce results very similar to those obtained classifying countries based on the value of the ratio at the time of the first crisis (first column). In this case, the robustness of results reflects the fact that, while the credit-to-GDP ratio of individual countries varies over time, only in few cases this variation is large enough to determine a change in the classification of countries from the first crisis to the most recent one. Specifically, in our sample only five countries move from Group 1 to Group 2, three countries move from Group 2 to Group 3 and one country from Group 1 to Group 3. These changes have limited impact on the median performance of the BIS and frontier gaps. Finally, the issue pertaining the uncertainty associated with the identification and timing of banking crises is already partially addressed by using a forecast interval, rather than a point-wise forecast horizon, as discussed at the beginning of this section. As additional robustness check, Table 5 reports the results obtained by using the crisis databases from Reinhart (2010) and from Laeven and Valencia (2013) separately (fifth and sixth column, respectively), rather than merging them in a unified database (first column). Also in this case, the overall thrust of results remains robust: both in terms of predictive power and statistical significance, the frontier gap is the dominant EWI for Group 1 and 3, while the BIS gap outperforms in Group 2. Overall, the outcome of the static assessment provides strong evidence that, for the fix forecast period under consideration, the frontier gap is a statistically superior EWI of financial crises in countries at an early stage of financial development, while the BIS gap outperforms in countries with a medium and high level of financial development. In countries with an over-sized financial sector, the frontier gap seems to be the dominant EWI, even if the size of the sample under observation is too small to perform proper statistical inference. 3.3.2 Dynamic forecasting performance of the BIS and frontier gaps The results in Section 3.3.1 are only indicative of the forecasting performance of the BIS and frontier gaps over a fix forecast window of 3 years but do not provide information on the performance at specific (point-wise) forecast horizons, which is relevant to test the forecasting power over time and assess the timing of the signal and its persistence across horizons. The timing of a signal is relevant because data are reported with lags, particularly in developing countries, and policy makers tend to adjust policies gradually to changes in macroeconomic conditions. In addition, changes in macro-prudential tools may take time to impact on their policy objective (Drehmann and Juselius, 2014; CGFS, 2012). It follows that ideal EWIs should start issuing signals well before a crisis. Also the persistence (or stability) of the signal is an important requirement because EWIs that issue stable signals reduce uncertainty regarding risk buildup, thus allowing for more decisive policy actions. Based on these considerations, in this section we assess the dynamic predictive performance of the BIS and frontier gaps in individual countries based on three criteria. To test the predictive power across time, we use again criterion (1), although applied to individual (point-wise) forecast horizons rather than to a forecast interval. The second criterion pertains to the timeliness of the signal and requires that the indicator starts issuing informative signals at least 3 years in advance: AUC(Ih,τ)>0.5withτ≥3. (2) Finally, the third criterion refers to the stability of the signal and requires that, after issuing the first informative signal, the indicator stays informative over the remaining forecast horizon. For instance, assuming that the first informative signal is issued 3 years before the crisis, the criterion requires that: AUC(Ih,3−j)>0.5∀j∈{1,2}. (3) To test criteria (1) to (3) on our sample, we derive the AUC for the BIS and frontier gaps for five consecutive point-wise horizons for individual countries.13 To have a consistent sample across forecast horizons, we drop six countries because their series of the credit-to-GDP ratio are not sufficiently long to assess the predictive performance of the gaps five years before a crisis. This reduces the sample size from 79 to 73.14 As in the static assessment, Figure 6 and Table 7 restrict the focus on the observations for which at least one of the two indicators is informative (i.e. the AUC is superior to 0.5) and synthesise the results for all forecast horizons and country groups. Figure 6: View largeDownload slide Signalling Quality, as Measured by the AUC, of BIS (Left) and Frontier Gaps (Right) at Different Levels of Financial Development and Different Forecast Periods. The Figure Shows that the Best Indicator in Each Country Group (i.e., the BIS Gap for Group 2 and the Frontier Gap for Groups 1 and 3) Has an Informative and Significant AUC across Projection Horizons. This Confirms that, in Each Group, the Best Indicator Issues Timely and Persistent Signals. Source: IMF IFS, Finstats and Authors’ Calculations. Figure 6: View largeDownload slide Signalling Quality, as Measured by the AUC, of BIS (Left) and Frontier Gaps (Right) at Different Levels of Financial Development and Different Forecast Periods. The Figure Shows that the Best Indicator in Each Country Group (i.e., the BIS Gap for Group 2 and the Frontier Gap for Groups 1 and 3) Has an Informative and Significant AUC across Projection Horizons. This Confirms that, in Each Group, the Best Indicator Issues Timely and Persistent Signals. Source: IMF IFS, Finstats and Authors’ Calculations. Table 7: Dynamic Predictive Performance of BIS and Frontier Gaps across Levels of Financial Development t−1 t−2 t−3 t−4 t−5 Group 1 Bis gap Median (AUC) 0.62 0.64 0.67 0.65 0.66 C.I. high1 0.74 0.77 0.79 0.78 0.76 C.I. low1 0.37 0.38 0.37 0.39 0.43 p-Value2 0.04 0.05 0.08 0.11 0.11 Frontier gap Median (AUC) 0.70*** 0.76*** 0.82*** 0.79*** 0.82*** C.I. high1 0.83 0.84 0.93 0.88 0.88 C.I. low1 0.56 0.54 0.71 0.63 0.63 p-Value2 0.01 0.01 0.00 0.00 0.01 Group 2 Bis gap Median (AUC) 0.85*** 0.82*** 0.81*** 0.75*** 0.72*** C.I. high1 0.95 0.92 0.90 0.84 0.80 C.I. low1 0.69 0.73 0.71 0.65 0.63 p-Value2 0.00 0.00 0.00 0.00 0.00 Frontier gap Median (AUC) 0.61 0.64 0.60 0.52 0.54 C.I. high1 0.77 0.79 0.82 0.73 0.76 C.I. low1 0.40 0.42 0.10 0.34 0.30 p-Value2 0.13 0.11 0.25 0.37 0.11 Group 3 Bis gap Median (AUC) 0.75 0.82* 0.88** 0.89** 0.91** C.I. high1 1.00 0.99 0.95 0.94 0.97 C.I. low1 0.06 0.15 0.26 0.35 0.44 p-Value2 0.20 0.06 0.05 0.04 0.05 Frontier gap Median (AUC) 0.92*** 0.94*** 0.96*** 0.98*** 1.00*** C.I. high1 1.00 1.00 1.00 1.00 1.00 C.I. low1 0.81 0.88 0.90 0.90 0.91 p-Value2 0.00 0.00 0.00 0.00 0.00 t−1 t−2 t−3 t−4 t−5 Group 1 Bis gap Median (AUC) 0.62 0.64 0.67 0.65 0.66 C.I. high1 0.74 0.77 0.79 0.78 0.76 C.I. low1 0.37 0.38 0.37 0.39 0.43 p-Value2 0.04 0.05 0.08 0.11 0.11 Frontier gap Median (AUC) 0.70*** 0.76*** 0.82*** 0.79*** 0.82*** C.I. high1 0.83 0.84 0.93 0.88 0.88 C.I. low1 0.56 0.54 0.71 0.63 0.63 p-Value2 0.01 0.01 0.00 0.00 0.01 Group 2 Bis gap Median (AUC) 0.85*** 0.82*** 0.81*** 0.75*** 0.72*** C.I. high1 0.95 0.92 0.90 0.84 0.80 C.I. low1 0.69 0.73 0.71 0.65 0.63 p-Value2 0.00 0.00 0.00 0.00 0.00 Frontier gap Median (AUC) 0.61 0.64 0.60 0.52 0.54 C.I. high1 0.77 0.79 0.82 0.73 0.76 C.I. low1 0.40 0.42 0.10 0.34 0.30 p-Value2 0.13 0.11 0.25 0.37 0.11 Group 3 Bis gap Median (AUC) 0.75 0.82* 0.88** 0.89** 0.91** C.I. high1 1.00 0.99 0.95 0.94 0.97 C.I. low1 0.06 0.15 0.26 0.35 0.44 p-Value2 0.20 0.06 0.05 0.04 0.05 Frontier gap Median (AUC) 0.92*** 0.94*** 0.96*** 0.98*** 1.00*** C.I. high1 1.00 1.00 1.00 1.00 1.00 C.I. low1 0.81 0.88 0.90 0.90 0.91 p-Value2 0.00 0.00 0.00 0.00 0.00 This table shows the median AUC of the BIS and frontier gaps at different levels of financial development across five forecast horizons. Result confirms that the BIS (frontier) gap issues timely and persistent signals in Group 2 (Groups 1 and 3). 1The 95% bootstrap confidence intervals are estimated by the cutoff values for the middle 95% of each bootstrap distribution. 2The bootstrap p-Values are computed as the proportion of bootstrap samples where the simulated median AUCs are equal or inferior to 0.5. *, ** and *** denote statistical significance at the 10, 5 and 1% level, respectively. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5 for both the frontier and BIS gaps. Source: IMF IFS, Finstats and Authors’ Calculations. View Large Table 7: Dynamic Predictive Performance of BIS and Frontier Gaps across Levels of Financial Development t−1 t−2 t−3 t−4 t−5 Group 1 Bis gap Median (AUC) 0.62 0.64 0.67 0.65 0.66 C.I. high1 0.74 0.77 0.79 0.78 0.76 C.I. low1 0.37 0.38 0.37 0.39 0.43 p-Value2 0.04 0.05 0.08 0.11 0.11 Frontier gap Median (AUC) 0.70*** 0.76*** 0.82*** 0.79*** 0.82*** C.I. high1 0.83 0.84 0.93 0.88 0.88 C.I. low1 0.56 0.54 0.71 0.63 0.63 p-Value2 0.01 0.01 0.00 0.00 0.01 Group 2 Bis gap Median (AUC) 0.85*** 0.82*** 0.81*** 0.75*** 0.72*** C.I. high1 0.95 0.92 0.90 0.84 0.80 C.I. low1 0.69 0.73 0.71 0.65 0.63 p-Value2 0.00 0.00 0.00 0.00 0.00 Frontier gap Median (AUC) 0.61 0.64 0.60 0.52 0.54 C.I. high1 0.77 0.79 0.82 0.73 0.76 C.I. low1 0.40 0.42 0.10 0.34 0.30 p-Value2 0.13 0.11 0.25 0.37 0.11 Group 3 Bis gap Median (AUC) 0.75 0.82* 0.88** 0.89** 0.91** C.I. high1 1.00 0.99 0.95 0.94 0.97 C.I. low1 0.06 0.15 0.26 0.35 0.44 p-Value2 0.20 0.06 0.05 0.04 0.05 Frontier gap Median (AUC) 0.92*** 0.94*** 0.96*** 0.98*** 1.00*** C.I. high1 1.00 1.00 1.00 1.00 1.00 C.I. low1 0.81 0.88 0.90 0.90 0.91 p-Value2 0.00 0.00 0.00 0.00 0.00 t−1 t−2 t−3 t−4 t−5 Group 1 Bis gap Median (AUC) 0.62 0.64 0.67 0.65 0.66 C.I. high1 0.74 0.77 0.79 0.78 0.76 C.I. low1 0.37 0.38 0.37 0.39 0.43 p-Value2 0.04 0.05 0.08 0.11 0.11 Frontier gap Median (AUC) 0.70*** 0.76*** 0.82*** 0.79*** 0.82*** C.I. high1 0.83 0.84 0.93 0.88 0.88 C.I. low1 0.56 0.54 0.71 0.63 0.63 p-Value2 0.01 0.01 0.00 0.00 0.01 Group 2 Bis gap Median (AUC) 0.85*** 0.82*** 0.81*** 0.75*** 0.72*** C.I. high1 0.95 0.92 0.90 0.84 0.80 C.I. low1 0.69 0.73 0.71 0.65 0.63 p-Value2 0.00 0.00 0.00 0.00 0.00 Frontier gap Median (AUC) 0.61 0.64 0.60 0.52 0.54 C.I. high1 0.77 0.79 0.82 0.73 0.76 C.I. low1 0.40 0.42 0.10 0.34 0.30 p-Value2 0.13 0.11 0.25 0.37 0.11 Group 3 Bis gap Median (AUC) 0.75 0.82* 0.88** 0.89** 0.91** C.I. high1 1.00 0.99 0.95 0.94 0.97 C.I. low1 0.06 0.15 0.26 0.35 0.44 p-Value2 0.20 0.06 0.05 0.04 0.05 Frontier gap Median (AUC) 0.92*** 0.94*** 0.96*** 0.98*** 1.00*** C.I. high1 1.00 1.00 1.00 1.00 1.00 C.I. low1 0.81 0.88 0.90 0.90 0.91 p-Value2 0.00 0.00 0.00 0.00 0.00 This table shows the median AUC of the BIS and frontier gaps at different levels of financial development across five forecast horizons. Result confirms that the BIS (frontier) gap issues timely and persistent signals in Group 2 (Groups 1 and 3). 1The 95% bootstrap confidence intervals are estimated by the cutoff values for the middle 95% of each bootstrap distribution. 2The bootstrap p-Values are computed as the proportion of bootstrap samples where the simulated median AUCs are equal or inferior to 0.5. *, ** and *** denote statistical significance at the 10, 5 and 1% level, respectively. Note: Statistics in the table are computed excluding observations corresponding to countries for which the area under the ROC curve is equal to or below 0.5 for both the frontier and BIS gaps. Source: IMF IFS, Finstats and Authors’ Calculations. View Large These results permit to derive three main findings pertaining the dynamic performance of the BIS and frontier gaps: In terms of predictive power, as defined by criterion (1), the BIS gap dominates the frontier gap at all horizons in countries with an intermediate and high level of financial development (Group 2), while the frontier gap outperforms in countries at an early stage of financial development (Group 1) and with an over-sized financial sector (Group 3). In terms of timing, the BIS and frontier gaps satisfy criterion (2) in all groups in the sample under observation. However, considering only results that are statistically significant, the BIS gap satisfies criterion (2) only in Groups 2 and 3, while the frontier gap in Groups 1 and 3. Pertaining to stability, the BIS and frontier gaps satisfy criterion (3) in all groups in the sample under observation. However, if we limit the analysis to statistically significant results, the BIS gap issues stable signals only in Group 2 and the frontier gap in Groups 1 and 3. In summary, the results in this section show that BIS (frontier) gap has the highest dynamic forecasting performance in Group 2 (Group 1) based on all criteria: predictive power, timeliness, and stability. Results for Group 3 suggest that the frontier gap has the best dynamic performance but the size of the sample under observation is too small to perform proper statistical inference and draw firm conclusions. 3.3.3 Overall assessment The evidence in Sections 3.3.1 and 3.3.2 provides strong support to the idea that the assessment of systemic risk buildup associated with excess credit should be performed with different EWIs at different levels of financial development. In countries at an early stage of financial development, the frontier gap outperforms the BIS gap in both the static and the dynamic assessments. In this country group, the frontier gap starts issuing informative signals 5 years ahead of a crisis and records a consistently high predictive power across forecasting horizons (with the AUC fluctuating between 0.70 and 0.82 over the 5 years). In addition, results are strongly significant for all forecasting horizons. The predictive power of the BIS gap is remarkably lower and median AUC estimates are not significant. This result is likely to reflect higher capacity of the frontier gap to discriminate financial deepening from systemic risk buildup compared to the BIS gap, which is relevant in developing countries. A healthy process of financial deepening, indeed, is associated with a situation where the frontier gap is negative and the acceleration of credit reflects a process of catch up with respect to the financial possibility frontier. This process would be incorrectly signalled as systemic risk buildup by the BIS gap but not by the frontier gap. In countries at an intermediate and high level of financial development, the results are inverted: the BIS gap has the highest static and dynamic predictive performances. These results are in line with previous findings in the literature according which the BIS gap outperforms other indicators as predictor of banking crises in advanced economies and emerging markets (Drehmann and Juselius, 2014; Drehmann and Tsatsaronis, 2014; Detken et al., 2014). This might reflect the fact that in countries with a developed financial sector the risks of additional financial deepening exceed the benefits. Finally, in countries with over-sized financial sectors (Group 3), the frontier gap seems to be the most powerful predictor but results should be taken with caution as they are based on a very small sample. A few caveats are worth mentioning when assessing the results for Group 1. First, the relation between credit and GDP should be expected to be less stable in developing countries than in other economies due to large output volatility. Drops in GDP, which are not mirrored by a fall in credit to the private sector of the same size, would result in an increase of both the BIS and frontier gaps that might, therefore, erroneously signal overheating.15 In addition, data limitations are extensive in developing countries. Quarterly series of credit and GDP, commonly used to test the performance of the BIS gap, are available only for a limited number of developing countries. Also, in those countries the series of credit and GDP may be marred by data gaps and structural breaks (due to civil wars, changes in the compilation of statistics and large swings in exchange rates) that can significantly impact the measurement of the credit-to-GDP ratio (Drehmann and Tsatsaronis, 2014). It must also be noted that the benchmark estimates used in this paper, and derived from FinStats, are based on a limited number of structural variables, while a vast array of other relevant factors, that might be critical to the development of the financial system, are excluded. 4. Conclusion We have provided evidence that systemic risk assessment based on measures of excess credit should be performed with different indicators at different stages of financial development. Measures that compare the level of credit with its historical trend are good proxies for systemic risk buildup in countries at an advanced stage of financial development. Instead, indicators that contrast actual credit to a benchmark level based on a country’s structural characteristics work better for countries with an underdeveloped financial sector. This fills a void in the EWS literature, as the existing analytical work on this subject has mainly tested the predictive performance of leading indicators of financial crises in advanced economies, where crises are expected to present common patterns and similarities. The findings in this paper also shed some light on the relationship between credit, financial stability, and deepening. In particular, In countries at an early stage of financial development, credit expansion is mainly positively associated with economic growth (financial deepening) (Rajan and Zingales, 1998; Levine et al., 2000; Levine, 2005). Systemic risk may, however, arise when credit expansion exceeds the absorption capacity of the economy, as determined by its structural characteristics. In countries with a developed financial system, on the other hand, rapid credit growth is frequently associated with excessive risk taking, weakening of lending standards, and the likelihood of a systemic banking crisis. In these countries, an acceleration of credit growth above its secular trend is likely to reflect systemic risk buildup rather than additional financial deepening. This explains the higher predictive performance of the BIS gap. This result is also in line with the empirical finding showing that booms that start at a higher level of financial depth are more likely to turn into a financial crisis (Dell’Ariccia et al., 2012; Meng and Gonzalez, 2017). While results for countries with an over-sized financial sector are not conclusive, the size of the financial sector seems to be a source of systemic risk per se, independent from acceleration in credit growth. This may reflect the fact that large financial systems are likely to be accompanied by riskier business models, larger interconnectedness, and higher vulnerability to global shocks. In addition, larger financial systems tend to have more systemically important banks that, on average, create more individual and systemic risk than smaller banks (Laeven et al., 2014). We also found that for some developing countries systemic risk buildup stems from sources other than excess credit. This paper showed, indeed, that in these countries banking crises may occur even if credit is not excessive by any measure. This may reflect the fact that banking sectors in developing countries are frequently more fragile than in advanced economies. Thus, there remains a need for the development of additional leading indicators and early warning models that are specific for those countries and focus on sources of vulnerability that are different from excess credit. The identification of these risk factors is important also in the case of crises preceded by excessive credit expansion, as they could heighten the severity of financial disruption connected to the bursting of the credit bubble. We leave a systematic examination of this issue for future research. In future work, it might also be interesting to explore whether, in developing countries, the assessment of systemic risk buildup based on the frontier gap should be complemented with signals from the BIS gap. In fact, while in our analysis the frontier gap showed a higher median predictive performance in countries at an early stage of financial development, the BIS gap outperformed the frontier gap in several cases (35% in the static assessment, Table 3). This may reflect the fact that despite financial development being below capacity (negative frontier gap), the acceleration of credit was too fast (strongly positive BIS gap) and led to the buildup of vulnerabilities, particularly in the absence of strong supervision. Finally, this paper has not addressed the issue of determining the threshold levels of the BIS and frontier gaps that may pose significant risks to financial stability in developing countries. Previous empirical research on this subject is based on the evidence of advanced economies and large emerging markets, or on large samples including countries at different levels of financial development. Two factors, pulling in opposite directions, should be taken into consideration when extending the analysis to developing countries. On the one side, financial deepening could justify higher BIS gap thresholds than in advanced economies. On the other, financial systems and supervisory and regulatory frameworks are frequently more weaker in developing countries. In these circumstances, even small deviations of the credit-to-GDP ratio from its long-term trend, or from the frontier, can be sufficient to spur a banking crisis. This issue also remains open for future research. References Al-Hussainy E. , Coppola A. , Feyen E. , Ize A. , Kibuuka K. , Ren H. ( 2011 ) FinStats2011: A Ready-to-Use Tool to Benchmark Financial Sectors Across Countries and Over Time . Washington : World Bank . Arcand J. , Berkes E. , Panizza U. ( 2015 ) ‘ Too Much Finance? ’, Journal of Economic Growth , 20 ( 2 ): 105 – 48 . Google Scholar CrossRef Search ADS Arena M. , Bouza S. , Dabla-Norris E. , Gerling K. , Njie L. ( 2015 ) Credit Booms andMacroeconomic Dynamics: Stylized Facts and Lessons for Low-Income Countries, IMF Working Papers, 15/11. Barajas A. , Beck T. , Dabla-Norris E. , Yousefi S. R. ( 2013 ) Too Cold, Too Hot, Or Just Right? Assessing Financial Sector Development Across theGlobe, IMF Working Papers, 13/81. Barajas A. , Dell’Ariccia G. , Levchenko A. ( 2007 ) Credit Booms: The Good, the Bad, and the Ugly. Unpublished manuscript, International Monetary Fund. Basel Committee on Banking Supervision ( 2010 ) Guidance for National Authorities Operating the Countercyclical Capital Buffer . Basel : Basel Committee on Banking Supervision . Beck T. , Feyen E. , Ize A. , Moizeszowicz F. ( 2008 ). Benchmarking Financial Development, World Bank Policy Research Working Paper, No. 4638. Bernanke B. , Gertler M. , Gilchrist S. ( 1999 ) ‘The financial accelerator in a quantitative business cycle framework’, in Taylor J. , Woodford M. (eds) , Handbook of Macroeconomics , Vol. 1C . Amsterdam : Elsevier , pp. 1531 – 1614 . Borio C. ( 2012 ) The financial cycle and macroeconomics: What have we learnt?, BIS Working Papers, No 395. Borio C. , Drehmann M. ( 2009 ) ‘ Assessing the Risk of Banking Crises—Revisited ’, BIS Quarterly Review , March : 29 – 46 . Borio C. , Lowe P. ( 2002 ) ‘ Assessing the Risk of Banking Crises ’, BIS Quarterly Review , December : 43 – 54 . Cecchetti S. , Kharroubi E. ( 2012 ) Reassessing the impact of finance on growth, BIS Working Papers, No. 381. Committee on the Global Financial System (CGFS) ( 2012 ) Operationalising the selection and application of macroprudential instruments, CGFS Publications, No 48. Cottarelli C. , Dell’Ariccia G. , Vladkova-Hollar I. ( 2005 ) ‘ Early Birds, Late Risers, and Sleeping Beauties: Bank Credit Growth to the Private Sector in Central and Eastern Europe and in the Balkans ’, Journal of Banking & Finance , 29 ( 1 ): 83 – 104 . Google Scholar CrossRef Search ADS Crowe C. , Dell’Ariccia G. , Igan D. , Rabanal P. ( 2013 ) ‘ How to Deal with Real Estate Booms: Lessons from Country Experiences ’, Journal of Financial Stability , 9 : 300 – 19 . Google Scholar CrossRef Search ADS Dembiermont C. , Drehmann M. , Muksakunratana S. ( 2013 ) ‘ How Much Does the Private Sector Really Borrow—A New Database for Total Credit to the Private Non-Financial Sector ’, BIS Quarterly Review , March : 65 – 81 . Dell’Ariccia G. , Igan D. , Laeven L. , Tong H. ( 2012 ). Policies for Macrofinancial Stability: Options to Deal with Credit Booms, IMF Staff Discussion Note, 12/06. Demirgüç-Kunt A. , Detragiache E. , Gupta P. ( 2006 ) ‘ Inside the Crisis: An Empirical Analysis of Banking Systems in Distress ’, Journal of International Money and Finance , 25 ( No 5 ): 702 – 18 . Google Scholar CrossRef Search ADS Detken C. , Weeken O. , Alessi L. , Bonfim D. , Boucinha M. M. , Castro C. , Frontczak S. , Giordana G. , Giese J. , Jahn N. , Kakes J. , Klaus B. , Lang J. H. , Puzanova N. , Welz P. ( 2014 ). Operationalising the countercyclical capital buffer: indicator selection, threshold identification and calibration options, European Systemic Risk Board Occasional Paper Series, No 5, June 2014. Drehmann M. , Borio C. , Gambacorta L. , Jimenez G. , Trucharte C. ( 2010 ). Countercyclical capital buffers: Exploring options, BIS Working Paper Series, No 317. Drehmann M. , Borio C. , Tsatsaronis K. ( 2011 ). Anchoring countercyclical capital buffers: the role of credit aggregates, BIS Working Paper Series, No 355. Drehmann M. , Juselius M. ( 2014 ) ‘ Evaluating Early Warning Indicators of Banking Crises: Satisfying Policy Requirements ’, International Journal of Forecasting , Elsevier, 30 ( 3 ): 759 – 80 . Google Scholar CrossRef Search ADS Drehmann M. , Tsatsaronis K. ( 2014 ) ‘ The Credit-to-GDP Gap and Countercyclical Capital Buffers: Question and Answers ’, BIS Quarterly Review , March : 55 – 73 . Edge R. , Meisenzahl R. ( 2011 ) ‘ The Unreliability of Credit-to-GDP Ratio Gaps in Real-Time: Implications for Countercyclical Capital Buffers ’, International Journal of Central Banking , December : 261 – 98 . European Systemic Risk Board . 2014 . IsEurope Overbanked?, Reports of the Advisory Scientific Committee, No 4, June. Feyen E. , Kibuuka K. , Sourrouille D. ( 2015 ) FinStats 2015: A Ready-to-Use Tool to Benchmark Financial Sectors across Countries and over Time User Guide and Benchmarking Methodology . Washington : World Bank . April. Gambacorta L. ( 2009 ) ‘ Monetary Policy and the Risk-Taking Channel ’, BIS Qarterly Review , December : 43 – 53 . Geršl A. , Seidler J. ( 2012 ) ‘ Excessive Credit Growth and Countercyclical Capital Buffers in Basel III, an Empirical Evidence from Central and East European Countries ’, ACTA VSFS Economic Studies and Analyses , University of Finance and Administration, 6 ( 2 ): 91 – 107 . Gilchrist S. , Ortiz A. , Zakrajsek E. ( 2009 ) Credit Risk and the Macroeconomy: Evidence from an Estimated DSGE. Unpublished manuscript. Boston University. Gourinchas P. , Obstfeld M. ( 2012 ) ‘ Stories of the Twentieth Century for the Twenty First ’, American Economic Journal: Macroeconomics , Volume 4 ( No.1 ): 226 – 65 . Google Scholar CrossRef Search ADS Gourinchas P. , Valdes R. , Landerretche O. ( 2001 ) ‘ Lending Booms: Latin America and the World ’, Economia , 1 ( 2 ): 47 – 99 . Hodrick R. , Prescott E. C. ( 1997 ) ‘ Postwar U.S. Business Cycles: An Empirical Investigation ’, Journal of Money, Credit, and Banking , 29 ( 1 ): 1 – 16 . Google Scholar CrossRef Search ADS International Monetary Fund ( 2011 ) Towards Operationalizing Macroprudential policy: When to Act?, Global Financial Stability Report, September, Washington: IMF. International Monetary Fund ( 2014 ) Staff Guidance Note on Macroprudential Policy—Considerations for Low Income Countries . Washington : IMF . Janes H. , Longton G. , Pepe M. ( 2009 ) ‘ Accommodating Covariates in ROC analysis ’, Stata Journal , 9 ( 1 ): 17 – 39 . Google Scholar PubMed Jorda O. , Schularick M. , Taylor A. ( 2011 ) ‘ Financial Crises, Credit Booms, and External Imbalances: 140 Years of Lessons ’, IMF Economic Review , 59 : 340 – 78 . Google Scholar CrossRef Search ADS Kaminsky G. , Reinhart C. ( 1999 ) ‘ The Twin Crises: The Causes of Banking and Balance-of-Payments Problems ’, American Economic Review , 89 : 473 – 500 . Google Scholar CrossRef Search ADS Kiyotaki N. , Moore J. ( 1997 ) ‘ Credit Cycles ’, Journal of Political Economy , 105 ( 2 ): 211 – 48 . Google Scholar CrossRef Search ADS Laeven L. , Ratnovski L. , Tong H. 2014 . Bank Size and Systemic Risk. IMF Staff Discussion Notes, 14/04. Laeven L. , Valencia F. ( 2013 ) ‘ Systemic Banking Crises Database ’, IMF Economic Review , 61 : 225 – 70 . Google Scholar CrossRef Search ADS Levine R. ( 2005 ) ‘Finance and growth: theory and evidence’, in Aghion P. , Durlauf S. N. (eds) , Handbook of Economic Growth . Amsterdam : Elsevier , pp. 865 – 934 . Levine R. E. , Loayza N. V. , Beck T. ( 2000 ) ‘ Financial Intermediation and Growth: Causality and Causes ’, Journal of Monetary Economics , 16 ( 1 ): 31 – 77 . Google Scholar CrossRef Search ADS Mendoza E. G. , Terrones M. E. 2008 . An Anatomy of Credit Booms: Evidence from Macro Aggregates and Micro Data, NBER Working Papers, No 14049. Meng C. , Gonzalez R. L. ( 2017 ) ‘ Credit Booms in Developing Countries: Are they Different from Those in Advanced Economies and Emerging Market Countries? ’, Open Economies Review , No 28 : 547 – 79 . Google Scholar CrossRef Search ADS Rajan R. G. , Zingales L. ( 1998 ) ‘ Financial Dependence and Growth ’, The American Economic Review , 88 ( 3 ): 559 – 86 . Ravn M. O. , Uhlig H. ( 2002 ) ‘ On Adjusting the Hodrick-Prescott Filter for the Frequency of Observations ’, Review of Economics and Statistics , 84 ( 2 ): 371 – 5 . Google Scholar CrossRef Search ADS Reinhart C. 2010 . This time is different chartbook: country histories on debt, default, and financial crises, NBER Working Papers, No 15815. Reinhart C. , Rogoff K. ( 2009 ) This Time is Different: Eight Centuries of Financial Folly . Princeton : Princeton University Press . Sanusi L. S. 2010 . Global Financial Meltdown and the Reforms in the Nigerian Banking Sector, Speech at the Public Lecture delivered at the Convocation Square, Abubakar Tafawa Balewa University , Bauchi . Schularick M. , Taylor A. M. ( 2012 ) ‘ Credit Booms Gone Bust: Monetary Policy, Leverage Cycles, and Financial Crises, 1870–2008 ’, American Economic Review , 102 ( 2 ): 1029 – 61 . Google Scholar CrossRef Search ADS World Bank . 2010 . Comments on the consultative document countercyclical capital buffer proposal. Footnotes 1 The consolidation was forced by the introduction of stricter capital requirements. The Central Bank of Nigeria announced on 2 July 2004 that banks would be required to achieve a minimum capital level of Naira 25 billion up from Naira 2 billion, by 31 December 2005. 2 Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Greece, Iceland, India, Indonesia, Ireland, Italy, Japan, Korea, Mexico, Netherlands, Portugal, Singapore, South Africa, Spain, Sweden, Switzerland, Thailand, UK and the USA. 3 BIS data cover total credit, including also cross-border finance to the private non-financial sector (Dembiermont et al., 2013). 4 The indicator covers only domestic credit, thus omitting credit booms driven by cross-border finance. According to Drehmann and Tsatsaronis (2014), credit gaps based on total credit outperform those based on narrower measures of credit. The discrepancy is, however, expected to be larger in countries with market-based financial systems and open capital accounts. For those financial sectors, we mainly use the BIS series of total credit. In developing economies, credit to the private non-financial sector is mainly extended by domestic banks and total domestic credit may thus be considered a reliable proxy for total credit in those countries. 5 This implies that to estimate the value of the long-term trend in a certain year the filter uses only the information up to that year, even if information for the following years is available. This allows reprising the operative environment of supervisors and policy makers, who must decide when to activate the countercyclical capital buffer without knowing the future values of the credit-to-GDP ratio. Edge and Meisenzahl (2011) observed that the real-time estimates of the gap are not reliable because they suffer of the end-of-sample problem, meaning that each point estimate might be revised when new information becomes available. Drehmann and Tsatsaronis (2014), however, showed that for policy-relevant horizons the gap computed using a trend derived from a double-sided HP filter performs much worse than the gap obtained from a backward-looking one-sided HP filter. 6 For business cycle analysis, the standard smoothing parameter for quarterly data is 1,600 (Hodrick and Prescott, 1997). To adjust λ when the frequency of the data is different from quarterly, Ravn and Uhlig (2002) multiplied λ with the fourth power of the frequency ratio. Thus, in case of annual data the parameter λ should be set at (1/4)4∗1600=6.25. This value, however, would be inappropriate for the analysis of the financial cycle that is on average much longer than the business cycle (Drehmann et al., 2010). Assuming that credit cycles are four times as long as business cycles, we adopt a value of λ equal to 44∗6.25=1600. 7 In Laeven and Valencia (2013), policy interventions in the banking sector are considered significant if at least three out of the following six measures have been used: (a) deposit freezes and/or bank holidays, (b) significant bank nationalisations, (c) bank restructuring gross costs corresponding at least to 3% of GDP, (d) extensive liquidity support (5% of deposits and liabilities to nonresidents), (e) significant guarantees put in place and (f) significant asset purchases (at least 5% of GDP). 8 For countries with multiple occurrences of banking crisis, the credit-to-GDP ratio used corresponds to the first crisis. 9 For each series, we do not consider the signals issued during crisis periods because there is evidence (Demirgüç-Kunt et al., 2006; Laeven and Valencia, 2013) that credit-to-GDP ratios do not always fall during crises episodes. This might happen when the decline in GDP growth due to the crisis is higher than the slowdown in lending because of preexisting credit lines that are drawn upon during the crisis. This statistical increase in the credit-to-GDP during crisis episodes, due to a contraction in GDP rather than to an expansion in credit, may induce noise in the credit-to-GDP gap indicator that might incorrectly signal another future crisis (false alarm). 10 Non-parametric estimation allows also to address the complication arising from the fact that the data are likely to be correlated over time within individual countries (Janes et al., 2009). 11 In the paper, a result deemed significant indicates at least 5% significance if not otherwise indicated. 12 For instance, in Group 1, the C.I. of the BIS gap's AUC is [0.39, 0.76], while the C.I. of the frontier gap's AUC is [0.62, 0.85]. 13 The computation of the AUC differs from the static procedure in that the signal is calculated for any of the five forecasting horizons separately rather than over an interval. This entails that for any point-wise forecast horizon τ, a signal is considered correct if it forecasts a crisis τ years in advance. It is worth noticing that, while the point-wise forecast analysis allows assessing the timing and stability of the signals issued by EWIs, it may lead to inaccurate AUC estimates when there is uncertainty on crisis dates. 14 The six countries are Brazil, Cape Verde, Kazakhstan, Mauritania, Niger and Singapore. 15 World Bank (2010) notices that in low- and middle-income countries, the application of the rule for the countercyclical buffer based on the BIS gap may result in erratic activation and deactivation of the buffer due to higher volatility of credit and GDP growth in those countries. © The Author(s) 2018. Published by Oxford University Press on behalf of the Centre for the Study of African Economies, all rights reserved. For Permissions, please email: journals.permissions@oup.com This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices)

Journal

Journal of African EconomiesOxford University Press

Published: Apr 3, 2018

There are no references for this article.

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create lists to
organize your research

Export lists, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off