Early Warning Systems
Nordic banking crises
Introduction
Financial crises, such as the current global financial crisis, bear large direct and indirect costs in the form of bankruptcy, increased unemployment, and a worldwide loss of welfare. The costs of the banking crises of 1992/1993 in the Nordic countries were according to Caprio and Klingebiel (1996) for Finland 8 %, for Norway 4 % and for Sweden 6.4 % of GDP. According to Reinhart and Rogoff (2009) a serious financial crisis leads to three major consequences:
1. Asset market collapses that are deep and prolonged. Real housing prices decrease on average by 35 percent in a period of 6 years. Equity prices decrease by 56 percent over a period of three and a half years;
2. Unemployment increases by 7 percentage-points during the down phase of the cycle. Output falls from peak to trough more than 9 percent on average (in two years).
3. Government debt tend to increase by 86 percent.
Predicting these type of crises might help to prevent them or at least decrease their intensity and thereby their welfare costs. This is the reason that so called Early Warning Systems (EWS) are constructed to serve as a monitoring tool to predict crises.
The International Monetary Fund (IMF) uses an Early Warning System to monitor currency crises but has no explicit EWS for banking crises; see Davis and Karim (2008). The aim of this thesis is to develop a EWS for banking crises with, instead of the
The focus of the thesis is on developing an EWS for a set of developed countries. There are differences between the financial systems of developing and developed countries. By constructing an EWS model using data of developed countries we expect better results when predicting future crises in developed countries. Therefore the Nordic banking crises from the early 1990s is chosen as an example. These are developed countries that went to a banking crisis recently. We therefore expect great similarities between the financial system then and now. This will be an advantage when using the model to predict for example the current global crisis.
The banking industries in the Nordic countries, as in several other industrialized countries, underwent considerable change in the 1980s; see Drees and Pazarbaşioğlu (1998). There were economic deregulation, removal of cross-border restrictions on capital flows, financial innovation, and increased competition in financial services. With the blurring distinctions between the different types of financial intermediaries and financial fragility after a credit boom, these conditions led to a banking crises in Finland, Norway and Sweden
We explain how an EWS is developed prior to presenting our results. Since the main goal of an EWS is to predict crises in-sample and out-of-sample predictions are crucial. In-sample the model is tested to see if it would have predicted the crises in the Nordic countries. Out-of-sample the model is used to make predictions on data of Japan and the United States.
2. Literature review
2.1 Nordic banking crises
In the 1980s the banking industries in the Nordic countries, as in other industrial countries, changed drastically. It was a period marked by economic deregulation1, the removal of cross-border restrictions on capital flows, financial innovation, and increased competition in financial services; see Drees and Pazarbaşioğlu (1998). The key
differences between types of financial intermediaries became increasingly vague. This was accompanied by a credit boom. The ability of households to obtain credit had been constrained due to regulation. Because of the deregulation, and low real after-tax interest rates2 households started to borrow aggressively and reduced their savings. The increase in credit led to a jump in asset prices, and in particular real estate prices. Because the borrowers expected the upward trend in asset prices to continue, they were willing to incur heavy debt burdens because of the perceived upside potential. So there was a substantial stock-adjustment to more speculative and risky finance; see Minsky (1977). This led to an increase in the domestic credit to GDP ratio in all three the countries; see Figures 1 to 3, where the domestic credit to GDP ratio rises from the time of the
deregulations in the early 1980s till the time of the crises early 1990s.
Debt financing has its pros and cons. One of the pros is the tax advantage of debt financing. Another one is the leverage effect, where debt is issued because it is assumed to earn a greater return than the cost of interest. Therefore the debt will earn its own interest costs and some additional profits. The disadvantage of debt is that it can become a burden. If the interest costs exceed the revenues the debt will accumulate and if this can not be turned around will lead to bankruptcy.
1
See Table 5, 6 and 7 in Drees and Pazarbaşioğlu (1998) for a chronological overview of selected liberalization measures in Sweden, Finland and Norway.
2 The marginal tax rates were high, and there was full tax deductibility of interest payments in all three
Figure 1 1 2 3 4 1970q1 1980q1 1990q1 2000q1 Finland d o me st ic cre d it t o G D P ra ti o Time Graphs by Country Figure 2 2 2 .5 3 1970q1 1980q1 1990q1 2000q1 Norway d o me st ic cre d it t o G D P ra ti o Time Graphs by Country Figure 3 1 .5 2 2 .5 3 1980q1 1985q1 1990q1 1995q1 2000q1 Sweden d o me st ic cre d it t o G D P ra ti o Time Graphs by Country
This credit boom was followed by period of financial fragility due to lower asset quality, and lower interest margins weakening banks’ balance sheets. In Finland, Norway and Sweden this led to a significant increase in bank loan losses. Because of the thin (low) capitalization of banks the financial position of the banking system consequently was adversely affected. Apart from the losses on real estate loans there were also financial problems in other sectors. In Norway credit exposure to the primary retail and service sectors caused problems. In Sweden the loans backed by commercial real estate and in Finland the loans denominated in foreign currency were problematic. Although the banking crises in the Nordic countries were preceded by financial deregulation, this does not necessary mean that this is the sole cause. There were economic problems,
causes of the banking crises were the delayed policy responses, the structural characteristics of the financial systems, and – last but not least – banks’ inadequate internal risk-management controls.
The banking crises in Finland, Norway and Sweden show that, if economic incentives are distorted by for example deregulations, a negative shock may jeopardize the financial stability of a country. The lack of strict banking supervision coupled with a lack of expected government intervention removed the incentive for the market to impose
discipline on weak banks in times of an economic boom. This led to excessive increase in bank lending which resulted in a loss of efficiency in capital allocation. Instead of
increased savings people responded to the lifting of credit rationing by incurring debt. These debt burdens turned out to be unsustainable.
2.2 An overview of Early Warning Systems
The literature on Early Warning Systems is broad and diverse. As stated by Berg et al. (2004), the specification of EWS models involves a number of decisions that, while guided in some way by economic theory, are largely empirical and judgmental by nature. An example is the definition of a banking crisis. Financial crises are not precisely defined events, but the models must nonetheless define crisis dates in terms of the data. Therefore the chosen definition has a direct effect on the crises data and thus on the estimation of EWS.
Table 1 Overview characteristics reviewed models
Author Type of crises Definition/ dating crisis Indicators Method Time span
prediction Lestano et al. (2003) Currency crises, Banking crises, Debt crises
Kaminsky and Reinhart (1999) and correspondence with central banks, country reports and various financial publications
Divided into four major groups: -External
-Financial
Domestic (real and public) -Global -Binominal multivariate qualitative response approach -reduce information set into limited number of factors using factor analysis
Zero to four quarters Bussiere and Fratzscher (2006) Focus primarily on currency crises
Define currency crisis as the event when the exchange market pressure variable is two standard deviations or more above its country average EMP
-REER overvaluation -Current account (%GDP) -Short-term debt/reserves -Real GDP growth rate -Domestic credit to private and government sector (level and growth rate)
-Financial interdependence
Multinomial logit model with three regimes (a tranquil regime, a pre-crisis regime, and post-crisis/recovery regime) 12 months Demirgüç-Kunt and Detragiache (1997)
Banking crises At least one of following four conditions has to hold:
1 .ratio of nonperforming assets to total assets in banking system exceeded 10% 2. Cost of rescue operation was at least 2 % of GDP
3. Banking sector problems resulted in large scale nationalization of banks 4. extensive bank runs took place or emergency measures such as deposit freezes, prolonged bank holidays, or generalized deposit guarantees were enacted by the government in response to the crisis
-Growth -Tot change -Depreciation -real interest rate -inflation Surplus/GDP -M2/reserves -Private/GDP Cash/bank -Credit growth Deposit insurance -Law and order
Multivariate logit model
Kaminsky and Reinhart (1999) Twin crises: banking crises and balance-of-payments crises
Mark the beginning of a banking crisis by two events: (1) bank runs that lead to the closure, merging or takeover by the public sector of one or more financial institutions; and (2) if the are no runs, the closure, merging, takeover, or large-scale government assistance of an important financial institution (or group of institutions) that marks the start of a string of similar outcomes for other financial institutions.
Also date when the banking crisis hits its peak, defined as the period with the heaviest government intervention and/or bank closures.
Financial liberalization: -M2 multiplier
-domestic credit/ GDP -real interest rate
-lending-deposit rate ratio Other financial: -excess M1 balances -M2/ reserves -Bank deposits Current account: -exports -imports -terms of trade
-real exchange rate deviation Capital account:
-reserves
-real interest-rate differential Real sector:
-output -stock prices Fiscal:
Early Warning Systems are constructed for different types of financial crises. There are EWS models constructed for banking crises (for example Demirgüç-Kunt and
Detragiache, 2000), currency crises (Kaminsky et al., 1998), debt crises (Marchesi, 2003) or a combination of those (Lestano et al. 2003).
The main difference in the various early warning systems for financial crises lies in the type of econometric EWS model that is used. The most popular model is the qualitative response (logit or probit) model as used by for example Demirgüç-Kunt and Detragiache (1997). Other models used in the literature are cross-country regression models with dummy variables, see Sachs et al. (1996); graphical event studies as suggested by Eichengreen et al. (1995) and the signal extraction approach, used by Kaminsky and Reinhart (1999). According to Davis and Karim (2008) the logit model is the most appropriate approach for a global EWS and the signal extraction model for a country specific EWS.
When constructing an EWS model an important choice is the definition/ dating of the crisis. Boyd et al. (2009) distinguish four classifications when it comes to dating banking crises. These four classifications are all updates, modifications and/or expansions of the classification of banking crises first compiled by Caprio and Klingebiel (1996, 1999). Caprio and Klingebiel’s classification relies upon the assessment of a variety of finance professionals in pulling together characterizations of factors that have caused crises. The first definition is due to Demirgüç-Kunt and Detragiache (2002, 2005), the second one is compiled by Caprio et al. (2005), the third is compiled by Reinhart and Rogoff (2008) and based on Kaminsky and Reinhart (1999) and the fourth is constructed by Laeven and Valencia (2008). All four classifications are primarily constructed on the basis of information about government actions undertaken in response to banking distress obtained from bank regulators and/or central banks3. The crisis definition has a direct influence on the dependent variable in the model. The dependent variable is usually a dummy variable that takes on value 1 if there is a crisis in a certain country in a certain
timeframe, and 0 otherwise. The definition of a crisis has a direct influence when the dummy variable takes the values 1 or 0.
Early Warning Systems furthermore differ in the time horizon of the predictions. With the time horizon of the prediction there is a trade-off between predicting a crisis well in advance and the chance of false alarms. When a crisis is predicted far in advance there is a lot of time to react. However the signals will be weak or not existing. Predicting a crisis for example one month before it takes place is easier because the signals will be strong as a country is almost in crisis. With such a short time horizon there is no time to react. Therefore there is the trade-off between time to react and the correct prediction of an upcoming crisis.
Table 2 continued
3. Methodology
3.1 Data
Quarterly data from Norway, Sweden and Finland from 1975 to 2006 is used. The choice of these three countries is largely influenced by the fact that the goal of the thesis is to create an Early Warning System for banking crises in developed countries. Another factor that came into account is the recent nature of the crises in these countries. One of the advantages of using recent crises for the estimations is the availability of data.
Another advantage is the similarity in financials system then and now, which will benefit the prediction of contemporaneous crises in a similar financial climate. Most of the data is from the International Financial Statistics service of the International Monetary Fund; the rest from the Advance/DataStream.
3.2 Dependent variable
As dependent variable a multinomial dummy variable is created as in Bussiere and Fratzscher (2006), which takes on the following values:
Multinomial crisis variable: Value 0 in tranquil times Value 1 pre-crisis
Value 2 crisis/ post crisis
The advantage of such a multinomial crisis variable, as mention in Bussiere and Fratzscher is that in this way the problem of the post-crisis bias is overcome. This bias arises when no distinction is made between tranquil periods, when economic
fact the macroeconomic indicators are still adjusting from the crisis toward their steady state values.
The starting dates of the three crises are taken from Kaminsky and Reinhart (1999). They mark the start of banking crises by events that point at (i) bank runs that lead to closure, merger or takeovers by the public sector of one or more financial institutions, or (ii) a large-scale government bail-out of one or more financial institutions (possibly followed by more bail-outs). For the duration of the crises table A1 from Boyd et al. (2009) is used. In that table Boyd et al. make an overview of starting dates and duration of systemic banking crises as used in five different papers. The duration time adapted by most of these papers is used for our multinomial crisis variable.
As for the time horizon of predictions four quarters is chosen. This means that the four quarters before the crisis and the quarter where the crisis began, take on value 1. During the crisis and in the recovery period the dummy variable takes on value 2. All the other periods are seen as tranquil times and take on value 0.
The pre-crisis and crisis/ post-crisis dates are:
For Finland: 1990q3 – 1991q3 pre-crisis; 1991q4 – 1994q4 crisis/ post-crisis
For Norway: 1987q4 – 1988q4 pre-crisis; 1989q1 – 1994q4 crisis/ post-crisis
3.3 Explanatory variables
As explanatory variables variables are used that according to theory have an effect on the occurrence of a banking crisis as mentioned in Table 1 in the paper by Lestano et al. (2003). These variables are retrieved if proper data was available4:
Discount rate (IFS-60);
CA / GDP, the seasonally adjusted moving average of the yearly current account balance as a percentage of GDP; Source OECD Economic Outlook;
(1 year) change in bank assets (IFS-21 plus IFS-22a to IFS-22f) to bank reserves (IFS-20) ratio;
(1 year) change in CPI (IFS-64);
(1 year) change in money (IFS-35L for Norway and Sweden and data of Finland form the Bank of Finland) to international reserves (IFS-1L.D) ratio;
(1 year) change in real effective exchange rate, source Main Economic Indicators, OECD;
(1 year) change in terms of trade, unit value exports (IFS-74.D) divided by unit value imports (IFS-75.D);
(1 year) change in domestic credit (IFS-32) to GDP (IFS-99B) ratio, and the
(1 year) change in GDP (IFS-99B) per capita (IFS-99Z).
Apart from the Current Account to GDP ratio and the discount rate, all the data are transformed to obtain the quarterly and annual change of the variables. This annual change is the percentage change in the level of the variable with respect to its level four quarters earlier. Kaminsky et al. (1998) argue that this filtering of the data by using the 12-month percentage change ensures that the units are comparable across countries and that the transformed variables are stationary, with well-defined moments, and free from seasonal effects.
The correlation coefficients between these explanatory variables (both for the growth rates on a quarterly and annual basis) are shown in the Tables 3 and 4.
Table 3 Correlation explanatory variables, quarterly change
bank infl mon ex tot cre gro cagdp discou~e
bank 1.000 infl 0.004 1.000 mon 0.091 0.004 1.000 ex -0.043 0.022 -0.089 1.000 tot 0.014 0.025 0.004 0.142 1.000 cre 0.107 0.034 -0.011 -0.022 -0.100 1.000 gro -0.119 0.125 0.158 0.043 0.126 -0.817 1.000 cagdp 0.0167 -0.468 -0.093 0.113 0.151 -0.023 -0.041 1.000 discountrate -0.038 0.506 0.014 0.005 -0.024 0.012 0.047 -0.361 1.000
Table 4 Correlation explanatory variables, yearly change
bank4 infl4 mon4 ex4 tot4 cre4 gro4 cagdp discou~e
bank4 1.000 infl4 -0.020 1.000 mon4 0.089 0.027 1.000 ex4 -0.100 0.078 -0.167 1.000 tot4 -0.010 0.003 -0.022 0.162 1.000 cre4 -0.039 0.116 0.227 0.070 -0.072 1.000 gro4 0.003 0.584 -0.127 0.171 0.436 -0.121 1.000 cagdp 0.043 -0.540 -0.140 0.051 0.254 -0.110 -0.192 1.000 discountrate -0.101 0.583 -0.037 0.121 -0.014 0.050 0.175 -0.351 1.000 Where bank is the change in bank balance, infl is the change in the CPI, mon is the change in the M2 to GDP ratio, ex is the change in the real effective exchange rate, tot is the change in terms of trade, cre is the change in domestic credit to GDP ratio, gro is the change in GDP per capita, and cagdp is the current account to GDP ratio. A 4 behind a variable denotes a 4 quarter/ 1 year change variable.
4 Estimation and testing of the EWS
We estimate panel data of the Nordic countries and use a multinomial logit model and STATA-application GLLAMM.
4.1 Description of the models
4.1.1 The multinomial logit model
The multinomial logistic model is a discrete choice model and so comparable to the logistic regression model. In the case of the logistic regression model the dependent variable is binomial, so it takes on value 0 or 1. With the multinomial logistic model the dependent variable can take on more discrete values. So the (binomial) logistic regression can be seen as a restricted multinomial logistic model. The discrete values in the
multinomial logistic regression are said to be “unordered categorial values”, so they have no order; see Hilbe (2009, p.385). For example they could take on value 0=blue, 1=green and 2=yellow. In the model used in this thesis, the dependent variable takes on the values: 0 for tranquil periods, 1 for pre-crisis periods and 2 for crisis/ post-crisis periods.
We estimate the model using the econometric package STATA. STATA does not
recognize panels of multinomial logistic regressions, so the data is automatically pooled. Therefore it is not possible to account for fixed or random effects. Random effects can be included using the GLLAMM model, developed by Rabe-Hesketh as part of joint work with Skrondal and Pickles5. With GLLAMM it is not possible to estimate a fixed effects model.
4.1.2 The GLLAMM model
The GLLAMM (Generalized Linear Latent and Mixed Models) can, among other things, be used to analyze multinomial logistic panel models. This can be done by transforming the data. The data is expanded as follows6:
where the variable choice is the multinomial dependent variable; in our case crisis.
Response is the same as the original crisis variable. Altern give the alternative values of
the dependent variable, so in our case the three possible crisis scenario’s (no crisis; pre-crisis; crisis/ post-crisis). Selected is an indicator for the response that was given. So it takes on value 1 if the response is equal to the alternative and 0 otherwise. In this way the data is expanded by a factor three. That is the reason that the number of observations used in the GLLAMM estimations is 3 times larger than in the multinomial logit
regressions. Instead of the variable crisis as in the multinomial logit model, the variable
altern is used as the dependent variable. The variables response and selected are also
used in the model estimation, but as an "expanded option" in the syntax of the model.
The results of the estimations of both the multinomial logit and the random effects GLLAMM models are displayed in the Table 5.
Table 5 Estimation results of the multinomial logit and GLLAMM models (1) (2) (3) (4) multinomial logit multinomial logit, 1 year change GLLAMM GLLAMM, 1 year change
Pre-crisis period Yi,t=1
CA / GDP -0.353*** (-3.67) -0.810*** (-2.73) -0.393*** (-3.34) -0.872*** (-2.68) change in bank assets to
bank reserves ratio
0.132 (0.20) -0.394 (-0.40) 0.0101 (0.01) -0.406 (-0.40) change in CPI -89.10** (-2.35) 46.84** (2.01) -100.6*** (-2.58) 49.82* (1.93) change in money to
international reserves ratio
1.461 (0.64) -4.278 (-1.30) 1.880 (0.81) -3.859 (-1.16) change in real effective
exchange rate 12.86 (0.86) 22.45 (1.33) 16.92 (1.17) 26.54 (1.44) change in terms of trade 4.969
(0.57) 37.02*** (3.18) 2.609 (0.25) 38.67*** (3.10) change in domestic credit
to GDP ratio -24.49** (-2.57) -25.43** (-2.24) -24.64** (-2.47) -29.24** (-2.34) change in GDP per capita -25.99**
(-2.33) -119.8*** (-3.82) -24.02** (-2.19) -123.4*** (-3.61) Discount rate 0.683*** (4.34) 0.925*** (2.79) 0.819*** (4.15) 1.029*** (2.82) Constant -7.555*** (-4.97) -8.649** (-2.19) -9.109*** (-4.51) -9.909** (-2.26)
Crisis/ Post-crisis period
Yi,t=2 CA / GDP -0.114** (-2.44) -0.352*** (-3.19) -0.154*** (-3.13) -0.535*** (-3.55) change in bank assets to
bank reserves ratio
-0.0775 (-0.16) 0.334 (0.67) -0.0746 (-0.16) 0.568 (1.11) change in CPI -162.8*** (-5.08) -128.8*** (-3.63) -160.2*** (-4.94) -109.7*** (-2.79) change in money to
international reserves ratio
-0.572 (-0.41) -4.914*** (-3.11) -0.711 (-0.50) -5.066*** (-2.74) change in real effective
exchange rate -16.20*** (-2.72) -20.68*** (-3.85) -18.04*** (-2.93) -21.98*** (-3.71) change in terms of trade 0.729
(0.14) 18.74*** (3.35) 0.818 (0.16) 21.08*** (3.49) change in domestic credit
to GDP ratio -27.69*** (-4.39) -24.11*** (-4.72) -27.39*** (-4.36) -32.55*** (-4.55) Change in GDP per capita -27.37***
(-3.84) -65.00*** (-3.80) -27.93*** (-3.86) -76.53*** (-4.02) Discount rate 0.448*** (4.72) 1.228*** (5.06) 0.358*** (3.57) 0.753*** (2.78) Constant -3.105*** (-4.43) -2.872** (-2.23) -2.327*** (-3.06) 0.717 (0.41) Observations 322 316 966 948 Pseudo R2 0.341 0.709 AIC 297.4 149.7 301.0 151.4 BIC 372.9 224.8 413.0 263.0 t statistics in parentheses * p < 0.10, ** p < 0.05, *** p < 0.01
When comparing the multinomial logit estimations and the GLLAMM estimation results, the sign of the variables that have a significant impact are all the same. There are some differences in significance levels, but not too much. This implies that the taking into account of random effects does not have a decisive impact on the estimation results.
Considering the pseudo R2 the multinomial logit model with the 1 year change has a higher pseudo R2 (0.709 compared to 0.341) and thus a better fit. For the GLLAMM models de pseudo R2 is not available. The Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC) show lower values for the models with the better fit. Both the multinomial logit and GLLAMM model have a better fit with the 1 year change in the variables. The overall differences between the two type of models with respect to the AIC and BIC is small.
Since it is not possible to make predictions with the GLLAMM model when the data is expanded, the multinomial logit model will be used in the remainder of the analysis. An import aspect of an early warning system is the predictability of the model and if the differences in estimation between the two types of models are that small, the lack of random effects will not cause a problem.
4.2 Significance and signs
When looking at the results of the model in the second column of Table 3, the
multinomial logit model with 1 year change variables, most variables are significant, With logistic regressions, as opposed to linear regressions, the value of the coefficient does not tell anything about the size of the effect of marginal changes in the independent variables on the dependent variable. In this thesis we are only interested in the signs of the explanatory variables.
Pre-crisis period Yi,t=1
other hand, increase the chance of a banking crisis. Furthermore does a positive change domestic credit to GDP ratio and GDP per capita all decrease the chance that a banking crisis occurs.
Most of the variables have the expected signs, but not all. A positive change in terms of trade means that the unit value of exports increases compared to the unit value of imports. This would normally be associated with a strengthening of the real economy and
therefore decrease the probability of a banking crisis.
Another remarkable result is the negative sign of change in the domestic credit to GDP
ratio. One would expect the increase of domestic credit to GDP would stimulate and
propagate banking crises.
Post-crisis period Yi,t=2
During the post-crisis period the signs of a couple of variables change: inflation now has a negative sign as does the change in the real effective exchange rate. These results are intuitive. Inflation normally decreases in times of a crisis, and the same holds for the real effective exchange rate. Furthermore the GDP per capita is negative, as is the current account to GDP ratio, the money to international reserves ratio, and the change in domestic credit to GDP ratio. These variables all have the expected signs in the crisis/ post-crisis periods.
4.3 Predictions
observation (a particular country in a particular quarter) is categorized as to whether it is an alarm (i.e. whether the predicted probability is above the cutoff threshold) and also according to whether it is an actual pre-crisis quarter.
The choice of the cut-off level is a trade-off between number of pre-crises quarters correctly called and the number of false alarms. If the cut-off level is low, more quarters are called as pre-crisis quarters. This increases the chance that a pre-crisis quarter is correctly seen as such but will also lead to more false alarms and vice versa. If, for example, a cut-off level of 0 is chosen, the predictions values for all quarters are above 0 and all quarters are signaled as pre-crisis periods. Therefore all pre-crisis periods are correctly called, but the rest consists of false signals. With a cut-off level of 1 there are no crisis periods called and therefore there are no false alarms, however there are no pre-crisis periods correctly called. With all values between 0 and 1 there is a trade-off
between pre-crisis periods correctly called and the number of false alarms
As Berg et al. mention a threshold probability for an alarm can be chosen to minimize a “loss function” equal to the weighted sum of false alarms (as a share of total tranquil periods) and missed crises (as a share of total crisis periods). In their paper, equal weight is placed on the share of alarms that are false and the share of crises that are missed. (The former might be thought of as Type 1 errors and the latter as Type 2 errors, if the null hypothesis is of no crisis.) A higher weight on missed crises would imply a lower cut-off threshold for calling a crisis, and the model would generate both fewer missed crises and more false alarms. Note that only in-sample information can be used to calculate a threshold for actual forecasting purposes. When using the model to produce out-of-sample predictions, however, there is no guarantee that another threshold would not provide a better goodness-of-fit.
An important goal of this thesis is to produce out-of-sample predictions to see if for example the current global financial crisis could be predicted. Since the cut-off level with the best fit for the Nordic countries does not necessary provide the best fit for out of sample test, the cut-off level of 0.5 is arbitrarily chosen.
In-Sample Predictions with cut-off 0.5:
Table 6 Annual change variables
Si,t=0 Si,t=1 Si,t=2 Total
Yi,t=0 248 0 6 254
Yi,t=1 4 9 0 13
Yi,t=2 10 0 39 49
Total 262 9 45 316
Table 7 Quarterly change variables
Si,t=0 Si,t=1 Si,t=2 Total
Yi,t=0 250 2 6 258
Yi,t=1 11 2 2 15
Yi,t=2 29 0 20 49
Total 290 4 28 322
Where Si,t=0 , Si,t=1, Si,t=2 are respectively the number of quarters predicted as tranquil, pre-crisis, and crisis/ post crisis. Yi,t=0 , Yi,t=1, Yi,t=2 are the states of the quarters as observed in the data.
The performance of the multinomial logit model with the 1 year change variables is considerably better. 9 of the 13 (~69%) crisis quarters are predicted and there are no false signals where a crisis is predicted. Furthermore there are only 6 out of 254 (~2%)
quarters where a post-crisis period is predicted instead of a tranquil period. And 10 out of 49 (~20%) times were predicted as tranquil periods, while they were in fact in a post-crisis period.
The performance of the multinomial logit model with quarterly change variables is bad. Only 2 out of the 15 (~13%) crisis quarters are correctly predicted and more than half (~59%) of the post-crisis quarters are predicted as tranquil times.
This confirms the results of the pseudo R2, AIC and BIC that the model with the 1 year change variables has a better fit.
To test whether the 4 quarter change in variables yields the best results, in-sample
4.4 Transformations
4.4.1 Transformations
Several transformations of the variables were operated to check the robustness of the results. One transformation was the addition of lags to see whether that would have an effect on the sign of the domestic credit parameter. It turned out that with a lag 8, so a lag of two years, the growth in domestic credit to GDP variable has a positive sign and is significant. Another transformation was replacing the change in domestic credit to GDP variable with the level of domestic credit to GDP variable. This variable also has a positive sign and is significant, so it shows that a high domestic credit to GDP ratio has a positive influence on the occurrence of a banking crisis. The increase of domestic credit does not immediately have an effect on the occurrence of a crisis. Only when a part of this credit turns into bad loans because they can not be repaid, the probability of an occurrence of a banking crisis increases. But this effect does not occur until repayment is due. At this point the level of domestic credit is still high so that explains the positive sign associated with the level of domestic credit to GDP.
Another transformation was the replacement of the change in bank assets to bank
reserves variable with the level of bank assets to bank reserves ratio variable. In some
papers, see for example Lestano et al. (2003), the bank assets to bank reserve ratio is used instead of the change in that variable. The results of these estimations are shown in Table 8., where model (1) is the 1 year change multinomial logit, which is used as a benchmark. Models (2) and (3) are the models with a lagged domestic credit to GDP variable, and models (4) and (5) are the models where the level of the domestic credit to GDP ratio is used.
The table shows that, as a result of the transformations, there is a change of the sign of the domestic credit variable from a negative to a positive one. The signs of the other explanatory variables remain the same and there is some change in the significance levels of the variables that are significant in the benchmark model. Because of the
change in the real effective exchange rate only in case of the transformation with the lagged variable. The transformation of the bank variable does not turn out to be significant in any of the estimations.
The interpretation of the lagged domestic credit variable is that a rise in the domestic credit to GDP ratio has a positive effect on the occurrence of a banking crisis two years before any of the other variables signal a crisis, as shown in models (2) and (3) in Table 8. During the pre-crisis period the rise in the domestic credit ratio has a negative effect on the occurrence of a banking crisis as shown in benchmark model (1) in Table 8. When looking solely at the level of the domestic credit to GDP ratio, displayed in regressions (4) and (5) in Table 6, one can see that the height of this ratio has a positive effect on the occurrence of a banking crisis. So there is a higher probability if the domestic credit to GDP ratio had a sharp increase two years before the signaling of a crisis. And a high ratio has a positive effect on the occurrence of a banking crisis as well. This shows that the domestic credit rises a few years before a crisis; in the pre-crisis period the domestic credit ratio already decreases but is at a high level. This explains why the lagged change variable has a positive effect, the change variable a negative one and the level variable a positive one.
The same transformations that were made with the domestic credit variable were also done for the terms of trade variable. As with domestic credit, the sign changes after transformation. In case of the terms of trade from a positive one to the more intuitive negative one, for both the lagged transformation as the level transformation. Only in case of the lagged transformation the terms of trade variable was significant. One can
conclude that, as with domestic credit, changes in the terms of trade do have a two-year in-advance-impact on the probability of a banking crisis. The size of the terms of trade ratio however does not have a significant effect.
Table 8 – Models with transformed variables
(1) (2) (3) (4) (5)
Benchmark model Lagged model A Lagged model B Level model A Level model B
1 CA / GDP -0.810*** (-2.73) -0.625** (-2.34) -0.669** (-2.45) -0.379** (-2.46) -0.415** (-2.38) Discount rate 0.925*** (2.79) 1.008** (2.35) 0.994** (2.38) 0.511** (2.25) 0.533** (2.39) 1year-change in CPI 46.84** (2.01) 102.4** (2.28) 99.92** (2.29) 55.61** (2.38) 61.94** (2.27) 1year-change in money to
international reserves ratio
-4.278 (-1.30) -12.76** (-2.54) -13.19** (-2.31) -7.162** (-2.54) -8.504** (-2.36) 1year-change in real effective
exchange rate 22.45 (1.33) 51.59* (1.86) 51.41* (1.91) 18.04 (1.21) 23.20 (1.40) 1year-change in terms of trade 37.02*** (3.18) 36.91*** (2.71) 39.03*** (2.74) 20.76** (2.32) 22.65** (2.33) 1year-change in GDP per capita -119.8*** (-3.82) -162.9*** (-2.84) -166.8*** (-2.81) -90.08*** (-3.34) -97.93*** (-3.05) 1year-change in bank assets
to bank reserves ratio
-0.394 (-0.40) 0.727 (0.61) 0.0907 (0.10) 1year-change in domestic credit to GDP ratio -25.43** (-2.24) Lag 8; 1year-change in
domestic credit to GDP ratio
14.75** (2.40)
14.29** (2.46) bank assets to bank reserves
ratio
0.00411 (0.30)
0.00653 (0.86)
domestic credit to GDP ratio 2.457**
1year-change in money to international reserves ratio
-4.914*** (-3.11) -5.910*** (-4.02) -6.313*** (-4.11) -4.876*** (-3.60) -5.233*** (-3.73) 1year-change in real effective
exchange rate -20.68*** (-3.85) -20.77*** (-4.45) -20.87*** (-4.52) -22.24*** (-4.48) -21.43*** (-4.51) 1year-change in terms of trade 18.74*** (3.35) 15.65*** (3.41) 17.95*** (3.52) 13.53*** (3.21) 15.65*** (3.43) 1year-change in GDP per capita -65.00*** (-3.80) -40.57*** (-3.19) -45.69*** (-3.24) -34.07*** (-2.84) -39.74*** (-3.09) 1year-change in bank assets
to bank reserves ratio
0.334 (0.67) 0.263 (0.65) 0.168 (0.42) 1year-change in domestic credit to GDP ratio -24.11*** (-4.72) Lag 8; 1year-change in
domestic credit to GDP ratio
-7.760** (-1.99)
-8.646** (-2.09) bank assets to bank reserves
ratio
0.00975*** (2.73)
0.00863** (2.51)
domestic credit to GDP ratio 0.324
(0.55) -0.375 (-0.68) Constant -2.872** (-2.23) -3.273*** (-2.61) -3.491*** (-2.87) -4.160** (-1.98) -2.748 (-1.44) Observations 316 298 298 316 316 Pseudo R2 0.709 0.653 0.671 0.624 0.642 AIC 149.7 163.2 156.8 181.6 174.9 BIC 224.8 237.1 230.7 256.7 250.1 t statistics in parentheses * p < 0.10, ** p < 0.05, *** p < 0.01
4.4.2 Predictions with the models using the transformed variables
To test whether the transformations to the variables increase the performance of the model, in-sample predictions were made. As before the cut-off level of 0.5 was chosen. Tables 9 to 12 present the results.
Table 9 Lagged model A
Si,t=0 Si,t=1 Si,t=2 Total
Yi,t=0 232 1 7 240
Yi,t=1 4 9 0 13
Yi,t=2 11 0 34 45
Total 247 10 41 298
Table 10 Lagged model B
Si,t=0 Si,t=1 Si,t=2 Total
Yi,t=0 230 1 9 240
Yi,t=1 4 9 0 13
Yi,t=2 14 0 31 45
Total 248 10 40 298
Table 11 Level model A
Si,t=0 Si,t=1 Si,t=2 Total
Yi,t=0 240 2 12 254
Yi,t=1 7 6 0 13
Yi,t=2 17 0 32 49
Total 264 8 44 316
Table 12 Level model B
Si,t=0 Si,t=1 Si,t=2 Total
Yi,t=0 245 2 7 254
Yi,t=1 7 6 0 13
Yi,t=2 12 0 37 49
Total 264 8 44 316
Where Si,t=0 , Si,t=1, Si,t=2 are respectively the number of quarters predicted as tranquil, pre-crisis, and crisis/ post crisis. Yi,t=0 , Yi,t=1, Yi,t=2 are the states of the quarters as observed in the data.
The lag transformation outperforms the level transformation when it comes to
predictability. Compared to the benchmark estimation, there is a small difference. 9 out of 13 pre-crisis periods are predicted in all three models. With the transformed domestic credit variable however, there is also a false pre-crisis signal.
4.5 Out-of-sample predictions
In-sample the model performs rather well, most of the pre-crisis quarters are predicted and there are not so many false signals. To test how the model performs out of sample, data from 1970q1 – 2009q2 of the United States was used to see if the Savings and Loans crisis and the Subprime Mortgage crisis could be predicted. The predictions were made using the benchmark 1 year change model and the two models with the lagged domestic
credit to GDP variable. As with the in-sample predictions the cut-of level is 0.5.
4.5.1 United States
The quarters that are predicted as pre-crisis periods and the ones that are crisis/ post-crisis periods are:
Benchmark multinomial logit:
Figure 4 0 .2 .4 .6 .8 1 1970q1 1980q1 1990q1 2000q1 2010q1 time
Pre-crisis Crisis/ post-crisis cut-off
Pre-crisis: 1981q3 – 1982q4 2008q4 – 2009q2
Multinomial logit dom. cre. lag 8 and change in bank: Figure 5 0 .2 .4 .6 .8 1 1970q1 1980q1 1990q1 2000q1 2010q1 time
Pre-crisis Crisis/ post-crisis cut-off
Pre-crisis: 1980q3 – 1980q4 1981q3 – 1982q4 2009q1 – 2009q2 Crisis/ Post-crisis: 1983q2 – 1983q3 1985q4 – 1987q1 1995q2 – 1995q3
2006q4 – 2007q3
Multinomial logit dom. cre. lag 8 and bank level:
Figure 6 0 .2 .4 .6 .8 1 1970q1 1980q1 1990q1 2000q1 2010q1 time
Pre-crisis: 1980q4 1981q3 – 1982q4 2008q4 – 2009q2 Crisis/ Post-crisis: 1983q2 – 1983q3 1985q4 – 1987q1 1995q2 – 1995q3
2006q4 – 2007q3
What these predictions show is that the saving and loans crisis in the US at the end of the 1980s is to some extent predicted by our models. All three the models recognize a pre-crisis period in the early 1980s. This would signal a banking pre-crisis taking place within a year. However, as can be seen in Figures 4 to 6, this predicted pre-crisis period is not directly followed by a predicted crisis/ post-crisis period. The crisis period is recognized by the models years later: at the end of 1985. So the crisis itself is recognized at the correct date but the predictions according to the models are that it would take place at an earlier date. An explanation might be that in the early 1980s there were signals of an upcoming crisis (as recognized by the models) and the United States government tried to avert it. This was initially successful but it was not enough since the crisis took place after all. The ongoing financial crisis that started in the beginning of 2007 is recognized by all three the models but was not predicted by any of them. So there were no signals according to our models that a crisis was about to take place. When it did take place it was recognized by the models however. This can be interpreted as a crisis that came as a surprise and could not be predicted.
4.5.2 Japan
When considering the Japanese data an out-of-sample prediction yields the following results:
Figure 7 0 .2 .4 .6 .8 1 1970q1 1980q1 1990q1 2000q1 2010q1 Time
Pre-crisis Crisis/ post-crisis cut-off
Pre-crisis: 1978q3 1986q2 – 1986q4 1993q3 1994q1
1998q1
Crisis/ Post-crisis: 1995q4 1996q2 1998q2 – 1998q3 2001q1 – 2002q3 2003q3
Multinomial logit dom. cre. lag 8 and change in bank:
Figure 8 0 .2 .4 .6 .8 1 1970q1 1980q1 1990q1 2000q1 2010q1 Time
Pre-crisis Crisis/ post-crisis cut-off
1999q3 2008q4
Crisis/ Post-crisis: 1995q4 – 1996q1 2001q1 – 2002q2
Multinomial logit dom. cre. lag 8 and bank level:
Figure 9 0 .2 .4 .6 .8 1 1970q1 1980q1 1990q1 2000q1 2010q1 Time
Pre-crisis Crisis/ post-crisis cut-off
Pre-crisis: 1978q2 – 1978q3 1983q4 1986q2 – 187q1
1993q2 – 1994q 1999q3 2008q4
Crisis/ Post-crisis: 1995q3 – 1996q3 2001q1 – 2002q2
With these out-of-sample predictions the banking crisis that started in 1997 is predicted to some extent, as is the case with the US Savings and Loans crisis. Again the quarters that predict a crisis are a couple of years earlier than in reality. The same explanation as with the US Savings and Loans crisis can be given: the government tried to prevent a crisis from taking place and managed to postpone the crisis but in the end could not avert it. There are however a lot more false signals in the form of predictions of crises that did not take place, as is shown in Figures 7 to 9. So the models perform considerably worse than in case of the US.
also the case with the in-sample predictions. So although the sign of the domestic credit variable changed from negative to a more intuitive positive sign, this does not increase the performance of the model. Since the goal is to create an early warning system model that can predict crises the best, the transformed models are dismissed. These models do not have an advantage in the number of correctly called crises compared to the
5 Conclusion
The goal of this thesis is to create a multinomial logit model for banking crises, instead of the more used binomial logit model. We follow the methods of Bussiere and Fratzscher (2006) for currency crises. Data from Finland, Norway and Sweden, where banking crises took place at the beginning of the 1990s, were used.
The model with the best results was the one that includes explanatory variables defined as 1-year change variables. There were two variables that had the opposite sign of what is expected. The most remarkable is the negative sign of the change in domestic credit to
GDP ratio variable. So an increase in the domestic credit to GDP ratio should decrease
the occurrence of a banking crisis, although an increase is expected. By adding a lag of eight quarters the sign of this variable changed and gets the intuitive positive sign. This means that it takes some time before an increase in the domestic credit to GDP ratio has a positive effect on the occurrence of a banking crisis. This is intuitive since there is an increased chance of a crisis if these loans cannot be repaid.
The within-sample predictions for the pre-crisis quarters were high, with about 70% of the pre-crisis periods correctly called and no false signals in the benchmark model and one false signal in the models with the transformed variables. Out of sample, the banking crises in Japan of the late 1990s and the Savings and Loans crisis of the 1980s and the current crisis in the United States can be predicted to some extent. Both crises were predicted to take place earlier than they did. There also were in case of both the countries a period between the predicted pre-crisis prediction and crisis/ post-crisis periods. This could mean that the government recognized an upcoming crisis and tried to prevent it, however only succeeded in postponing it. The current crisis in the US is recognized as a crisis period, but it was not preceded by a pre-crisis period in the predictions of the model, so there are no immediate signs of it. So according to the model did the US Subprime Mortgage crisis that led to the current worldwide crisis came as a total surprise.
In future research the sample of countries could be extended to see how this will
Appendix
In-Sample Predictions with 1 to 6 quarter change variables with cut-off level 0.5
Table 13; 1 Quarter change
Si,t=0 Si,t=1 Si,t=2 Total
Yi,t=0 250 2 6 258
Yi,t=1 11 2 2 15
Yi,t=2 29 0 20 49
Total 290 4 28 322
Table 14; 2 Quarters change
Si,t=0 Si,t=1 Si,t=2 Total
Yi,t=0 246 3 7 256
Yi,t=1 12 0 2 14
Yi,t=2 20 0 29 49
Total 278 3 38 319
Table 15; 3 Quarters change
Si,t=0 Si,t=1 Si,t=2 Total
Yi,t=0 247 3 5 255
Yi,t=1 10 2 1 13
Yi,t=2 13 0 36 49
Total 270 5 42 317
Table 16; 4 Quarters change
Si,t=0 Si,t=1 Si,t=2 Total
Yi,t=0 248 0 6 254
Yi,t=1 4 9 0 13
Yi,t=2 10 0 39 49
Total 262 9 45 316
Table 17; 5 Quarters change
Si,t=0 Si,t=1 Si,t=2 Total
Yi,t=0 245 3 5 253
Yi,t=1 13 0 0 13
Yi,t=2 10 0 39 49
Total 268 3 44 315
Table 18; 6 Quarters change
Si,t=0 Si,t=1 Si,t=2 Total
Yi,t=0 244 3 5 252
Yi,t=1 11 2 0 13
Yi,t=2 8 0 41 49
Total 263 5 46 314
References
Berg, Andrew, Eduardo Borensztein, and Catherine Pattillo. 2004. “Assessing Early Warning Systems: How have they worked in practice?” IMF Working Paper, WP/04/52, International Monetary Fund, Washington.
Boyd, John, Gianni De Nicolò, and Elena Loukoianova. 2009. “Banking crises and crisis dating: theory and evidence.” IMF working paper, WP/09/141, International Monetary Fund, Washington.
Bussiere Matthieu, and Marcel Fratzscher. 2006. “Towards a new early warning system of financial crises.” Journal of International Money and Finance, 25: 953-973.
Caprio, Gerard, and Daniela Klingebiel. 1996. “Bank Insolvencies: Cross-Country Experience.” Policy Research Working Paper PRWP1620, Washington, D.C.: World Bank.
Caprio, Gerard, and Daniela Klingebiel. 1999. “Episodes of Systemic and Borderline Financial Crises.” Washington, D.C.: World Bank Mimeo.
Caprio, Gerard, Daniela Klingebiel, Luc Laeven, and Guillermo Noguera. 2005. “Banking Crises Database.” in Systemic Financial Crises, P. Honahan and L. Laeven eds. , Cambridge University Press, Cambridge, U.K.
Davis, E. Philip, and Dilruba Karim. 2008. “Comparing Early Warning Systems for banking crises.” Journal of Financial Stability, 4, June: 89–120.
Demirgüç-Kunt, Asli, and Enrica Detragiache. 2000. “Monitoring banking sector fragility: amultivariate logit approach.” World Bank Economic Review, 14(2): 287–307.
Demirgüç-Kunt, Asli, and Enrica Detragiache. 2005. “Cross-Country Empirical Studies of Systemic Bank Distress: A Survey.” National Institute of Economic Review, no.192, April.
Drees, Burkhard, and Ceyla Pazarbaşioğlu. 1998. “The Nordic banking crises. Pitfalls in financial liberalization.” Occasional Paper 161, IMF, Washington DC.
Eichengreen, Barry, Andrew K. Rose, Charles Wyplosz, Bernard Dumas, and Axel Weber. 1995. “Exchange rate mayhem: the antecedents and aftermath of speculative attacks.” Economic Policy, 21: 251–312.
Hilbe, Joseph. 2009. Logistic regression models. Boca Raton: Taylor & Francis Group.
Kaminsky, Graciela, Saul Lizondo, and Carmen Reinhart. 1998. “Leading indicators of currency crisis.” IMF Staff Papers 45/1, International Monetary Fund, Washington, D.C.
Kaminsky, Graciela. and Carmen Reinhart. 1999. “The Twin Crises: The Causes of
Banking and Balance-of-Payment Problems.” American Economic Review, Vol. 89, No. 3: 473-500.
Laeven, Luc and Fabian Valencia. 2008. “Systemic Banking Crises: A New Database.” IMF Working Paper 08/224, September.
Lestano, Jan Jacobs, and Gerard Kuper. 2003. “Indicators of financial crises do work! An early-warning system for six Asian countries.” Department of Economics, University of
Marchesi, Silvia. 2003. “Adoption of an IMF programme and debt rescheduling.” Journal
of Development Economics, 70(2): 403–423.
Minsky, Hyman. 1977. “A theory of systematic fragility.” in Financial crises: Institutions
and markets in a fragile environment, ed. by Edward I. Altman, and Arnold W. Sametz
(New York)
Reinhart, Carmen, and Kenneth Rogoff. 2008. “This time is different: A panoramic view of eight centuries of financial crises.” NBER Working Paper 13882, March.
Reinhart, Carmen, and Kenneth Rogoff. 2009. This time is different: A panoramic view of
eight centuries of financial crises. New Jersey: Princeton University Press.
Sachs, Jeffrey, Aaron Tornell, and Andres Velasco. 1996. “Financial crises in emerging markets: the lessons from 1995 (with comments and discussion).” Brookings Papers on
