Chapter IV: Early Warning System Models: The Next Steps Forward
- International Monetary Fund. Monetary and Capital Markets Department
- Published Date:
- March 2002
The Mexican (1994-95) and Asian (1997-98) crises stimulated a variety of empirical studies designed to identify both the causes of these crises and the determinants of the associated spillover effects (Kaminsky and Reinhart, 2000). To the extent that past crises can yield useful lessons about factors that contribute to a country's vulnerability to future crises, scholars and policymakers quickly realized that these empirical studies could be one element in a forwar-looking early warning system (EWS). As a result, a growing number of international financial institutions (IFIs) and central banks are using EWS models in their surveillance activities. Similarly, several investment banks have developed in-house EWS models aimed at providing foreign exchange trading advice to their clients and complementing their economic analysis of emerging markets.
These models typically have an empirical structure that attempts to forecast the likelihood of a certain type of “crisis” using factors such as country fundamentals, developments in the global economy and/or global financial markets, and, in some cases, political risks. A number of considerations have influenced the development of existing EWS models. First, despite a common origin in the academic models of the mid-1990s, existing EWS models differ sharply among themselves in terms of their definition of a “crisis” and in terms of the time horizon over which they attempt to forecast a crisis. Not surprisingly, these differences reflect the different interests of the various end users. For example, investment bank models define a crisis primarily in terms of variables such as changes in exchange rates and interest rates that are likely to affect the profitability of foreign exchange trading or investment positions. Moreover, investment bank models typically focus on one- to three-month forecast horizons that are regarded as most relevant for foreign exchange trading and investment positions. In contrast, EWS models used by IFIs and central banks in surveillance exercises tend to focus on variables associated with major balance of payments crises, namely large changes in exchange rates and/or central bank foreign exchange reserves. In addition, surveillance-linked EWS models typically attempt to forecast a country's vulnerability to crisis over a much longer time horizon than investment bank models. These models can have a forecast horizon of up to 24 months, reflecting in part the desire to have enough time to formulate corrective policy adjustments.
A second consideration that has affected the specification of the models and the interpretation of their forecasts has been the need to confront the trade-off between so-called type I and type II errors associated with the estimation of the statistical models. Type II errors (the acceptance of a false hypothesis of no crisis) can be minimized if EWS models are designed so that they have a low probability of missing a crisis. Unfortunately, adopting estimation procedures that minimize type II errors typically result in larger type I errors (the rejection of a true hypothesis of no crisis). In other words, the models can be calibrated to catch most crises, but only at the cost of many false claims.
A third consideration has been the availability and timeliness of data. For example, the absence of adequate historical time series on market interest rates and other financial variables has led to the exclusion of such variables in some EWS models. Moreover, variables available only with long reporting lags complicate the updating of forecasts and again discourage the use of what would otherwise be important explanatory variables.
The current EWS models have a mixed record in terms of forecasting accuracy, but they offer a systematic, objective and consistent method to predict crisis that avoids analysts' biases. Nonetheless, these models are just one of the many inputs into the IMF's surveillance process, which encompasses a comprehensive and intensive policy dialogue. As is the case with risk management models, they are not a substitute for sound and balanced judgments on financial weaknesses. This chapter briefly reviews the nature of the core EWS models used by the IMF in surveillance exercises, examines the performance of these models in terms of the accuracy of both in-sample and out-of sample forecasts, and considers various avenues for improving the usefulness of these models as surveillance tools. In particular, the use of alternative “building blocks” for predicting foreign exchange, debt, and banking crises, as well as the more efficient use of the information embedded in forward-looking asset prices, is discussed.
Current EWS Models at the IMF
As part of the IMF's surveillance activities, the IMF maintains two core EWS models: the Developing Countries Studies Division (DCSD) model and the Kaminsky, Lizondo, and Reinhart (KLR) Crisis Signals model (see Berg and others, 1999).1 In addition, the results of a number of external EWS models are monitored on an ongoing basis.
The core models attempt to forecast a country's vulnerability to a foreign exchange crisis de-fined as a large depreciation of the exchange rate and/or extensive losses of foreign exchange reserves over a 24-month forecast horizon. In this context, a crisis is said to have occurred when the “exchange market pressure” index—a weighted average of one-month changes in the exchange rate and foreign exchange reserves—is more than three (country-specific) standard deviations above the country average value.
The core models have parsimonious structures. For example, the DCSD model includes only five explanatory variables (real exchange rate overvaluation, current account, foreign exchange reserve losses, export growth, and the ratio of short-term debt to foreign exchange reserves).2 In contrast, the KLR Crisis Signals model uses twice as many variables (the first four variables used in the DCSD model as well as the ratio of foreign exchange reserves to M2, the growth of the ratio of reserves to M2, domestic credit growth, change in the money multiplier, real interest rate, and “excess” Ml balances—defined as actual Ml less an estimated demand for money). Box 4.1 describes some of the statistical properties of the two core models.
The output of the core models is complemented by monitoring the results from three investment bank models: Credit Suisse First Boston Emerging Markets Risk Indicator (CSFB-EMRI), Deutsche Bank Alarm Clock (DBAC), and Goldman Sachs GS-Watch.3 The forecasting horizon is one month for CSFB-EMRI and DBAC, and three months for GS-Watch. The choice of explanatory macroeconomic and financial variables is similar to those in the core IMF models. Among the three, there are some differences in the estimation methodology, but the major difference is in how each model defines crisis events. CSFB-EMRI defines a crisis as an exchange rate depreciation that exceeds 5 percent and is at least double the preceding month's depreciation. In contrast, DBAC defines exchange rate and interest rate events as currency devaluations and interest rate increases exceeding exogenous thresholds (typically a depreciation of more than 10 percent and an interest rate increase of more than 25 percent) and estimates the probability of both events simultaneously using a two-equation framework. Finally, GS-Watch defines a crisis event as one in which a financial price index (FPI) crosses an endogenous threshold, the latter being determined as a function of lagged values of the FPI.4
Box 4.1.The IMF's Core Early Warning System Models—A Primer
The core Early Warning System (EWS) models used in the IMF are the Developing Country Studies Division (DCSD) and the “modified” Kaminsky, Lizondo, and Reinhart (KLR) models.1 Both models define a crisis as an event during which an “exchange market pressure” index (EMPI)—a weighted average of monthly percentage depreciations in the nominal exchange rate and monthly percentage declines in foreign exchange reserves—exceeds its mean by more than three standard deviations.2 The EMPI is used to create the dependent binary variable, a crisis indicator, that is equal to one if a crisis occurs in the subsequent 24 months, and equal to zero otherwise.
The DCSD model uses a multivariate panel pro-bit regression technique to estimate the monthly probability that a country would suffer a crisis in the following 24 months. Explanatory variables include real exchange rate overvaluation, current account balance, foreign exchange reserve losses, export growth, and the ratio of short-term debt to foreign exchange reserves, measured in percentile terms. The model coefficients corresponding to the most recent estimation for the period December 1985 to July 1997 are shown in the table below.
The probabilities obtained from the model are converted into a binary indicator, an early warning indicator that signals a crisis and also allows for the statistical evaluation of different models. The indicator is equal to one (and is said to “call” a crisis) if the probability exceeds a cutoff threshold probability, and equal to zero otherwise. When the indicator is equal to one, the model correctly calls or signals a crisis if one ensues within 24 months, and gives a false alarm otherwise. When the indicator is zero, the model correctly calls a tranquil period if no crisis ensues within 24 months, and misses a crisis otherwise.
A statistical evaluation of the different models' accuracy relies heavily on the choice of the cutoff threshold probability. Clearly, the choice of a low threshold would lead to many false alarms but few missed crises.3 Alternatively, the choice of a large probability threshold would lead to few false alarms but many missed crises. In DCSD, the optimal selection of the threshold is obtained by minimizing an equally weighted sum of false alarms and missed crisis.
|Foreign exchange reserves||0.007|
|Short-term debt to reserves||0.002|
The econometric methodology of the KLR signals model is somewhat different from that of the DCSD model, except for the final stage that determines the threshold probability for an aggregate crisis index and a crisis is called. The KLR model assumes that each individual explanatory variable signals a crisis if its mean exceeds a variable-specific optimal threshold and a crisis occurs in the next 24 months. This threshold, which is expressed in percentile terms and is assumed equal across countries, is determined by minimizing the noise-to-signal ratio: the number of months during which the variable signaled a crisis incorrectly (false alarm or noise) divided by the number of months during which the variable signaled a crisis correctly.
KLR constructs a single composite crisis indicator equal to the weighted-sum of the explanatory variables, with the weights being equal to the inverse of each indicators' noise-to-signal ratio. The probability of crisis for each value of the aggregate index is then obtained by observing how often, within the sample, a given value of the aggregate index is followed by a crisis within 24 months, and the optimal probability threshold for the KLR model is determined in a similar way as for the DCSD model.1 The current versions of the models were developed in the IMF Research Department and are currently maintained by the International Capital Markets Department.2 Means, standard deviations, and weights are country-specific. Weights are calculated so that the variance of the two components of the index are equal.3 From a statistical point of view, the former might be thought of as type I errors and the latter as type II errors, under the null hypothesis of no crisis.
How well have the EWS models performed, especially in periods outside the original sample period used for estimation of the relevant parameters, in terms of both the appropriate forecast of an actual crisis and the avoidance of false signals (either missing an actual crisis or falsely forecasting a crisis that did not occur)?
A thorough evaluation of the IMF's EWS models recently concluded that the core models' forecasts are significant predictors of actual crises but that they still generate a substantial number of false alarms and missed crises (see Berg, Borensztein, and Pattillo, 2001). The study stresses the importance of out-of-sample prediction for an EWS model to be a useful surveillance tool, as well as the trade-off between missing crises and generating false alarms. The relatively long 24-month prediction horizon led the authors to evaluate the models according to the forecasts issued in July 1999. Overall, the authors concluded that the DCSD model performed reasonably well, as countries with a predicted probability of crisis above 50 percent subsequently had crises, and no crisis country had a probability of crisis below 26 percent. Furthermore, the DCSD model called 59 percent of the crises correctly but issued a large number of false alarms, 78 percent.5
In contrast, investment bank models do not perform as well as the core models when predicting exchange rate crises out-of-sample, but they appear to have satisfied their commercial objectives. Both the GS-Watch and CSFB models have an adequate in-sample performance (the percent of crises correctly called is 66 percent and 61 percent, respectively), but their out-of-sample predictions are much weaker (the percent of crises correctly called falls to 54 percent and 27 percent, respectively).6 However, investment bank models are regarded by the investment banks as performing reasonably well when evaluated solely on the merits of their short-term trading and/or investment recommendations. For example, the expected return of currency portfolios based on GS-Watch have consistently outperformed market neutral portfolios of emerging markets currencies (see Ades, Masih, and Tenengauzer, 1998).
The more recent performance of EWS models can be examined in terms of their forecasts in the periods surrounding two episodes that took place in 2001 and satisfy the EWS models definition of a crisis: namely, the devaluation of the Turkish lira in February 2001 and the devaluation of the Argentine peso in early January 2002.
The predicted probabilities of crisis and cutoff probabilities for the IMF's DCSD and KLR models before these episodes are shown in Figure 4.1. Both the DCSD and KLR Crisis Signals models correctly called for a crisis during the entire year preceding the Turkey crisis of February 2001. More precisely, both models correctly called the Turkey crisis in respectively 19 and 14 of the 24 months preceding the crisis. For the January 2002 crisis in Argentina, the DCSD model only started to signal problems in March 2001, when the run on the banks and foreign exchange reserves began. By October 2001, the DCSD model still called the crisis but only marginally and after a counterfactual decline in the probability of a foreign exchange crisis. The KLR model has not called a crisis in Argentina since August 2OOO.7
Figure 4.1.The IMF Early Warning System Models: Developing Country Studies Division (DCSD) and Kaminsky, Lizondo, and Reinhart (KLR) Probabilities of Crisis
Source: IMF staff calculations.
The performance of investment bank models during these two crises was also mixed. CSFB-EMRJ missed the Turkish episode, since it was predicting a decline in risk one month ahead of the crisis. In Argentina, the model indicated a significant increase in the country's risk score. In October 2001, DBAC called the Argentine episode correctly and predicted a possible devaluation as large as 20 percent, in part owing to the inclusion of interest rates as part of the definition of crisis. However, DBAC also missed the devaluation of the Turkish lira. Finally, GS-Watch correctly signaled events in Turkey three months ahead. For Argentina, GS-Watch did call a crisis from September 2001 to early December 2001. However, as in the case of the DCSD model, the GS-Watch model indicated that the crisis probability in Argentina declined in November and early December 2001. Also, as in the DBAC model, the GS-Watch model has issued a significant number of false alarms in the last quarter of 2001, as it called crisis in the next three months in almost every emerging market analyzed with the exception of Bulgaria, China, Chile, and Peru.
In sum, the current EWS models show mixed results in terms of forecasting accuracy, but they offer a systematic, objective, and consistent method to predict currency crisis, that helps avoid analysts' biases. Moreover, such models offer a single measure of risk in a statistically optimal way, that can be easier to interpret than, for instance, a large number of indicators giving different signals. Some analysts, noting the diverse group of countries singled out as most (and least) vulnerable by the different models, have suggested averaging the models' predictions. However, recent experience does not suggest that averaging leads to significant improvements.8 As a result, the next section discusses other avenues of improvement of EWS models.
EWS: A Way Forward
The recent evaluations of the performance of EWS models, as well as the experience with emerging markets crises around the turn of the century, suggest two potential avenues for increasing the useful ness of these models as surveillance tools. First, the different nature of recent financial crises suggests the desirability of developing a set of “building blocks” that would help forecast not only foreign exchange but also debt and banking crises, and identify the linkages between them. Second, EWS models could be improved or augmented by a more efficient use of the information embedded in forward-looking asset prices to anticipate financial market pressures. Although several crises took market participants by surprise, the increasing number of financial instruments available in emerging markets, combined with new techniques for extracting information from the prices of such instruments, suggest that this could be a fruitful avenue to pursue.9
Foreign Exchange Crises
Most of the existing EWS models omit the use of short-term interest rates as either a component of the definition of a foreign exchange crisis or as a determinant of the crisis. Domestic interest rates could be included as an independent event/equation along the lines of the DBAC model, or they could be combined with exchange rates in a more complete measure of “financial market pressures.” Similarly, excessive money or domestic credit creation is a key determinant of “first generation” foreign exchange crises, but the lags in the availability of monetary aggregates make them a less useful predictor of such crises; interest rates could reflect money market pressures in a more timely fashion.10
Stock market prices have some degree of predictive power in all market models, as well as in KLR,11 and sectoral stock prices show promising results that are worth pursuing further. In particular, Becker, Gelos, and Richards (2000) argued that sectoral differences in stock market performance may constitute valuable leading indicators of currency crises in emerging markets. Using company level data, the study indicates that around a year before the 1994-95 Mexican currency crisis net importers and financial companies began to continuously underperform the market, while net exporters showed continuously high abnormal stock returns.
New methods to extract information about market expectations from derivative prices could also be used as part of an EWS. Traditionally, forward exchange rates have been used in the mature markets to extract information about expected future spot rates.12 The proliferation of offshore nondeliverable forward markets for emerging market currencies, together with the deepening of onshore markets in some countries, increases the feasibility of applying these methods to emerging markets. For instance, a substantial increase in the probability of devaluation of the Argentine peso was priced in the 12-month nondeliverable forward rates In hue July-early August 2001 (see Figure 4.2). Also, the experience with the Korean won shows that once restrictions are removed and the market deep ens, the offshore nondeliverable forward market leads the domestic spot market-suggesting that price discovery happens primarily offshore and that information from offshore markets could be used to predict onshore financial pressures (Park, 2001).
Figure 4.2.Forward Exchange Rates for Argentine Peso
Source: IMF staff calculations based on data from Bloomberg L.P.
More recently, new techniques based on foreign exchange option prices provide market expectations on the whole risk-neutral probability distribution of future exchange rates (see Soderlind and Svensson, 1997; and Annex II of IMF, 1997). Unlike traditional exchange rate forecasts that provide only a point estimate of the mean of future exchange rates, option prices allow for the derivation of the probabilities associated with different ranges for the value of the underlying exchange rate. An example of the potential usefulness of these techniques is provided in Figure 4.3, which plots the probability density functions for the Brazilian real at different dates in July, October, and December of 2001.13
Figure 4.3.Spot Price and Implied Probability Density Function for the Brazilian Real
Source: IMF staff estimates based on JP Morgan data and Bloomberg.
By the end of July, forward rates on the Argentine peso had increased sharply. The implied probability density function for the real on July 31, 2001, shows that markets underestimated somewhat the contagion effects from Argentina. The Brazilian real exchange rate hit 2.60 on September 10, 2001, a level that at the end of July was considered by market participants to be likely with only a 2 percent probability. However, by mid-October, the implied probability density function displayed negative skewness, indicating the market was pricing currency options as if it expected a reversal of the depreciation of the real that had taken place between the end of July and mid-October. Since the Argentine peso continued to trade in the forward market at a substantial discount from spot, this suggests that market expected the real to decouple from the peso. The real did appreciate between mid-October and mid-December, and the December probability function shifted to the left and its dispersion dec reased.
There have been important recent crises (such as in Pakistan and Argentina) in which debt crises occurred either without or well before foreign exchange crises, and it may be that predicting debt crises is a worthwhile goal in and of itself. Recent research has highlighted the role of short-term external debt and rollover difficulties during such crises (Detragiache and Spilimbergo, 2001). However, while the inclusion of short-term debt in an early warning system has improved the performance of the DCSD model relative to others, the changing nature of crises suggests that it may be too restrictive to focus just on foreign exchange crises models. In particular, the short-term debt/reserves variable is a good indicator of liquidity problems, but it is not necessarily a good predictor of external solvency crises. This point is illustrated in Figure 4.4, which shows liquidity ratios, as well as debt service ratios (a traditional determinant of country risk) for Argentina, Turkey, and their peer group. While Argentina had a good external liquidity position relative to Turkey, the country's debt-service indicator was more than twice that of its peer group in 1999-2000. This reflected, and anticipated, the sovereign's solvency problems that would become more evident in 2001.
Figure 4.4.Argentina and Turkey: External Liquidity and Debt Service Ratios1
Sources: Fitch Research; and IMF staff estimates based on Fitch Research
1 The figures for 2001 are Fitch Research estimates. Liquidity ratio is defined as official reserves incl. gold plus banks' foreign assets/debt service plus liquid external liabilities; debt service ratio is defined as debt service/current receipts; Argentina's peer group by December 2000 included: Brazil, Colombia, Costa Rica, El Salvador, India, Kazakhstan, Lebanon, Panama, Peru, Philippines, Slovakia, Turkey, Venezuela.
Traditional models of country (credit) risk and external solvency indices could be used as part of an EWS for emerging market crises.14 An example of a country risk model is provided by Eichengreen and Mody (2000), who developed an econometric model of the determinants of emerging market debt spreads. The model was estimated for the years 1991-95 and produced satisfactory out-of-sample forecasts for 1996.15 The debt service to exports ratio is statistically significant in the regression,16 and has a sample average of 0.20; Argentina's figure for 1999-2000 is more than three times that average. An example of a model iusing external solvency indices is provided by Cohen (1991), who develops a simple solvency index that allows him to empirically assess whether or not indebted nations may have passed the point where they would default on their debt-service obligations.
The above models could be complemented with information on the term structure of emerging market bond spreads as well as on credit default swaps. The term structure of sovereign yield spreads could reflect expectations about the probable timing of any default, and the path of resolution and recovery of market access. In a recent study, Cunningham, Dixon, and Hayes (2001) show that Argentina's zero-coupon spread curve was sharply inverted by the end of July 2001 while the Brazilian curve was upward sloping, suggesting a heightened short-term credit risk in Argentina that was somewhat—but not totally—reduced after the June 2001 debt swap. The same pattern of curve inversion is indicated by default swap spreads: the differential between 10-year and one-year Argentina default swap spreads became strongly negative in the second half of last year (Figure 4.5), implying a higher short-term probability of default. Moreover, the differential is several times larger than the similar negative differentials observed during the Brazilian crises of 1998-99.
Figure 4.5.Slope of Default Swap Curve
Sources: IMF staff estimates based on JP Morgan data.
Recent experience has shown that temporary market closures (also referred to as “sudden stops” in capital inflows) have become a feature of international capital markets, and that this could increase the vulnerability of countries with relatively large external financing needs. Market closures are systemic rather than country-specific events that can be defined as the weeks when aggregate gross flows to all emerging markets are below 20 percent of average issuance levels (see IMF, 2001b). While in some cases market closures start with difficulties in a particular emerging market, in others they are the result of conditions in mature markets that constrain the supply of funds to the emerging market asset class as a whole. Preliminary studies have shown that mature market factors—such as the closure in the U.S. high-yield bond market and global equity market volatility—increase the probability of closure for emerging market issuers. More generally, studies show that factors like high U.S. interest rates and high-yield corporate bond spreads are associated with less issuance of emerging market debt. At the same time, domestic emerging market factors, such as high local market returns and high debt amortizations, are associated with higher issuance levels and lower probabilities of bond market closures (see Annex III of IMF, 2001a).
Banking crises share several common determinants with foreign exchange crises, and are sometimes a main cause of an exchange crisis, but recent efforts to predict banking crises have met with limited success. Kaminsky and Reinhart (1999) and Kaminsky (1998) find that excessive credit growth, recessions, and the burst of asset price bubbles tend to precede banking crises. However, with the exception of stock market prices, these variables (in particular, monetary and credit aggregates) are available with relatively long lags in emerging markets. More important, models that appropriately assign a higher weight to type II errors for banking crises, had low predictive power relative to the Asian banking crises.17 Also, studies that look at individual bank balance sheet indicators (see Gonzalez-Hermosillo, 1999), which could potentially identify problems in systemically important banks, find that loan quality and equity deteriorate rapidly before a bank fails. However, the lack of consistent cross-country results and the scarcity and lack of timeliness of data on these variables in emerging markets make individual bank failure quite difficult to predict.18
Bank stock prices may provide useful predictive information, and while only some emerging markets banks are publicly traded, these are in general the ones that would eventually cause systemic banking problems. Studies for the United States, for instance, show that stock prices help predict the financial condition of individual banks, even after taking into account past rating and financial statement information (see Berger, Davies, and Flannery, 2000; and Gunther, Levonian, and Moore, 2001). There is no systematic evidence for emerging markets, but Figure 4.6 demonstrates the potential usefulness of forward-looking bank stock prices to predict banking crises: the financial sector subcomponent of Korea's stock market index fell by more than three standard deviations almost a year before the onset of the foreign exchange crisis in Korea, a forewarning of potentially serious problems in the banking system.
Figure 4.6.Korea: Financial Institutions Index (1991-2001)1
Source: Korea Stock Exchange.
1 Sectoral index based on the set of financial companies in KOSPI of the Korea Stock Exchange. The thick solid line represents the average of the index through 05/30/96; the thin lines represent the corresponding standard deviation bands.
Banking crises are intimately linked to corporate financial stress, and new tools developed to estimate forward-looking measures of credit risk of both banks and corporates could be used as part of an EWS. These methods combine structural models of default with standard portfolio theory to assess the probability of extreme portfolio losses related to emerging market credit events. Structural models of default rest on the premise that a firm defaults on its debt when the market value of assets falls below the book value of its liabilities; viewing equity claims as a call option on the underlying asset value allows for the use of stock market prices to estimate marketbased default probabilities.19 A preliminary application of these methods to selected emerging markets, using KMV LLC proprietary software and techniques (see Bohn and Chai, 2001), suggest that they are not only usable in emerging markets but also that they lead to results that the authors regard as superior to those of econometrically-fitted models given the problems with the availability and quality of the data.
Financial Market Linkages
The development of separate “building blocks” for predicting foreign exchange, debt, and banking crises should take into account the close links between financial markets, and that one should be able to capture spillovers across bond, equity, and loan markets. Some of these linkages are already incorporated into each building block, but others are more subtle and may require a second stage of (joint) estimation and/or the use of scenario analyses. For example, Flood and Marion (2001) argue that studying currency and banking crises either in isolation or in perfect correlation with each other is inappropriate, producing biased estimates of the likelihood of crises. A key linkage between the two crises derives from the fact that government guarantees to depositors weaken the government's ability to fulfill other guarantees, such as that of maintaining a fixed exchange rate. Chang and Velasco (2000) relate bank and debt crises by jointly modeling the behavior of domestic bank depositors and foreign debt creditors. Finally, Christiano, Gust, and Roldos (forthcoming) show how a foreign exchange crisis could exacerbate the effects of a “sudden stop” in capital inflows (i.e., a debt crisis), by reducing the value of domestic assets that could be used as collateral in international credit markets. Although incorporating these linkages into a general EWS may be challenging, the insights could be taken into account in scenario analyses.
Contagion and Cross-Country Linkages
Contagion measures are included in some of the investment bank EWS models, but the adhoc treatment in these models contrasts sharply with the research on the issue. Traditional measures of contagion compare sample correlations among asset returns of different countries during tranquil and crises periods (see, for example, IMF, 2001b). However, unconditional correlations ignore the existence of interdependences, both through international trade and financial linkages across countries.20 There are two ways to overcome these limitations.
One approach is to assume that standard interdependencies are associated with small shocks to fundamentals whereas large crises and sporadic shocks may generate panics and herding behavior unrelated to fundamentals. This approach assumes that linkages across asset markets in periods of stress can be characterized by means of a measure derived from extreme value theory, one that captures the dependence or correlation between the extreme values of returns (or the tails of the distribution of returns). The approach allows for the derivation of the probability that a crisis occurs in one country, conditional on a crisis happening in another one. Estimates for the G-5 countries suggest that simultaneous crashes in stock markets are about two times more likely than in bond markets (Hartmann, Straetmans, and de Vries, 2001). A recent application of these techniques to emerging markets (Box 4.2) suggests that the intensity of stock market linkages during crises periods in Latin America has increased in recent years, a phenomenon that could be explained by shifts in the investors base holding emerging market assets.
Box 4.2.Alternative Measures of Contagion
The major emerging market crises of the 1990s were often associated with extensive spillover effects (contagion) across countries and markets. Such contagion has been viewed as arising from trade linkages, financial linkages (such as a common lender), a “wake-up” call that leads investors to reevaluate their view on countries that are “similar” to the crisis countries, and herding behavior.
The extent and pattern of contagion during a crisis has often been measured by examining changes in the level of correlation between financial variables in different markets and countries such as bond spreads and equity returns. For example, in the aftermath of the Long-Term Capital Management (LTCM) debacle and Russian default in the fourth quarter of 1998, the correlations between equity markets jumped upwards for all equity markets (see the first Figure). In contrast, the continuing problems faced by Argentina during the last quarter of 2001 did not affect stock markets in the region substantially as witnessed by the decline in correlations (see Chapter II).
Although the changing pattern of correlations provides one indicator of contagion, these correlations do not directly measure the behavior that is more relevant during crisis periods, namely the degree to which large negative returns in one country or market are associated with large negative returns in another country or market.1 Understanding how large shocks are transmitted across markets, and characterizing this transmission mechanism quantitatively is important if the effects of small financial shocks in one country propagate to another country in a different manner than large shocks. Indeed, large negative returns might have a higher level of correlation across countries if their occurrence leads investors to reevaluate the risks associated with investing in certain groups of countries or classes of assets.
Correlation of Daily Equity Returns During Russian Crisis and Long-Term Capital Management Failure1
Recent developments in extreme value theory (EVT) have allowed for a more precise measure of what can be called “extreme correlation”—the likelihood that a large negative financial return in one country is accompanied by a large negative return in another country. To derive this extreme correlation, one needs first to specify what constitutes a “large” or extreme negative return. In many studies, the criteria has been to focus on the bottom 5 percent of the negative returns over a specified sample period.2 Given this criteria, one can show that the univariate distribution of extreme returns is well captured by the class of generalized Pareto distributions, and that the degree of extreme correlation between two series of returns is given by the so-called Chi dependence measure, χ.3
The dependence measure χ can be roughly regarded as the conditional probability that, when a return in one market is the lower say 5 percent of all its outcomes, the other return will also be in the lower 5 percent of its outcomes. A higher value of χ implies an increasing likelihood that large negative returns in one market will be associated with large negative returns in the other.
While the value of χ can be examined during any given period to measure the degree of contagion as represented by the extreme correlation of asset returns, one can also consider how the value of χ has evolved over time to see if the degree of contagion has been rising or falling. As an example, the second figure portrays the evolution of the average χ values for a number of Latin American countries during the 1990s. Using weekly returns on stock indices over a 15-year period between December 31, 1987 and October 25, 2001, the extreme correlation measure is estimated for the bottom 5 percent of the negative returns using five-year rolling windows.4 A number of results stand out. First, there has been a secular increase in the values of χ throughout the 1990s. Second, the values of x become statistically different from zero (at the 5 percent level of confidence) only in the period since the Russian crisis of late 1998 (see Chan-Lau, Mathieson, and Yao, 2002). These results suggest that, for at least the Latin American economies included in the sample, a large negative return in the equity market in one country has become increasingly likely to be associated with large negative outcomes in the other markets. The reasons for this secular increase in the extreme correlation between the Latin American markets are as yet not fully understood. One possibility is that the nature of the investor base for Latin American equities has changed over the course of the 1990s, with socalled “crossover” investors playing an increasingly important role.5 Such investors will place a relatively small fraction of large portfolios in emerging market investments if they expect them to offer an attractive return. However, since the benchmarks used to evaluate the performance of portfolio managers of crossover investors typically do not encompass emerging market assets, they can abruptly reduce or eliminate their holdings of emerging market assets if the outlook for emerging markets deteriorates or if managers become more risk adverse and seek lower overall volatility of their holdings.
Average Extreme Correlations (x) of Weekly Equity Returns for Selected Latin American Countries1
Source: IMF staff calculations based on data from Primark Datastream LLC.
1 Average of five-year rolling windows for estimates of X for the following country pairs: Argentina-Brazil, Argentina-Chile, Argentina-Mexico, Brazil-Chile, and Brazil-Mexico.
Another approach estimates the empirical relevance of the different channels of transmission of shocks across countries and argues that only the unexplained residual correlation across returns should be regarded as contagion. For example, Kaminsky and Reinhart (2000) examine the role of international bank lending, the potential for cross-market hedging, and international trade in the transmission of crises across countries. The authors conclude that contagion is more regional than global, and that, although it is sometimes difficult to distinguish between channels, financial sector linkages appear to improve forecasting performance relatively more than trade linkages. Also, they claim that even though these estimates reflect past contagion, to the extent that current cross-hedging strategies use historical correlations, they could be a good forecast of future contagion. The incorporation of these linkages to an EWS should take into account the evolving nature of the investor base for emerging market instruments in order to assess potential changes in the relative importance of different channels. For instance, the change in investor behavior toward sectoral (rather than country) allocations is likely to have changed the nature of cross-country equity market correlations.
Going forward, EWS models will continue to be one element in the IMF's multilateral surveillance activities. To enhance the usefulness of these models, the IMF staff will focus on incorporating more information from forward-looking asset prices as well as developing “building blocks” for the prediction of foreign exchange, debt, and banking crises. In addition, there will be further analyses of the determinants of the extent and scope of contagion during crises. The IMF staff will report periodically on the progress made in developing these additional tools of analysis.
These EWS models are just one of the many inputs into the IMF's surveillance process, which encompasses a comprehensive and intensive policy dialogue.
The FPI, a weighted average of three-month exchange rate and reserve changes, is similar to the “exchange market pressure index” used in DCSD.
A false alarm may not necessarily be bad if it signals real risks are eliminated through, for example, policy adjustments.
The corresponding number of false alarms as percent of total alarms are 74 percent and 94 percent (in-sample) and 87 percent and 96 percent (out-of-sample).
One of the reasons for the models' contrasting performance in the Argentina and Turkey crises is explored in the discussion on debt crises in the next section.
See Forbes (2001). For example, the three most vulnerable countries in October 2001 were Malaysia, Israel, and Mexico for CSFB-EMRI; Turkey, South Africa, and the Czech Republic for DBAC; and Poland, India, and Argentina for GS-Watch. Furthermore, on many occasions the models' predictions of changes in the average vulnerability of emerging markets move in different directions. From February 200 1 to August 2001, average vulnerability according to CSFB-EMRI increased by 26 percent. In the same period, average vulnerability according to GS-Watch declined by 16 percent.
The use of financial market data suggests that the definition of a crisis and the predicition horizon may have to be adjusted to the shorter history and higher frequency of the data. Ideally, from the policymaker's point of view, it would be optimal to have indicators that signal a crisis several months or even years in advance. However, in the KLR model, all the indicators send their first crisis signal between a year and a year-and-a-half before the crisis erupts, and most of them give persistent signals that grow in intensity as one approaches the crisis. It seems that not much would be lost by moving to a 12-month horizon.
First generation models of crises are driven by excessive domestic credit creation needed to finance budget deficits, while second generation models explain crises as the result of shifts in investors' expectations whenever there is a conflict between a fixed exchange rate and other government objectives. Third generation models of crises emphasize financial frictions (Krugman, 1999).
Berg and Pattillo (1999) find that a rerun of the KLR model yields a noise-to-signal ratio greater than one for stock prices; this is in part due to a change in sample that removes some European countries and adds other emerging markets.
Although forward rates are generally biased predic-tors of future spot rates, they could be used togetherwith survey expectat ions of exchange rates and other factors that explain systematic forecast errors in foreign exchange markets (Lewis, 1995).
While these methods have been applied mostly to ma-ture market currencies, Campa, Chang, and Refalo (1999) used them to study the credibility of the Brazilian Real Plan of 1994-97. The method used to derive the probability density functions in Figure 4.3 differs from the ones used in that paper and follows that suggested in Malz (1996). In both cases, these are risk-neutral probability density functions, and there may be a bias derived from the need to compensate risk averse investors (Breuer, 2002).
Just as in the case of a balance of payments crisis, one would need an operational definition of what constitutes a “debt” or “sustainability” crisis.
Deviations of spreads from the levels predicted by credit ratings could also signal the need for further scrutiny of a country's fundamentals and market technicals, warning about potential reassessments of a country's creditworthiness (Sy, 2001).
Other significant determinants of spreads included factors such as the ratio of external debt to GDP, debt reschedulings, maturity of the instruments, and the existence of a private placement.
Demirguc-Kunt and Detragiache (1999) stress that, the higher the cost of missing a crisis relative to the cost of taking preventive action, the more concerned the policymaker will be about type II errors relative to type I errors.
The Financial Soundness Indicators (FSI) database, currently under cons truction in the IMF's Monetary and Exchange Affairs Department, could provide a useful input for a banking crises building block of an enhanced EWS.
Ades, Alberto, RumiMasih, and DanielTenengauzer, 1998, “GS-Watch: A New Framework for Predicting Financial Crises in Emerging Markets,” Emerging Markets Economic Research, New YorkGoldman Sachs,December.
Becker, Torbjorn, GastonGelos, and AnthonyRichards, 2000, “Devaluation Expectations and the Stock Market: The Case of Mexico in 1994–95,” IMF Working Paper No. 00/28 WashingtonInternational Monetary Fund.
Berg, Andrew, and CatherinePattillo, 1999, “Are Currency Crises Predictable? A Test,” IMF Staff Papers, International Monetary Fund, Vol. 46, pp. 107–38.
Berg, Andrew, EduardoBorensztein, Gian Maria, Milesi-Ferretti, and CatherinePattillo, 1999, “Anticipating Balance of Payments Crises—The Role of Early Warning Systems,” IMF Occasional Paper No. 186WashingtonInternational Monetary Fund.
Berg, Andrew, EduardoBorensztein, and CatherinePattillo, 2001, “Assessing Early Warning Systems: How Have They Worked in Practice?” (unpublished;Washington: International Monetary Fund).
Berger, Allen N., Sally M.Davies, and Mark J.Flannery, 2000, “Comparing Market and Supervisory Assessments of Bank Performance: Who Knows What When?” Journal of Money, Credit, and Banking, Vol., 32August, pp. 641–67.
Black, Fisher, and MyronScholes, 1973, “The Pricing of Options and Corporate Liabilities,” Journal of Political Economy, Vol. 81May–June, pp. 637–59.
Bohn, Jeffrey, and JingqingChai, 2001, “Sharpening Credit-Portfolio Risk Analysis in Emerging Asia,” (unpublished; KMV LLC).
Boyer, Brian H., MichaelGibson, and MicoLoretan, 1997, “Pitfalls in Tests for Changes in Correlations,” International Finance Discussion Paper No. 597WashingtonBoard of Governors of the Federal Reserve.
Breuer, Peter, 2002, “How Does the Volatility Risk Premium Affect the Informational Content of Currency Options?” (unpublished; Washington:International Monetary Fund).
Campa, José, KevinChang, , and JamesRefalo, 1999, “An Options-Based Analysis of Emerging Market Exchange Rate Expectations: Brazil's Real Flan, 1994–1997,”NBER Working Paper No. 6929Cambridge, MassachusettsNational Bureau of Economic Research.
Chang, Roberto, and AndresVelasco, 2000, “Banks, Debt Maturity and Financial Crises,” Journal of International Economics, Vol. 51June, pp. 169–94.
Chan-Lau, Jorge A., DonaldMathieson, and JamesYudong Yao, 2002, “Extreme Contagion across Equity Markets,” (unpublished; Washington: International Monetary Fund).
Christiano, Lawrence, ChrisGust, and JorgeRoldos, , forthcoming, “Monetary Policy in a Financial Crisis,” Journal of Economic Theory.
Cohen, Daniel, 1991, Private Lending to Sovereign States: A Theoretical AutopsyCambridge, MassachusettsMIT Press.
Coles, Stuart, JanetHeffernan, and JonathanTawn, 1999, “Dependence Measures for Extreme Value Analyses,” Extremes2, pp. 339–65.
Corsetti, Giancarlo, MarcelloPericoli, and MassimoSbracia, 2001, “Correlation Analysis of Financial Contagion: What You Should Know Before Running a Test,” (unpublished).
Cunningham, Alastair, LizDixon, and SimonHayes, 2001, “Analysing Yield Spreads on Emerging Market Sovereign Bonds,”in Financial Stability Review,Issue No. 11LondonBank of England.
de, Faria, JoseCarlos, 2001, “The Unbearable Lightness of the Real,” Deutsche Bank Global Markets ResearchOctober.
Demirguc-Kunt, Asli, and EnricaDetragiache, 1999, “Monitoring Banking Sector Fragility: A Multivariate Logit Approach,” IMF Working Paper No. 99/147WashingtonInternational Monetary Fund.
Detragiache, Enrica, and AntonioSpilimbergo, 2001, “Crises and Liquidity: Evidence and Interpretation,” IMF Working Paper 01/2WashingtonInternational Monetary Fund.
Eichengreen, Barry, and AshokaMody, 2000, “What Explains Changing Spreads on Emerging Market Debt: Fundamentals or Market Sentiment?” in Capital Flows and the Emerging Economies: Theory, Evidence, and Controversies, ed. by Sebastian EdwardsChicagoUniversity of Chicago Press.
Embrechts, Paul, AlexanderMcNeil, and DanielStraumann, 1999, “Correlation: Pitfalls and Alternatives,” RiskMay, pp. 69–71.
Flood, Robert, and NancyMarion, 2001, “A Model of the Joint Distribution of Banking and Exchange Rate Crises,”IMF Working Paper No. 01/213WashingtonInternational Monetary Fund.
Forbes, Kristin J., 2001, Anticipating Crises: Model Behavior or Stampeding Herds, presentation at the IMF Economic ForumNovember.
Forbes, Kristin J., and RobertoRigobon, 2001, “Measuring Contagion: Conceptual and Empirical Issues,” in International Financial Contagion, ed. by Stijn Claessens and Kristin J. ForbesBostonKluwer Academic Publishers.
Garber, Peter M., Robin L.Lumsdaine, and PaoloLongato, 2001, “Deutsche Bank Alarm Clock: Descriptive Manual Version 3,” Global Markets ResearchLondonDeutsche Bank,October.
Gonzalez-Hermosillo, Brenda1999, “Determinants of Ex-Ante Banking System Distress: A Macro-Micro Empirical Exploration of Some Recent Episodes,” IMF Working Paper No. 99/33WashingtonInternational Monetary Fund.
Gunther, Jeffrey W., Mark E.Levonian, and Robert R.Moore, 2001, “Can the Stock Market Tell Bank Supervisors Anything They Don't Already know?” Economic and Financial Review, Federal Reserve Bank of Dallas, Second Quarter.
Hartman, Philipp, StefanStraetmans, and Casper G.de Vries, 2001, “Asset Market Linkages in Crisis Periods,” CEPR Discussion Paper Series No. 2916LondonCentre for Economic Policy Research.
International Monetary Fund, 1997, International Capital Markets: Developments, Prospects, and Key Policy Issues, World Economic and Financial Surveys, Annex IIWashington, November.
International Monetary Fund, 2001a, International Capital Markets: Developments, Prospects, and Key Policy Issues, World Economic and Financial Surveys,Washington, August.
International Monetary Fund, 2001b, Emerging Market Financing, May issue, available on the Internet at http://www.imf.org/external/pubs/ft/emf/index.htm.
Kaminsky, Graciela, SaúlLizondo, and Carmen M.Reinhart, 1998, “Leading Indicators of Currency Crises,” Staff Papers, International Monetary Fund, Vol. 45, pp. 1–48.
Kaminsky, Graciela, 1998, “Currency and Banking Crises: The Early Warnings of Distress,” International Finance Discussion Paper No. 629WashingtonInternational Monetary Fund.
Kaminsky, Graciela, and CarmenReinhart1999, “The Twin Crisis: The Causes of Banking and Balance-of-Payments Problems,” American Economic Review, Vol. 89June, pp. 473–500.
Kaminsky, Graciela, and CarmenReinhart2000, “On Crises, Contagion, and Confusion,” Journal of International Economics, Vol. 51June, pp. 145–68.
Kell, Michael, and AxelSchimmelpfennig, 2002, “Fiscal Indicators in an EWS—Some Empirical Results,” (unpublished; Washington: International Monetary Fund).
Krugman, Paul, 1999, “Balance Sheets, the Transfer Problem, and Financial Crises,” in International Finance and Financial Crises—Essays in Honor of Robert P. Flood, Jr.LondonKluwer Academic Publishers;WashingtonInternational Monetary Fund.
Lewis, Karen, 1995, “Puzzles in International Financial Markets,” Chapter 37 in Grossman and Rogoff, eds., Handbook of International Economics, Vol. III, Elsevier Science.
Longin, Francois, and BrunoSolnik, 2001, “Extreme Correlation of International Equity Markets,” Journal of Finance, Vol. 56April, pp. 649–76.
Malz, Allan, 1996, “Options-based Estimates of the Probability Distribution of Exchange Rates and Currency Excess Returns” (unpublished;New York: Federal Reserve Bank).
Merton, R.C., 1973, “Theory of Rational Option Pricing,” Bell Journal of Economics and Management Science 4,Spring,141–83.
Mulder, Christian, RobertoPerrelli, and ManuelRocha, forthcoming, “The Role of Corporate, Legal and Macro Balance Sheet Indicators in Crisis Detection and Prevention,” IMF Working PaperWashingtonInternational Monetary Fund.
Park, Jinwoo, 2001, “Information Flows between Nondeliverable Forward (NDF) and Spot Markets: Evidence from Korean Currency,” Pacific-Basin Finance Journal, Vol. 9August, pp. 363–77.
Poon, Ser-Huang, MichaelRockinger, and JonathanTawn, 2001, “New Extreme-Value Dependence Measures and Finance Applications,” CEPR Discussion Paper No. 2762LondonCentre for Economic Policy Research.
Roy, Amlan, 2001, “Emerging Markets Risk Indicator (EMRI),” Global Emerging Markets StrategyLondonCredit Suisse First Bank, February.
Soderlind, Paul, and LarsSvensson, 1997, “New Techniques to Extract Market Expectations from Financial Instruments,” Journal of Monetary Economics, Vol. 40October, pp. 383–429.