IV EWS Models and the Severity of Currency Crises

Richard Hemming, Axel Schimmelpfennig, and Michael Kell
Published Date:
May 2003
  • ShareShare
Show Summary Details

This section summarizes the results of various statistical and econometric analyses of the dataset described in the previous section. First, the extent to which fiscal variables can help to predict crises is examined using two popular EWS models. The second part investigates whether fiscal variables can help to explain the severity of currency crises. Further explanation of the methodologies and detailed results are given in Appendix IV.

Using Fiscal Variables to Predict Crises

EWS models provide a systematic empirical framework for estimating the likelihood of a crisis over a given time horizon based on a combination of vulnerability indicators.29 The number of EWS models has grown rapidly in recent years, and they have become an important part of IMF work on crisis prevention. The basic approach with EWS models is to determine empirically the relationship between past crises and a range of factors—such as country fundamentals, developments in the global economy and financial markets, and political risks—and then to use the latest values of these variables to predict the probability of future crises. EWS models should be evaluated primarily on the basis of their ability to predict future crises (i.e., out-of-sample testing), rather than on how well the model fits the observations from which it was estimated (i.e., in-sample testing).30

EWS models are far from perfect forecasting tools. They have a track record of both missing crises and sending false alarms, and they are certainly not sufficiently accurate to be used as the sole method of predicting crises. But they can contribute to the analysis of vulnerability in conjunction with more traditional surveillance methods and other indicators; in particular, the “mechanical” approach of EWS models can provide a relatively objective and systematic starting point for crisis prediction.31 Moreover, Berg, Borensztein, and Pattillo (2003) conclude that the best EWS models performed markedly better than other predictors—such as spreads, ratings, and the assessments of market analysts—particularly over the period of the Asian crisis.32

The added value of fiscal variables in predicting crises is unclear from the literature. Appendix I includes a survey of the EWS literature, concentrating on those studies that have included at least one fiscal variable. Some EWS models of currency crises find an important role for fiscal variables, but others do not. The evidence from banking crisis EWS models is clearer: fiscal variables do not seem important. However, the range of fiscal variables that have been tested is rather narrow. There do not appear to be any EWS models of debt crises, despite the fact that fiscal variables should be relatively useful in predicting debt crises. More generally, the literature review reveals no EWS studies that have specifically focused on fiscal variables, or indeed tested more than a few standard deficit and debt variables. However, there is sufficient evidence of a role for fiscal variables in predicting crises to merit a more systematic examination of fiscal vulnerability indicators using EWS models.

This study uses two different EWS methodologies—the signals approach and the more widely used probit approach. In the signals approach, the idea is that when a variable departs significantly from its normal historical behavior, it may be sending a signal of an impending crisis. Historical data are used to assess the accuracy of the signals sent by particular variables prior to actual crises, with the expectation that past relationships will provide a reliable basis for deciding which current signals may be indicating future crises.33

The signals approach is closely related to the event study approach. But it extends and refines the event studies by allowing a more precise comparison between variables in terms of their leading indicator properties. In particular, the approach allows the direct comparison and ranking of alternative indicators, in terms of their track record in failing to signal crises (Type I errors) and sending false positives (Type II errors). However, the signals approach is typically univariate, and not amenable to tests of statistical significance.34

The probit approach is multivariate and therefore accounts for the correlations and interactions between different variables in forecasting crises. The approach uses probit (or sometimes logit) regressions to identify the factors that jointly explain the occurrence of crises. Latest values of the explanatory variables are then combined with the estimated coefficients to give a predicted probability of a crisis happening during a given window, typically 6, 12, or 24 months ahead. Further details about both the signals and probit approaches are provided in Appendix IV.

Signals EWS Results

Separate signals EWS models are estimated for currency, debt, and banking crises. Detailed results are given in Appendix IV. Table 4.1 summarizes the findings for the six best performing indicators of currency, debt, and banking crises. A signal counts as good when a crisis occurs in either of the two subsequent years (column 2) and as bad when no crisis occurs in either of the two subsequent years (column 3). The ratio of column 3 to column 2 gives the noise-to-signal ratio (column 4), by which indicators are ranked. Other information that can be used to compare indicators is also shown. The main points to note are the following:

Table 4.1.Summary Results from the Signals Approach
Percentage of

Crises for

Which Data

Are Availabe1

Good Signals as

a Percentage

of Possible

Good Signals2

Bad Signals as

a Percentage

of Possible

Bad Signals2


Signal Ratio





Percentage of

Crises with

No Signal

Percentage of

Crises with

at Least

One Signal

Percentage of

Crises with

at Least

Two Signals

Currency crises
Primary balance3624100.4390673314
Short-term debt862090.4493703010
Defense expenditure741040.479881190
Social expenditure501470.489572280
International trade taxes83740.549885150
Revenue buoyancy83740.559888130
Debt crises
Foreign currency debt433130.1198445622
Social expenditure571340.319875250
Short-term debt1002790.3391574310
Revenue buoyancy81940.419882180
Change in net claims on government95830.439885150
Tax buoyancy81940.439882180
Banking crises
Foreign currency debt342550.229764369
Total debt562260.288067336
Overall balance881130.299079210
Short-term debt9734110.3392584226
Foreign debt501450.339575250
Source: Appendix IV.

Out of a maximum of 58 currency crises, 21 debt crises, and 32 banking crises.

See Appendix IV for further explanation.

Source: Appendix IV.

Out of a maximum of 58 currency crises, 21 debt crises, and 32 banking crises.

See Appendix IV for further explanation.

  • Some fiscal variables are good predictors of currency crises. Short-term debt, foreign currency debt, the overall and primary balances, and some financing variables perform as well as the best indicators of currency crises in other (annual) signals EWS models, such as the current account deficit. These variables are less good, but still useful, predictors of debt crises.
  • A few of the expenditure and revenue variables are also useful predictors of currency and debt crises. These include defense and social spending, and international trade taxes. This is in contrast to the findings of the event studies in Section III, but provides some empirical support for the suggestions in Hemming and Petrie (2002) that a high share of expenditure that is nondiscretionary and reliance on volatile revenue sources can add to fiscal vulnerability.
  • Fiscal variables have a similar signaling performance overall for banking crises as for debt and currency crises. Despite weaker theoretical links between fiscal policy and banking crises, and the different findings of the event studies, this suggests at least some role for fiscal variables in signaling banking crises.35
  • However, fiscal variables fail to signal a very high proportion of crises. Even the best performing indicator of currency crises misses around two-thirds of all crises. And the proportion of crises signaled in both years prior to the crisis is very low—just over a quarter for the best performing indicator, and more typically around 10 percent. So fiscal variables, in and of themselves, do not appear to be reliable crisis predictors.
  • At the same time, the better performing fiscal variables generally send very few false alarms. This implies that if a fiscal variable does exceed its critical threshold and signals an impending crisis, the warning that is being given should be taken seriously.

Probit EWS Results

The approach taken is to estimate the EWS model developed by the IMF’s Developing Countries Studies Division (DCSD), and then add fiscal variables to determine whether this improves the in-sample predictions of the model. The DCSD model predicts currency crises using the crisis definition set out in Section III, with a 24-month-ahead forecasting horizon. The parameters are generated by probit regressions, using monthly data, and there are six explanatory variables—exchange rate overvaluation, measured by the deviation of the real exchange rate from its long-term trend; the current account deficit relative to GDP; reserves growth; export growth; the ratio of short-term debt to reserves; and the ratio of M2 to reserves.36 Fiscal variables have not previously been examined systematically in the context of this model.

The data on which the DCSD model has been estimated cover the same 29 countries and the same period (1970–2000) as the fiscal dataset described in Section II.37 The DCSD dataset contains mostly financial and monetary variables that are available on a monthly basis, and also some variables available on a quarterly basis, such as the current account balance. All variables have been percentiled to remove country-specific effects. The fiscal variables are, in almost all cases, only available on an annual basis, and therefore have to be converted to a monthly frequency to be used with the DCSD model.

Results are presented for three models. These are the benchmark specification, which includes only the core DCSD variables and no fiscal variables (Model 1); a specification including the DCSD variables plus the two best performing fiscal variables (the change in net claims on government and foreign currency debt) (Model 2); and a specification with only the best performing fiscal variables among each subgroup of deficit, debt, expenditure, and revenue variables (Model 3). Detailed results are presented in Appendix IV. Table 4.2 summarizes the results for each specification.

Table 4.2.Summary of Probit EWS Results
Model 1:

Model 2:

Best Fiscal Variables
Model 3:

Only Fiscal Variables
Pseudo R20.220.240.17
Correctly called (percent)
Tranquil periods89.890.587.8
Exchange rate overvaluation0.0210.60.0211.5
Current account deficit0.
Reserves growth0.
Export growth0.
Short-term debt to reserves0.
M2 to reserves0.
Change in net claims on government0.
Foreign currency debt0.
Interest expenditure–0.01–9.6
Social expenditure0.016.6
Total revenue0.019.3
International trade taxes0.016.8
Source: Appendix IV.
Source: Appendix IV.

The main results are as follows:

  • All of the deficit and financing variables enter the DCSD model significantly at the 5 percent level and with the expected sign, but none leads to a significant improvement of the model’s in-sample predictive power.
  • Among the debt variables, only foreign currency debt enters the DCSD model significantly at the 5 percent level and with the expected sign. This tends to confirm the finding of the signals EWS approach that the composition of public debt matters for predicting crises.
  • The revenue variables add very little to the model. Two of the six revenue variables—total revenue and international trade taxes—enter the model significantly, but there is no improvement in predictive power.
  • The results for the expenditure variables are mixed. Interest expenditure enters significantly and improves the predictive power of the model. However, the estimated coefficient is not robust. Social expenditure improves the predictive power of the model and enters significantly at the 5 percent level, and this result is stable.
  • Of the best performers from each group of fiscal variables, only the change in net claims on government and foreign currency debt are significant and robust. Compared with the DCSD specification, including these two variables marginally reduces the number of correctly called crisis signals, but it improves the model’s predictive power for tranquil episodes, and thus its overall predictive power.
  • A specification including only fiscal variables performs as well as the DCSD model in terms of calling tranquil periods. This specification performs less well in terms of signaling crises, but still correctly calls around 42 percent of crisis episodes.

Overall, there is robust evidence that currency crises are correlated with some fiscal variables. But in terms of predicting crises, fiscal variables add little to the existing DCSD specification, although they do help signal tranquil episodes.38 One way of reconciling these findings is to note that the effects of fiscal variables on crisis vulnerability may operate through variables already included in the DCSD model. It seems plausible that loose fiscal policy could lead to overvaluation of the real exchange rate and a wider current account deficit, depletion of reserves, and a buildup of short-term debt. In other words, the findings presented here are consistent with the notion that fiscal variables affect crisis vulnerability indirectly.

Fiscal Variables and the Severity of Currency Crises

In addition to influencing the likelihood and timing of financial crises, a country’s fiscal situation could affect the depth or severity of a crisis. For example, a sharp increase in interest rates resulting from pressure on the exchange rate could suddenly worsen public debt dynamics and exacerbate the flight of capital; or a banking crisis could crystallize large contingent liabilities, such as those needed to protect deposits or to recapitalize the banking system, adding to public debt. Alternatively, a weak precrisis fiscal position could make it difficult to respond to a crisis by supporting aggregate demand; conversely, structural fiscal rigidities such as weak capacity for tax administration or budget management could make it hard to tighten fiscal policy quickly and thus restore confidence when a crisis hits.

A number of studies attempt to explain the severity of currency crises conditional on a crisis occurring elsewhere. Sachs, Tornell, and Velasco (1996) (STV) examine the spread of crises during the six months following the onset of the Mexican crisis in December 1994. The severity of crisis is measured using an index of exchange market pressure similar to the FMP variable described in Section III. The explanatory variables are exchange rate overvaluation, measured by the depreciation of the trade-weighted real effective exchange rate; banking sector weakness, proxied by growth of credit to the private sector; and vulnerability to capital inflow reversals, proxied by the ratio of M2 to reserves. STV also include two dummy variables for weak fundamentals and low reserves. STV conclude that their model fits the data well. In addition, they include a measure of government consumption to proxy the extent to which lax fiscal policy prior to the crisis explains the severity of pressure on the exchange rate during the crisis and find that the percentage change in government consumption in the period 1990–94 is statistically significant, but only in countries with weak fundamentals and low reserves. They also find some evidence that lax fiscal policy contributes indirectly to the severity of crises by influencing the degree of exchange rate overvaluation prior to the crisis.

To investigate further the role of fiscal factors in explaining changes in the FMP index, the first step is the estimation of a variant of the STV model to which fiscal variables are added. The sample of countries (the 58 currency crises in the dataset), the fiscal variables, and the currency crisis index are all as described in Section III. Detailed results are given in Appendix IV, but Table 4.3 shows the benchmark specification and various alternative specifications, including the three best performing fiscal variables. While there is evidence that the severity of crisis is positively correlated with the change in net claims on government (Fiscal Models 1 and 2 in Table 4.3), this relationship becomes insignificant when other explanatory variables are included (Fiscal Models 3 and 4). This suggests that, consistent with similar empirical studies using the STV approach, fiscal variables have at most an indirect impact on crisis severity.

Table 4.3.Explaining the Severity of Currency Crises (STV Approach)
Benchmark ModelFiscal Model 1Fiscal Model 2Fiscal Model 3Fiscal Model 4
Test for joint significanceF(3,53)27.9F(1,52)4.9F(3, 50)36.0F(5,32)17.3F(5,39)82.2
Adjusted R20.
Exchange rate overvaluation (in the previous period)0.008.620.007.850.006.800.007.20
M2 to reserves0.
Weak fundamentals1.462.
Change in net claims on government0.
Total debt–0.00–0.01
Long-term debt0.021.19
Source: Appendix IV.
Source: Appendix IV.

As an alternative, panel estimation is used to explain changes in the FMP index over the entire pooled sample of crisis and tranquil periods combined. This is akin to modeling changes in the exchange rate, with all its attendant difficulties; however, it provides a much larger sample for estimation. The methodology and results are discussed in Appendix IV, with the key results shown in Table 4.4. The panel approach finds stronger evidence of a role for fiscal variables in influencing pressure in the foreign exchange market. When the best performers among the fiscal variables are entered jointly in the panel model, elevated deficit, debt, and interest variables all increase pressure in the foreign exchange market, over and above the effect of other factors. The relationships are robust over different specifications, measures of fiscal variables, and sample composition.

Table 4.4.Explaining Changes in the FMP Index (Panel Approach, Fixed Effects)
Benchmark ModelFiscal Model 1Fiscal Model 2Fiscal Model 3Fiscal Model 4
Countries included2927272727
Per country
Minimum observations64444
Average observations24.42120.720.120.4
Maximum observations2929292929
Test for joint significanceF(6,674)73.66F(9,530)43.43F(9,523)44.74F(9,508)44.00F(9,515)42.67
Hausman test for random effectsχ2(6)26.77χ2(9)13.81χ2(9)14.83χ2(9)87.92χ2(9)13.85
Results similar for random effectsYesYesYesYesYes
Exchange rate over valuation (in the previous period)0.002.320.002.350.002.350.
Change in M2 to reserves0.0114.260.0113.200.0113.390.0113.380.0113.11
Weak fundamentals0.453.620.312.360.312.340.322.360.322.37
Export growth–0.02–7.08–0.01–4.92–0.01–4.56–0.01–4.29–0.01–4.68
Short-term debt to reserves0.
Real GDP growth–0.02–2.63–0.02–2.14–0.02–2.21–0.02–2.15–0.02–1.99
Actuarial deficit0.022.670.011.980.022.390.023.12
Public external debt0.
Long-term debt0.
Interest expenditure10.062.390.052.07
Interest payments10.
Source: Appendix IV.

In percent of GDP; interest expenditure from the IMF’s Government Finance Statistics (GFS) functional classification and interest payments from the GFS economic classification.

Source: Appendix IV.

In percent of GDP; interest expenditure from the IMF’s Government Finance Statistics (GFS) functional classification and interest payments from the GFS economic classification.

Summary of Findings

A thorough examination of the univariate leading indicator properties of a range of fiscal variables finds some that are potentially useful for signaling crises. The best vulnerability indicators—short-term debt, foreign currency debt, and various deficit measures—perform as well as the best (annual) leading indicators in other signals EWS studies. But even the best performing indicators fail to signal around two-thirds of crises, although they send very few false alarms. These findings apply as much to banking crises as currency and debt crises, suggesting that fiscal variables may indeed have a part to play in predicting all three types of financial crisis.

In a multivariate context, there is robust evidence that fiscal variables are correlated with currency crises after controlling for other variables. But in terms of predicting crises, fiscal variables add relatively little to the IMF’s main EWS model. It seems plausible that the variables already included in the model are capturing the main fiscal effects on crisis vulnerability—through exchange rate overvaluation, depletion of reserves, buildup of short-term debt, and so on. In probit EWS models, where parsimony is important, it is unlikely that fiscal variables will add sufficient power to the predictions to merit inclusion. This is reinforced by the relatively low frequency and significant time lags associated with most fiscal data. That said, the addition of certain fiscal variables to the probit EWS model does lead to some improvement in its ability to predict tranquil or noncrisis periods.

Finally, fiscal variables have limited value in explaining the severity of currency crises conditional on a crisis having occurred. But when a measure of exchange market pressure is estimated over the entire pooled data sample, there is robust evidence that loose fiscal policy, high public debt, and high interest expenditure are correlated with pressure in the foreign exchange market. This is further evidence that fiscal variables may have a role to play in explaining and predicting tranquil or noncrisis periods in currency markets, albeit indirectly through their effect on the exchange rate and reserves.

    Other Resources Citing This Publication