disaggregated data on external debt is based upon the suggestion given by various studies
on financial contagion. The literature on financial contagion puts the emphasis on the role
of the geographical composition of external debt (e.g., the common lender channel), and on
the maturity mismatch in explaining the spread of the crisis hitting one country to other
countries. Given that BIS external debt data are available for a relatively long data span
(starting from 1983) only at low frequency (bi-annual basis), the number of cross sections
exceed the time series observations, hence it is not practical to use standard state space
model methods to extract factors (especially, when one is interested in recursive estimation
in order to produce out-of-sample predictions for the forecast evaluation period). There-
fore, we use factors extraction based on principal components analysis as suggested by Stock
and Watson (2002)2. Furthermore, contrary to previous studies which have only explored
the out-of-sample forecasting performance of the composite indicator, we also assess which
group of variables (say, short term debt) play an important role in predicting out-of-sample
the crisis event through the composite leading indicator. The use of principal components
for the purpose of constructing leading indicators of currency crisis has already been put
forward by Cipollini and Kapetanios (2003) who produce out-of-sample point forecast of the
EMP index in East Asian countries for the 1997-1998 period, and also by Inoue and Rossi
(2006) and by Jacobs et al. (2008). While the probability forecast of currency crisis events
in the study of Inoue and Rossi (2006) rely on modelling the conditional second moments
of the nominal exchange rate through principal components, the probability forecasts we
produce are obtained by modelling the conditional first moment of the EMP index through
common factors3. As for the out-of-sample probability forecasts, Jacobs et al. (2008) focus
only on one year, 2002, (where there is no evidence of particular turbulence in the East Asian
currency markets) and there is no comparison with forecasts produced by other benchmark
models, including an AR for the EMP . In this paper, we produce out-of-sample probability
forecasts, first, by employing probit modelling (which is standard when using the parametric
approach to EWS). Second, we also use stochastic simulation of the Dynamic Factor model,
(DF ), following the suggestion of Forni et al. (2005) on how to retrieve a single common
shock which we interpret as a regional vulnerability indicator. The probability forecast ac-
curacy is assessed by using both the Kuipers Score and the Matthews correlation coefficient.
external debt. However, Frenkel and Rose (1996) focus on predicting large nominal exchange rate deprecia-
tions and not crisis events defined in terms of the EMP index.
2The Stock and Watson (2002) method is a time-domain based approach. In Forni, et al, (2000) the factor
extraction is obtained using an a frequency domain based approach. Finally, Kapetanios and Marcellino
(2003) use an approach based upon a state space model.
3 Inoue and Rossi (2006) probability forecasts are for crisis events defined in terms of nominal exchange
rate depreciations bigger than 20%. We argue that this type of currency crisis considered does not include
speculative attacks successfully warded off by the authorities through reserve sales