specification with the explanatory variables measured in percentile terms. Also, the private
sector models developed by Credit Suisse First Boston, Goldman Sachs and Deutsche Bank
use logit regressions. The study of Van Rijckeghem and Weder (2003) uses probit regression
to examine the role of a common lender channel in triggering crisis events. The authors
(op. cit.) rely on disaggregate data on external debt produced by the BIS to construct mea-
sures of competition for fund in order to explore the role played by a common lender channel.
While the previous studies (with the exception of Goldstein et al., 2000) are based upon
in sample prediction, the studies of Berg and Pattillo (1999) and of Berg et al. (2004) show
that the out-of-sample forecasts of East Asian 1997-1998 currency crisis events produced
through the signal approach are fairly good, with many of the most vulnerable countries
in fact being the hardest hit in terms of crisis severity. In Berg et al. (2004) the out of
sample forecasting performance for the post East Asian crisis period is also assessed and the
signal approach is found to perform better than the regression based approach. The short-
horizon private sector models developed by Credit Suisse First Boston, Goldman Sachs and
Deutsche Bank are shown to have a poor out of sample predictive performance. The study of
Fuertes and Kalotychou (2006a) concentrates on several logit model specifications (with dif-
ferent degree of unobserved heterogeneity) which are found to be less successful, in terms of
out-of-sample forecasting performance of sovereign default crisis, than a pooled logit model.
Further, Fuertes and Kalotychou (2006b) consider not only logit regression but also a non
parametric method based upon K-means clustering to predict crisis events. Fuertes and
Kalotychou (2006b) find that a combination of forecasts from the different methods gener-
ally outperform both the individual and naive forecasts. The empirical analysis reveals that
the best combining scheme depends also on the decision-makers preferences regarding the
desired trade-off between missed defaults and false alarms (see also the study of Bussiere
and Fratzscher, 2002, on the issue of designing the features of their EWS model according
to the preferences and to the degree of risk-aversion of policymakers).
Finally, there are studies that have constructed composite leading indicators of currency
crisis events using diffusion indices rather than the weighting scheme suggested by Kaminsky
(1998b) and by Goldstein et al. (2000). Beyond the studies which rely upon the construction
of diffusion indices using principal component analysis fitted to a large dataset (see the studies
of Cipollini and Kapetanios, 2003; Inoue and Rossi, 2006; Jacobs et al., 2008) there are few
studies which extract common factors from small datasets. Mody and Taylor (2003) use
Kalman filter estimation of state space models in order to extract a measure of regional
vulnerability in a number of emerging market countries, and, in order to produce in sample