scores or the Matthew correlation coefficients are always zero. Even though the forecasts
corresponding to these benchmark models do not lead to false alarms, they are not capable
to call correctly a crisis. In the rows labelled ARDL_pc and Probit_pc of Table 1 we report
the probability forecast performance (the KS scores are on the left hand side of every table
cell and the Matthews correlation coefficient are on the right hand side of every table cell)
corresponding with the predictions obtained using recursive OLS and ML estimation, re-
spectively, when using the principal components as regressors. More specifically, in Table
1, we provide results related to forecasting a crisis event defined as 1.5 standard deviation
above the mean. In Table 2 we report the probability forecast performance (associated to
the models already described for Table 1) of a crisis event defined as two standard deviation
above the mean. As we can observe from Tables 1 and 2, for most of the countries, both the
KS scores and the Matthews correlation coefficients are positive. This suggests that the use
of either a single composite leading indicator, modelled as a common shock, ut, underlying
the dynamics of the observables entering in the large dataset from which we extract the
principal components, or the direct use of the principal components in Probit regression
gives accurate forecasting results. The only exception, as suggested by Table 1 and 2, is the
EM P index for the Philippines, although, as shown in Table 2, the use of a probit model
specification with the common factors as regressors, gives a proportion of correct signals
exceeding the proportion of false alarms regarding a crisis event in terms of two standard
deviation above the EMP sample mean. These results are in line with the probability
forecast (over a 6 months horizon) performance of the models studied by Inoue and Rossi
(2006) to monitor nominal currency changes in East Asia. The authors (op. cit.) show that
diffusion index based probability forecasts are more accurate than other competing models
(except for South Korea). Furthermore, our forecast horizon (six months) are comparable
with those of private sector models (which are at one and three months horizons) for which a
poor out-of-sample forecasting performance (regarding the period pre and post Asian crisis)
has been shown by Berg et al. (2004).
Moreover, we argue that it is also interesting to assess the contribution of different set
of observable to overall forecasting performance of the Dynamic Factor model. For this pur-
pose, we remove a group of variables, for instance, the short term debt ratio of the stock of
international reserves in the five countries, from the large dataset x, and we apply again the
DF methodology. We then compare the forecasting performance of the new DF model with
the one associated with the whole dataset. More specifically, we are interested in assessing
whether and to what extent the new DF model forecasting performance is lower than the one
corresponding with diffusion indices obtained by using the whole sample of observations. In
14