is on crisis event defined in terms of 1.5 standard deviation. However, the short term debt
seems to be more important than long term debt only for Indonesia and Thailand when the
focus is on crisis events defined in terms of two standard deviations. As argued by Chang
and Velasco (1999) among the others, given that the developed countries’ loan contracts were
of short maturity, the lending country rebalancing needs might imply not only the refusal
to extend new credits to the other borrowers, but also the refusal to roll-over their existing
loans. Finally, from Table 9, we can also infer the important role played by the ratio of M2
to the stock of international reserves (used to proxy the potential for resident-based capital
flight from the currency) in predicting out-of-sample currency crisis events. These results
are in line with the results of Berg and Pattillo (1999) which show the high information
content of this particular indicator both in terms of in sample and out of sample forecasting
performance.
5 Conclusions
In this paper we are interested in the out-of-sample predictability of balance of payment
crises in a number of East Asian countries. The currency turbulence is proxied by the
EM P index exceeding a given threshold. For this purpose we construct diffusion indices
summarising the information conveyed by external debt disaggregate data from the BIS
using principal components. Compared to a number of competing models, diffusion indices
forecasts obtained either through Probit modelling or through stochastic simulation of a
Dynamic Factor model (see, Stock and Watson, 2002, and Forni et al., 2005), provide superior
probability forecast performance. We also find that groups of variables, such as the ratio
of external borrowing from European countries to international reserves, or the ratio of the
money supply aggregate M2 , to the stock of international reserves play an important role in
explaining out-of-sample the dynamics of the EM P index. Finally, we argue that superior
probability forecast performance of the principal components model relative to competing
models can be explained by recognising the capability of the principal component analysis
in filtering out the noise associated with each variable in the dataset x.
References
[1] Bai, J. (2003):“Inferential Theory for Factor Models of Large Dimensions,” Economet-
rica, 71(1), 135-173.
16