Designing an early warning system for currency crises: an empirical treatment

Abstract

This article investigates the predictive power of qualitative response models that can serve as the basis for an early warning system for currency crises. It employs a signals framework that monitors the behaviour of key economic variables and issues a warning when their values exceed certain critical levels. The analysis is carried out for a parsimonious specification with high-frequency data. Taking Egypt as a case study, it is shown that this class of probability models captures to a good extent the turbulence in the foreign exchange market and the onset of crises.

Notes

¹ It is noteworthy that alternative practical approaches to identify currency crises are often used in country group studies. See, for example, Eichengreen et al. (1995, 1996), Kaminsky et al. (1998) and Mody and Taylor (2007) on the use of indices of pressures on the foreign exchange market to identify currency crises for a panel of countries.

² See Demirguc-Kunt and Detragiache (2000) for an alternative treatment of the crisis threshold using a loss function for the decision-maker.

³ See Hall (1994) and Ng and Perron (1995) on merits of the general-to-specific method for lag selection in unit root tests. See also Hayashi (2000, Ch. 6) on the properties of the sequential t-rule for the choice of lag length in autoregressions.

⁴ The dependent variable in the QR models is serially correlated as it takes the value 1 if there is a crisis sometime in the next 15 months.

⁵ Amemiya (1981) found that scaling by 1.6 instead produced a closer approximation. However, this works best at the center of the distribution. See, e.g., Greene (2003, p. 676).

⁶ The forecast error statistics are least favourable for the stock price indicator. But this may be expected since price movements on the capital markets are generally hard to predict with a high degree of accuracy.