Modelling and forecasting exchange-rate volatility with ARCH-type models

The statistical analysis of short-run exchange-rate data shows that there is strong heteroskedasticity and serial dependence of volatility. In addition, the empirical distributions are leptokurtic. The model of generalized autoregressive conditional heteroskedasticity (GARCH) seems to be ideally suited to model these empirical regularities because the model incorporates autocorrelated volatility explicity and it also implies a leptokurtic distribution. The GARCH model does indeed achieve a reasonably good fit to the exchange-rate data. However, the GARCH model is not able to outperform the naive forecasts of volatility which use the current estimate of the variance from the past data.