LINEAR AND NONLINEAR (NON-)FORECASTABILITY OF HIGH-FREQUENCY EXCHANGE-RATES

This paper forecasts Daily Sterling exchange rate returns using variou s naive, linear and non-linear univariate time-series models. The accu racy of the forecasts is evaluated using mean squared error and sign p rediction criteria. These show only a very modest improvement over for ecasts generated by a random walk model. The Pesaran-Timmerman test an d a comparison with forecasts generated artificially shows that even t he best models have no evidence of market timing ability.