Predictability of foreign exchange rates : On the chaos ’ routes

Criteria usually used by economists to evaluate the accuracy of forecasts generated by a model are based on mean and quadratic errors. The ability to give directional forecast, raise or drop, is not very used whereas it is of big practical interest. If a model was able to give good directional forecasts with a rate greater than 50% over long periods, trading strategies with excess return could exist. Using this criterion, the predictability of daily major foreign exchange rates is examined. The forecasting model we used is non-parametric and based on nearest neighbors. First of all, a description of the data is done. We focus on possible chaotic properties and independence of the data. First results illustrate the need to select few neighbors and an embedding dimension adapted to the data, the relevance of the directional forecast criterion compared to MAE and RMSE. Ten years of daily forecasts, from 1989 to 1998, were computed. If it is not possible to find a model with significant results all over the period, some sub periods show results far from the behaviour of a random walk. Deutsche Mark and British Pound sign forecasting rate for the year 1991 is greater than 57%.