Neural networks for predicting the daily exchange rate

Structural models that deal with daily FX forecasts rarely achieve the precision of random walk models. This paper implements a forecasting exercise with a set of forecasting models. The estimation sample starts from January 2004 to September 2008. The daily FX depreciation forecasts generated by a random walk model are compared with those generated by AR(p) models, p-lag perceptron models and fractional AR(p) models. The results show that perceptrons are able to anticipate the pattern of daily FX movements, specifically with an information set including lags of the bid-ask FX spread as a percentage of average daily quotations, the daily Yen/US$ FX depreciation and the daily US$/sol domestic interbank interest rate differential.