Testing Currency Predictability Using An Evolutionary Neural Network Model

Two alternative learning approaches of a MLP Neural Network architecture are employed to forecast foreign currencies against the Greek Drachma, a Back-Propagation with a hyperbolic tangent activation scheme and an evolutionary trained model. Four major currency data series, namely the U. S. Dollar, the British Pound, the French Franc and the Deutsche Mark, are used in this forecasting experiment. Extended simulations have shown a high predictive ability, which is significantly better when using the actual rates compared to using the logarithmic returns of each series. The genetic algorithm performs best on FF and DM, while the back-propagation on USD and BP.