Exchange rate prediction: a wavelet-neural approach

The paper deals with the use of artificial neural networks (ANN) for predictions of economic time series. First, we revise the basic existing ANN architectures for time series forecasting and describe their application on CZK/EUR exchange rate prediction. Next, we introduce a hybrid version of the ANN that builds upon the same network strategy but tries to enhance the prediction accuracy by first decomposing the signal into individual frequency bands and then training the perceptron parameters on each frequency band separately. The time-frequency signal analysis is obtained using discrete wavelet transform. Different signal segments in various levels of decomposition are reconstructed separately. The resulting signals (in time domain again) are samples for ANNs as above. Finally, we compare the accuracy of the hybrid approach to that of the traditional ANN setting for the case of CZK/EUR exchange rate data.