Neural network forecasting of the British Pound/US Dollar exchange rate

Neural networks have successfully been used for exchange rate forecasting. However, due to a large number of parameters to be estimated empirically, it is not a simple task to select the appropriate neural network architecture for an exchange rate forecasting problem. Researchers often overlook the effect of neural network parameters on the performance of neural network forecasting. This paper examines the effects of the number of input and hidden nodes as well as the size of the training sample on the in-sample and out-of-sample performance. The British pound/US dollar is used for detailed examinations. It is found that neural networks outperform linear models, particularly when the forecast horizon is short. In addition, the number of input nodes has a greater impact on performance than the number of hidden nodes, while a larger number of observations do reduce forecast errors.