Performance Evaluation of Some Neural Network Models Used in Forex Forecasting

Abstract

This paper evaluates the performance of two neural network models used in Forex forecasting; neural network with wavelet de-noising based model and neural network with error back propagation algorithm to find out the model accuracy and the ability of reducing error rates to the minimum. Experiments have been conducted to obtain four important parameters necessary to design the neural networks, these parameters are; optimal number of neurons in the hidden layer, optimal number of both learning rates and momentum constant, activation function, and suitable data division. The neural networks were designed, trained, validated and tested for the parameters. The results shows that neural network with wavelet de-noising based model has the capacity to learn the data quickly and provides better accuracy than neural network with error back propagation algorithm. However, the results shows that neural networks are very sensitive, they easily respond to changes made to their architecture. Furthermore, the results reveal that, the performance of neural networks in Forex forecasting increase with the increase in the number of economic variables as inputs to the networks.