Optimal signal multi-resolution by genetic algorithms to support artificial neural networks for exchange-rate forecasting

Abstract

Detecting the features of significant patterns from historical data is crucial for good performance in time-series forecasting. Wavelet analysis, which processes information effectively at different scales, can be very useful for feature detection from complex and chaotic time series. In particular, the specific local properties of wavelets can be useful in describing the signals with discontinuous or fractal structure in financial markets. It also allows the removal of noise-dependent high frequencies, while conserving the signal bearing high frequency terms of the signal. However, one of the most critical issues to be solved in the application of the wavelet analysis is to choose the correct wavelet thresholding parameters. If the threshold is small or too large, the wavelet thresholding parameters will tend to overfit or underfit the data. The threshold has so far been selected arbitrarily or by a few statistical criteria. This study proposes an integrated thresholding design of the optimal or near-optimal wavelet transformation by genetic algorithms (GAs) to represent a significant signal most suitable in artificial neural network models. This approach is applied to Korean won/US dollar exchange-rate forecasting. The experimental results show that this integrated approach using GAs has better performance than the other three wavelet thresholding algorithms (cross-validation, best basis selection and best level tree).