A Neuro-wavelet Method for the Forecasting of Financial Time Series

We propose a wavelet neural network model (neuro-wavelet) for the short-term forecast of stock returns from high-frequency financial data. The proposed hybrid model combines the inherent capability of wavelets and artificial neural networks to capture non-stationary and non- linear attributes embedded in financial time series. A comparison study was performed on the modeling and predictive power among two traditional econometric models and four different dynamic recurrent neural network architectures. Several statistical measures and tests were performed on the forecasting estimates and standard errors to evaluate the predictive performance of all models. A Jordan net which used as input to the neural network the coefficients resulting from a non-decimated Haar wavelet-based decomposition of the high and low stock prices showed consistently to have a superior modeling and predictive performance over the other models. Reasonable forecasting accuracy for one, three, and five step-ahead horizons was achieved by the Jordan neuro-wavelet model.