Forecasting stock indices with wavelet domain kernel partial least square regressions

Financial time series are nonlinear and non-stationary. Most financial phenomena cannot be clearly characterized in time domain. Therefore, traditional time domain models are not very effective in financial forecasting. To address the problem, this study combines wavelet analysis with kernel partial least square (PLS) regressions for stock index forecasting. Wavelet transformation maps time domain inputs to time-frequency (or wavelet) domain, where financial characteristics can be clearly identified. Because of the high dimensionality and heavy multi-collinearity of the input data, a wavelet domain kernel PLS regressor is employed to create the most efficient subspace that maintains maximum covariance between inputs and outputs, and to perform final forecasting. Empirical results demonstrate that the proposed model outperforms traditional neural networks, support vector machines, GARCH models, and has significantly reduced the forecasting errors.