ε-descending support vector machines for financial time series forecasting

This paper proposes a modified version of support vector machines (SVMs), called epsilon-descending support vector machines (epsilon-DSVMs), to model non-stationary financial time series. The epsilon-DSVMs are obtained by incorporating the problem domain knowledge - non-stationarity of financial time series into SVMs. Unlike the standard SVMs which use a constant tube in all the training data points, the epsilon-DSVMs use an adaptive tube to deal with the structure changes in the data. The experiment shows that the epsilon-DSVMs generalize better than the standard SVMs in forecasting non-stationary financial time series. Another advantage of this modification is that the epsilon-DSVMs converge to fewer support vectors, resulting in a sparser representation of the solution.