epsilon-Descending Support Vector Machines for Financial Time Series Forecasting

This paper proposes Ɛ-descending support vector machines (Ɛ - DSVMs) to model non-stationary financial time series. The Ɛ -DSVMs are obtained by taking into account the problem domain knowledge of nonstationarity in the financial time series. Unlike the original SVMs which use the same tube size in all the training data points, the Ɛ-DSVMs use the tube whose value decrease from the distant training data points to the recent training data points. Three real futures which are collected from the Chicago Mercantile Market are examined in the experiment, and it is shown that the Ɛ-DSVMs consistently forecast better than the original SVMs.