Forecasting financial series using clustering methods and support vector regression

This paper proposes a two-stage model for forecasting financial time series. The first stage uses clustering methods in order to segment the time series into its various contexts. The second stage makes use of support vector regressions (SVRs), one for each context, to forecast future values of the series. The series used in the experiments is composed of values of an equity fund of a Brazilian bank. The proposed model is compared to a hierarchical model (HM) presented in the literature. In this series, the HM presented prediction results superior to both a support vector machine (SVM) and a multilayer perceptron (MLP) models. The experiments show that the proposed model is superior to HM, reducing the forecasting error of the HM by 32%. This means that the proposed model is also superior to the SVM and MLP models. An analysis of the construction and use of clusters associated with a series volatility study shows that data obtained from only one type of volatility (low or high) are enough to provide sufficient knowledge to the model so that it is able to forecast future values with good accuracy. Another analysis on the quality of the clusters formed by the model shows that each cluster carries different information about the series. Furthermore, there is always a group of SVRs capable of making adequate forecasts and, for the most part, the SVR used in forecasting is a SVR belonging to this group.