ICA-DEODA: An independent feature extraction model for stock index forecasting

In this paper, an integration of the independent component analysis (ICA) and an unsupervised feature selection scheme is proposed as a data engineering approach, called ICA-DEODA. The proposed feature extraction model enables beyond second order analysis which can powerfully underlie knowledge discovery methods. When applying the model for stock forecasting, firstly, independent components (ICs) are extracted from a set of influential factors of Security Exchange of Thailand (SET) using ICA. DEODA is then applied to select a number of highly informative independent components for being inputs of the forecasting model. Experimentally, the support vector regression (SVR) was taken for predicting movement of the Stock Exchange of Thailand (SET) using quarterly data. The results demonstrated that the proposed model outperforms the traditional PCA-SVR, SVR and also PCA-DEODA-SVR model.