A multi-indicator feature selection for CNN-driven stock index prediction

Abstract

Stock index prediction is regarded as a challenging task due to the phenomena of non-linearity and random drift in trends of stock indices. In practical applications, different indicator features have significant impact when predicting stock index. In addition, different technical indicators which contained in the same matrix will interfere with each other when convolutional neural network (CNN) is applied to feature extraction. To solve the above problem, this paper suggests a multi-indicator feature selection for stock index prediction based on a multi-channel CNN structure, named MI-CNN framework. In this method, candidate indicators are selected by maximal information coefficient feature selection (MICFS) approach, to ensure the correlation with stock movements while reduce redundancy between different indicators. Then an effective CNN structure without sub-sampling is designed to extract abstract features of each indicator, avoiding mutual interference between different indicators. Extensive experiments support that our proposed method performs well on different stock indices and achieves higher returns than the benchmark in trading simulations, providing good potential for further research in a wide range of financial time series prediction with deep learning based approaches.