Conducting stock market index prediction via the localized spatial–temporal convolutional network

Abstract

The stock index movement prediction is a fundamental research issue in financial investment. The high-noise and dynamic of stock data make this problem challenging. Efficiently modeling the underlying spatial–temporal correlations between indices is a potential study area for raising the prediction performance. Existing methods usually use separate modules to capture spatial and temporal correlations, which lose sight of local spatial–temporal correlation. In this paper, we propose a brand-new prediction model named the localized graph convolutional network (LoGCN) for stock index forecasting. This model can represent the intricate local spatial–temporal connections via a well-designed convolution mechanism. First, the localized graph construction method is proposed to model the local spatial–temporal correlations. Then, the localized graph convolutional module is presented to fuse the impact of different indices. Finally, the fully connected network is put forward to transform the features into the expected index trends. 42 Chinese stock market indices were selected to conduct extensive experiments. Three evaluation metrics including the increase rate regression, further trends classification, and stock market back-testing were adopted to the experimental comparison. The experiments show that our method is 4.2%, 3.1%, and 40% better than conventional methods under these three metrics.