Long Short-Term Memory Networks with Multiple Variables for Stock Market Prediction

Long short-term memory (LSTM) networks have been successfully applied to many fields including finance. However, when the input contains multiple variables, a conventional LSTM does not distinguish the contribution of different variables and cannot make full use of the information they transmit. To meet the need for multi-variable modeling of financial sequences, we present an application of multi-variable LSTM (MV-LSTM) network for stock market prediction in this paper. The network consists of two serial modules: the first module is a recurrent layer with MV-LSTM as its recurrent unit, which is able to encode information from each variable exclusively; the second module employs a variable attention mechanism by introducing a latent variable and enables the model to measure the importance of each variable to the target. With these two modules, the model can deal with multi-variable financial sequences more effectively. Moreover, a statistical arbitrage investment strategy is constructed based on the prediction model. Extensive experiments on the large-scale Chinese stock data show that the MV-LSTM network has a higher prediction accuracy and provides a better statistical arbitrage investment strategy than other methods.