Multi-task Recurrent Neural Networks and Higher-order Markov Random Fields for Stock Price Movement Prediction

Stock price movement not only depends on the history of individual stock movements, but also complex hidden dynamics associated with other correlated stocks. Despite the substantial effort made to understand the principles of stock price movement, few attempts have been made to predict movement direction based upon a single stock's historical records together with its correlated stocks. Here, we present a multi-task recurrent neural network (RNN) with high-order Markov random fields (MRFs) to predict stock price movement direction. Specifically, we first design a multi-task RNN framework to extract informative features from the raw market data of individual stocks without considering any domain knowledge. Next, we employ binary MRFs with unary features and weighted lower linear envelopes as the higher-order energy function to capture higher-order consistency within the same stock clique (group). We also derive a latent structural SVM algorithm to learn higher-order MRFs in a polynomial number of iterations. Finally, a sub-gradient algorithm is employed to perform end-to-end training of the RNN and high-order MRFs. We conduct thorough empirical studies on three popular Chinese stock market indexes and the proposed method outperforms baseline approaches. To our best knowledge, the proposed technique is the first to investigate intra-clique relationships with higher-order MRFs for stock price movement prediction.