Stock Return Prediction Using Dual-Stage Attention Model with Stock Relation Inference

Abstract

Deep learning models have become widely accessible for stock prediction tasks. However, most of the research in this area focuses on only a single stock or an index and often formulates the problem to optimize only on the accuracy. Our paper proposed a more profit-oriented framework by formulating the problem into multiple stock returns prediction as well as introducing a relation inference for stock ranking. This setup can diversify investment and eventually enhance trading profits while maintaining the regression accuracy. Moreover, it is become more challenging to process multiple time-series features simultaneously because of the great variety of available information in the financial market. We mitigate this with the state-of-the-art model for time-series forecasting, the Dual-stage attention recurrent neural networks (DA-RNN), and train them with the shared-parameter model setting. The attention layer within DA-RNN helps the model captures the relevance insight among the features. We conducted experiments on major 64 target stocks from the SET market with RMSE, mean reciprocal rank, and annualized profit returns as evaluation metrics. The results show that our proposed model framework (DA-RANK) can predict multiple stock returns in ranking order and able to produce a desirable improvement in profitability over other baseline models.