Thai stock return prediction using deep learning models with stock indicators and textual features

Abstract

Stock prediction task is notoriously challenging due to the uncertainty and dynamic external factor which influence the stock behavior. Recently, Deep learning research is gaining popularity on this task but often focuses on only a particular type of data; numeric indicators or textual information. Moreover, most researches focus on only a single stock or a market index. In this paper, we aim to predict multiple stock returns using both types of data. The model consists of dual-stage attention recurrent neural network, our proposed stock relation inference framework, and textual features integration. The proposed stock relation inference help tackles multiple time-series features (stocks indicators such as fundamentals or technicals) as well as add the ranking ability to the model from the combined ranking loss function. We demonstrate how to represent textual features with the hierarchical neural network and the BERT embedding method. Finally, we explore the approaches to integrate all elements in order to handle both types of data effectively, aiming to improve the performance of the stock prediction task.