Multi-Source Aggregated Classification for Stock Price Movement Prediction

Abstract

Stock price movement prediction remains a challenging research task. Existing studies mainly rely on numerical features of the target stock and news sentiment; however, semantics-based sentiment analysis cannot accurately capture real market sentiment. In addition, using only information about the target company is insufficient for explaining stock price fluctuations, because the price of a target stock may also be affected by related companies. To address these limitations, this paper proposes a Multi-source Aggregated Classification (MAC) method for stock price movement prediction. The proposed method jointly exploits numerical features of the target stock, market-driven news sentiment, and news sentiment from related stocks. To extract real market sentiment from news more effectively, we design a pretraining task for an embedding generator supervised by actual stock price movements, enabling the embeddings produced by the pretrained sentiment classifier to encode market sentiment information in vector space. Furthermore, MAC introduces a graph convolutional network to model the influence of news about related companies on the target stock and then predicts the stock price movement on the next trading day. Experimental results show that MAC outperforms the baselines considered in this study in stock price movement prediction, Sharpe ratio, and backtesting return.