关键词: 股票价格预测, 时间序列, 深度学习, Transformer, 注意力机制

Abstract: Stock price prediction is a key topic of common concern for financial research and quantitative investment, and the use of deep learning techniques to reveal the market patterns of stock markets has become a hot research topic in recent years. Most of the existing deep learning models for stock price prediction only study point-in-time data, and this structural shortcoming causes them to fail to reflect the impact of the cumulative effect of feature factors on stock prices. To address this, a Transformer-based stock price forecasting model Stockformer is proposed by redesigning the model to handle time series data. it mines the time-series dependence between stock prices and feature factors through a causal self-attentiveness mechanism, employs a trend enhancement module to provide the model with the trend features of the series, and uses encoder-specific inputs to provide the prediction direct a priori information of the input features. The experimental results show that the prediction accuracy of Stockformer is significantly better than that of existing deep learning models, and the average absolute error and root mean square error are reduced by 23.2% and 25.7%, respectively, compared with the classical Transformer prediction model, and the predicted values are more suitable to the real values; and the ablation experiments are conducted to evaluate the causal attention mechanism of Stockformer, the effects of the time-series feature extraction means and specific model inputs are evaluated by ablation experiments respectively, and the superiority and generalizability of the proposed model are verified.

Key words: stock price prediction, time series, deep learning, Transformer, attention mechanism