基于改进 Self-Attention 的股价趋势预测-《计算机技术与发展》

文章信息/Info

Title:: Research on Stock Price Trend Prediction Based on Self-Attention Model

Author(s):: ZHENG Shu-ting; XU Fei-fei; School of Computer Science and Technology,Shanghai University of Electric Power,Shanghai 200000,China

Keywords:: financial time series; stock price prediction; deep learning; self-attention; quantitative trading

摘要:: 近年来中国经济发展迅速,相应的,中国的金融市场也迅速发展,受到国内外投资者的关注,因此研究中国金融市场上股票价格趋势对学者、投资者和监管者具有重要的意义。随着量化交易等理念的兴起,越来越多的学者将深度神经网络(DNN)应用于金融领域。虽然近几年 DNN 在图像、语音以及文本等方面已经取得了极大的成功,但其在金融时间序列预测方面遇到了很多挑战,因为其数据本质上是高度动态性,且具有高噪声。作为 DNN 在时序数据处理的典型代表LSTM,由于该方法没有考虑不同时间点、不同来源数据的重要性程度,效果仍不理想。不同于在传统 LSTM 模型上引入Attention 机制,通过改进 Self-Attention 模型,分别对日线数据和分时线数据进行编码并融合,学习资金流变化对股票趋势变化的影响。实验结果表明,所提方法将对趋势判断的准确率提高到 63. 04% ,并在两个月的回测实验中获得了 6. 562% 的收益,证明了该模型在股价趋势预测上具有一定的有效性和实用性。

Abstract:: China’s economy has developed rapidly in recent years. Correspondingly,China’s financial market has also developed rapidly and attracted the attention of domestic and foreign investors. Therefore,studying the stock price trend in China’s financial market is of great significance to scholars,investors and regulators. With the rise of concepts such as Quantitative Trading,more and more scholars have applied deep neural networks (DNN) to the financial field. Although DNN has achieved great success in image,speech,and text in recent years,it has encountered many challenges in predicting financial time series because its data is highly dyn-amic in nature,and with high noise. LSTM is a typical representative of DNN in time series data processing, since this method does not take into account the importance of data from different time points and from dif-ferent sources,the effect is still not satisfactory. Different from the introduction of the Attention mechanism on the traditional LSTM model,by improving the Self-Attention model,the daily data and the time-sharing data are encoded respectively and fused,to learn about the effects of changes in capital flows on changes in stock trends. Experiment shows that the proposed method improves the accuracy of trend judgment to 63.04% and achieved 6.562% gain in two months of back testing experiments,which proves that the model is effective and practicality in predicting stock price trends.