Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (12): 249-259.DOI: 10.3778/j.issn.1002-8331.2112-0501

• Engineering and Applications • Previous Articles     Next Articles

Multi-Scale Stock Prediction Based on Deep Transfer Learning

CHENG Mengfei, GAO Shuping   

  1. School of Mathematics and Statistics, Xidian University, Xi’an 710126, China
  • Online:2022-06-15 Published:2022-06-15

基于深度迁移学习的多尺度股票预测

程孟菲,高淑萍   

  1. 西安电子科技大学 数学与统计学院,西安 710126

Abstract: The stock market is not only an important financing channel for listed companies, but also an important investment market. Stock prediction has always attracted people’s attention. In order to make full use of the information from different stock data and improve the prediction performance of stocks, a multi-scale stock price prediction model TL-EMD-LSTM-MA(TELM) is proposed. TELM decomposes the closing price into multiple time-scale components through empirical mode decomposition. Different time-scale components have different oscillation frequencies, reflecting different periodic information. Different components are predicted by different method according to oscillation frequencies. High-frequency component are predicted by the stacked LSTM which is trained by deep transfer learning method. Low-frequency component are predicted by the moving average method. The predicted closing price is calculated by summing the predicted values of all components. The stacked LSTM trained by deep transfer learning contains information from different stocks and has more industry or market knowledge, which can effectively reduce the predictive error. Using the moving average method to predict the low-frequency component can more effectively capture the general trend of stocks. 500 stocks in China’s A-share market, Shanghai stock index, SZSE component index, and other stock indexes are predicted. The results show that the error of TELM is the lowest and the goodness of fit is the highest compared with other models. By joining in the stock predicted closing price of TELM, the stock trading process is simulated. The results show that TELM has low investment risk and high return.

Key words: stock prediction, deep transfer learning, empirical mode decomposition, long short-term memory, moving average

摘要: 股票市场不仅是上市公司的重要融资渠道,也是重要的投资市场,股票预测一直受到人们的关注。为了充分利用来自不同股票价格的信息,提高股票的预测效果,提出一种多尺度股票价格预测模型TL-EMD-LSTM-MA(TELM)。TELM模型通过经验模态分解将收盘价分解为多个时间尺度分量,不同时间尺度分量震荡频率不同,反映了不同的周期性信息;根据分量的震荡频率选择不同方法进行预测,高频分量利用深度迁移学习的方法训练堆叠LSTM,低频分量利用移动平均法进行预测;将所有分量的预测值相加作为收盘价的最终预测输出。通过深度迁移学习训练的堆叠LSTM,包含来自不同股票的信息,具备更多行业或市场的知识,能有效降低预测误差。利用移动平均法预测低频分量,更有效捕获股票的总体趋势。对中国A股市场内500支股票以及上证指数、深证成指等指数进行预测,结果表明,与其他模型相比,TELM预测误差最低,拟合优度最高。根据TELM预测的股票收盘价模拟股票交易过程,结果表明TELM投资风险低、收益高。

关键词: 股票预测, 深度迁移学习, 经验模态分解, 长短期记忆网络, 移动平均方法