关于LSTM-TCN模型结合递归特征消除法的股票预测

Stock Prediction Based on LSTM-TCN Model Combined with Recursive Feature Elimination Method

摘要: 近年来，神经网络深度学习算法在金融领域研究中一直得到广泛应用。为了帮助股民深刻了解股票行情进而实现合理化投资，在融合长短时记忆网络(LSTM)和时序卷积神经网络(TCN)的优势构建的LSTM-TCN模型基础上，引入递归特征消除法，通过反复训练得到可以明显降低预测误差的特征向量组合作为LSTM-TCN模型的输入；其次利用LSTM网络提取数据局部时间特征，再将输出的特征矩阵输入TCN网络进一步挖掘长时间特征，从而构成LSTM-TCN联合模型来对股票收盘价进行预测，可实现时序特征的充分挖掘，获取更多的信息。中国平安股票数据的预测结果表明：LSTM-TCN联合模型的RMSE为0.773，MAPE为1.05%，均小于其他传统机器学习模型以及LSTM、TCN模型，说明了提出的LSTM-TCN联合模型在股价预测方面具有更高的预测精度。

Abstract: In recent years, neural network deep learning algorithm has been widely used in the research of fi-nancial field. In order to help investors deeply understand the stock market and realize rational investment, the Recursive Feature Elimination method is introduced on the basis of LSTM-TCN model established by combining the advantages of Long-Term And Short-Term Memory Network (LSTM) and Time-Series Convolution Neural Network (TCN), and the combination of feature vector that can significantly reduce prediction error is obtained through repeated training as the input of LSTM-TCN model; secondly, the LSTM network is used to extract the local time features of the data, and then its output feature matrix is input into the TCN network to further mine the long-time fea-tures, so as to form the LSTM-TCN joint model to predict the stock closing price, which can fully mine the time series characteristics and obtain more information. The prediction results of Ping An’s stock data show that the RMSE and MAPE of the LSTM-TCN joint model are 0.773 and 1.05% respec-tively, which are lower than those of other traditional machine learning models, LSTM and TCN models, indicating that the LSTM-TCN joint model has higher prediction accuracy in stock price pre-diction.