基于TRMF-LSTM组合模型的多维时间序列预测

Multi-dimensional Time Series Forecasting Based on TRMF-LSTM Combined Model

近年来,长短时记忆网络(LSTM)在处理时间序列的非线性部分有着巨大优势,但单一预测模型无法同时兼顾数据的线性与非线性特性。针对此问题,引入时序矩阵分解技术TRMF处理多元时间序列的线性主体部分,计算得到训练数据的残差后,输入LSTM模型进行非线性拟合,再将测试数据代入到训练好的TRMF-LSTM模型,将模型预测的线性主体与残差相加,得到组合预测值。选取沪深300、上证指数两支股指以及三一重工、中国人寿、农业银行、牧原股份、美的集团、隆基股份6支个股共8支股票的股价时间序列进行预测,以LSTM、Transformer、SVR作为对比模型,并选取MAPE与RMSE两项评价指标。实验结果表明,相较于对比模型,MAPE和RMSE的最小值均落在TRMF-LSTM组合预测模型中,充分验证了模型的有效性。

Multidimensional time series forecasting has always been an important research direction for scholars. In recent years,long shortterm memory network(LSTM)has a huge advantage in processing the nonlinear part of time series. However,a single predictive model cannot take into account the linear and nonlinear characteristics of the data at the same time. To solve this problem,the time series matrix factorization technology TRMF(Temporal Regularized Matrix Factorization)is introduced to process the linear main part of the multivariate time series. After the residual of the training data is calculated,it is input into the LSTM model for nonlinear fitting. Then,the test data is substituted into the trained TRMF-LSTM model,and the linear subject predicted by the model is added to the residual error to obtain the combined prediction value. Select the Shanghai and Shenzhen 300,Shanghai Stock Exchange Index two stock indexes,SANY Heavy Industry,China Life,Agricultural Bank,Muyuan,Midea Group,Longji Stocks 6 stocks,a total of 8 stock price time series for forecasting,using LSTM,Transformer,SVR as Compare the models,and select two evaluation indicators,MAPE and RMSE. The experimental results show that,compared with the comparison model,the minimum values of MAPE and RMSE all fall in the TRMF-LSTM combined prediction model,which fully verifies the effectiveness of the model.