基于Attention机制的GRU股票预测模型

An Attention-based GRU Model for Stock Predicting

股市行情随众多影响因子不断变化,现有基于时间序列预测的方法难以捕捉该非线性动力系统的复杂变化规律,预测效果并不理想。本文提出基于Attention机制的GRU预测模型,捕捉关键时间点特征信息以解决对时间特征不敏感导致预测精度不高的问题以提升预测精确度。首先使用LSTM和GRU构建基础预测模型;然后对输入特征进行统计处理和筛选,选取更能反映股价变动规律的特征;最后基于编码器-解码器框架,在GRU模型上加入Attention机制,使模型聚焦于重要时间点的股票特征信息。本文在科大讯飞股票数据上进行实验,实验结果表明基于Attention机制的GRU模型在MAPE,RMSE,R2 score三个评价指标上均优于其他模型,Attention机制能够捕捉重要时间点局部特征,对预测模型的优化是可行和有效的。

The stock market is constantly changing with many influencing factors.The existing time series forecasting method is difficult to capture the changing law of the nonlinear dynamic system,of which the performance is not ideal.In this paper,the attention mechanism is used to optimize the GRU model,so that the model would give higher weights to the data at important time points to improve the prediction effect.Firstly,the LSTM and GRU model have been used to construct the basic prediction model respectively;and then the input features arc processed and filtered to select the features that better reflect the law of stock price changes;finally, based on the encoder-decoder framework, the attention mechanism has been added on the GRU model,which can focus on the stock information at important time points.Experiments on the stock data of HKUST,the GRU model based on Attention is superior to other models in the three evaluation indicators of MAPE, RMSE, R2 score.The attention mechanism can capture important time point features,which is feasible and effective to optimize the prediction model.