基于注意力机制的循环神经网络对金融时间序列的应用

Application of attention mechanism based recurrent neural network in financial time series

金融时间序列由于高噪声性以及序列间的相关性,导致传统模型的预测精度和泛化能力往往较低。为了克服这一问题,提出一种基于注意力机制的循环神经网络预测模型。利用金融时间序列的技术指标作为特征序列,通过GRU得到隐藏状态,再利用注意力机制将其重构,并与目标序列一起作为新的GRU的输入,提高目标序列的预测效果。利用上证综指数据,分别使用加入注意力机制的门控循环网络与标准门控循环网络进行预测。在6个输入维度的情况下,基于注意力机制的GRU模型平均绝对百分比误差为0.76%,低于标准GRU模型的0.90%;在48个输入维度的情况下,基于注意力机制的GRU模型平均绝对百分比误差为0.73%,低于标准GRU模型的1.61%。结果表明,加入注意力机制后门控循环网络的预测效果得到提升,并且在特征序列的输入维度增大时,其预测效果提升更为明显。

Due to the high noise of financial time series and the correlation between the series,the prediction accuracy and generalization ability of the traditional models are often low.In order to overcome these problems,a recurrent neural network prediction model based on the attention mechanism is proposed in this paper.By taking the technical index of financial time series as the feature sequence,the hidden state is obtained by the gated recurrent unit(GRU).The obtained hidden state is reconstructed with the attention mechanism.Together with the target sequence,it is used as the input of a new GRU to improve the prediction effect of the target sequence.With the SSE composite index data,the GRU with the attention mechanism and the standard GRU are used to make predictions.In the case of 6 input dimensions,the average absolute percentage error of the GRU model based on the attention mechanism is 0.76%,which is lower than 0.90%of the standard GRU model.In the case of 48 input dimensions,the average absolute percentage error of the GRU model based on the attention mechanism is 0.73%,which is lower than 1.61%of the standard GRU model.The results show that the prediction effect of the gated recurrent network with attention mechanism is improved,and when the input dimension of the feature sequence increases,the improvement of the prediction effect is more obvious.