基于多尺度特征和注意力的金融时序预测方法

Financial Time Series ForecastingMethod Based onMulti-Scale Features and AttentionMechanism

金融时间序列预测是经济领域中一个非常重要的实际问题,然而,由于金融市场的噪声和波动性,当前存在方法的预测精度尚不能令人满意。为了提高金融时间序列的预测精度,提出了一种融合扩张卷积神经网络(dilated convolutional neural network,DCNN)、长短时记忆神经网络(long short term memory,LSTM)和注意力机制(attention mechanism,AT)的混合预测模型DCNN_LSTM_AT。该模型由两个部分组成:第一部分包含扩张卷积神经网络和基于LSTM的编码器,其功能在于提取原始序列数据中不同时间尺度的有效信息;第二部分由带注意力机制的LSTM解码器构成,其功能在于对第一部分提取的信息进行过滤并利用过滤后的信息进行预测。最后将所提模型在3支股指数据集和3支个股数据集上进行实验,并与其他常见的基准模型进行了对比,实验结果表明该模型相比于其他模型具有更好的预测精度和稳定性.

Financial time series forecasting is a very important practical problem in the economic field.However,due to the noise and volatility of financial markets,the forecasting accuracy of current existing methods is not yet satisfactory.In order to improve the prediction accuracy of financial time series,a hybrid prediction model DCNN_LSTM_AT integrating dilated convolutional neural network(DCNN),long short term memory(LSTM)and attention mechanism(AT)is pro-posed.The model consists of two parts:the first part contains a dilated convolutional neural network and an LSTM-based encoder,whose function is to extract effective information at different temporal scales in the original sequence data.The second part is composed of an LSTM decoder with an attention mechanism,its function is to filter the information extracted from the first part and use the filtered information to make predictions.Finally,the proposed model is tested on 3 index datasets and 3 individual stock datasets,and compared with other common benchmark models.The experimental results show that the model has better prediction accuracy and stability than other models.