Abstract: To address the problems of existing research on the inadequate ability to capture short-term change patterns of financial time series data and poor prediction accuracy, a financial time series prediction model（dual-attention MDWT-CVT-LSTM） based on dual-pathway attention mechanism and improved transformation-gated LSTM（VT-LSTM） is proposed. Firstly, this paper decomposes the stock index series using multi-level discrete wavelet transform（MDWT） to obtain high and low frequency data, and adds the transformation-gated mechanism to the LSTM of fusion gating unit to construct the VT-LSTM, which can effectively grasp short-term mutation information. Then, VT-LSTM is combined with one-dimensional temporal convolution（Conv1D） in a dual-pathway attention network to extract spatial local features and temporal features of different frequency data respectively, and finally, each sub-series is predicted to realize a multi-level and multi-pathway prediction study. The experimental comparison of different models on financial stock market index datasets and individual stocks datasets shows that the proposed model has better prediction accuracy than other methods and good feasibility.

Key words: financial time series, dual-attention mechanism, time-series convolution, multilevel discrete wavelet transform, long and short time memory networks

摘要： 针对现有研究对金融时序数据短期变化规律捕捉能力不足和预测精度不佳的问题，提出一种基于双通路注意力机制和改进转换门控LSTM（variant transformation-gated LSTM，VT-LSTM）的金融时序预测模型（dual-attention MDWT-CVT-LSTM）。使用多级离散小波变换（MDWT）分解股指序列得到高频和低频数据，并在融合门控单元的LSTM中加入转换门控机制，构造VT-LSTM，其能有效把控短期突变信息。在双通路注意力网络中结合VT-LSTM与一维时序卷积（Conv1D），分别提取不同频度数据的空间局部特征和时序特征，对各子序列进行预测，实现多层级多通路的预测研究。在金融股指数据集和个股数据集上对不同模型进行实验比较，结果表明提出模型预测精度优于其他方法，有良好的可行性。

关键词: 金融时间序列, 双通路注意力机制, 时序卷积, 多级离散小波变换, 长短时记忆网络