A deep learning model based on the combination of feature selection and self-attention mechanism is proposed in financial time series forecasting in view of the characteristics of financial time series data with large dimensionality, nonlinearity and strong time series correlation. In this paper, the filtering method, packing method, and embedded feature selection method are considered comprehensively, while PCA and Lasso are used for further compression and regularization to capture the important subset of features at a fine-grained level. Further, on this basis, LSTM is used to capture long-range correlations and learnable weights are employed to focus on each time point and dimension. In summary, the model proposed in this paper is trained in an end-to-end manner, and the combination of adaptive learning rate under the Adam optimizer enables the model to converge quickly. In practical tests, the method in this paper achieves good results for different financial market data and can be used in the field of financial time series forecasting.
