Mining Alpha Factors in Financial Markets using Multi-Frequency Timing Residual Networks

Time series forecasting in the stock market has always been a hot topic. Quantitative investment involves not only timing but also stock selection, asset allocation, and risk control. The choice among numerous potential investment targets is also crucial. Thanks to the robust time series feature extraction capability of Recurrent Neural Networks (RNN), combined with the multi-frequency data representation of stocks, the Multi-Frequency Time Series Residual Network (MFTR) proposed in this paper can simultaneously extract low-frequency and high-frequency data features. It uses the reciprocal of Pearson's correlation coefficient as the loss function. Experimental results show that the MFTR model has performed well in mining alpha factors in the China A-share market over the past five years. In a factor layered test, the annual return of the top layer reached 26.92%, representing a 46.46% increase over the daily frequency GRU model and a 22.36% increase over the GRU incremental learning model.