Forecasting Stock Return Based on the Multifactor Dynamic Attention Network

The application of multifactor models for forecasting stock returns has become a relatively mature practice in the financial field. The assumption of no collinearity among factors in some existing research may lead to the loss of stock information. In order to address this issue, a novel multifactor model, the multifactor dynamic attention network (MFDAN), is proposed, which includes a multiattention dynamic network and temporal dependence modeling. In the part of the multiattention dynamic network, a two‐dimensional attention mechanism is constructed. This mechanism assigns attention weights to each factor at different time steps without taking into account the constraint condition of collinearity between factors. Simultaneously, it addresses the potential ineffectiveness of certain factors at specific time steps and dynamically measures the influence of the factor on the stock price. In the part of temporal dependence modeling, the long short–term memory (LSTM) deep neural network model is used to model the data assigned by attention weights. The experiment results on the selected 30 stocks indicate that the prediction accuracy of the stock price for the MFDAN model achieved 90.42%. Compared to the traditional multifactor models and other neural network models such as LSTM, GRU, and LSTM + GRU, the prediction accuracy is improved by more than 20%. The proposed model can provide technical support for forecasting stock returns.