MAN: A Multidimension Attention-Based Recurrent Neural Network for Stock Price Forecasting

Recently, deep learning has been applied to many fields, including finance. One limitation of existing neural network-based solutions is that their optimization target is the accuracy of the classification or regression task, rather than the ultimate return on investment. Another limitation is that they often treat stocks as independent entities, ignoring the interrelationships among them that are informative for stock forecasting. We put forward a time-feature-stock-wise attention-based RNN to consider the relationship between stocks and distill valuable information from multi-feature temporal data to forecast the stock price. At the same time, we take both the prediction accuracy and the sorting error of the forecast result as the optimization goal to maximize the final yield. To validate our method, we use our model to predict the opening prices of 10, 20, and 40 stocks randomly selected from Chinese stock market and made investment based on the prediction results. The back-test results showed that the return rate of this model was better than that of traditional stock forecasting solutions.