Return forecasting of sub-new stocks via ADARNN Trained with Mixed Historical Data

Forecasting the return rates of sub-new stocks has always been a challenging problem. Typically, machine learning methods such as support vector machines and neural networks require a large amount of training data to ensure the effectiveness of the model. However, due to the limited historical trading data for sub-new stocks, these methods often fail to achieve satisfactory results. Based on the adaptive recurrent neural network (AdaRNN) model, a time concatenation method based on equal segmentation (ES-STC) is proposed to forecast the return rates of sub-new stocks. Our approach mixes old and sub-new stock data using a unique data fusion strategy to improve the model's performance on sub-new stocks. The feasibility of this method in forecasting the return of sub-new stocks is verified through experiments, and compared with baseline models such as ALSTM, Transformer, LSTM, and GRU. The results indicate that the performance of the ES-STC surpasses those of other baseline models. For example, compared to the ADARNN, our method shows an average improvement of 9.9% in the ICIR indicator and 3.7% in the IC indicator. Our findings reveal that: 1) Combining sub-new and old stocks in certain sectors (e.g., automotive and news media) can enhance the predictive performance of the model. 2) The ES-STC is better than ADARNN and other benchmark models in certain sectors (e.g., automotive and news media).