Stock Price Prediction Using BiLSTM and a Modified Transformer

Abstract

Stock price prediction remains a fundamental problem in financial forecasting. To improve predictive accuracy and stability, this paper proposes a bidirectional long short-term memory-modified Transformer-temporal convolutional network (BiLSTM-MTRAN-TCN). The Transformer is first redesigned and combined with a temporal convolutional network (TCN) to form a modified Transformer-temporal convolutional network (MTRAN-TCN). The resulting module is then integrated with a bidirectional long short-term memory network (BiLSTM) to construct a hybrid forecasting architecture. The proposed model combines the global dependency modeling capability of the Transformer, the bidirectional temporal feature extraction ability of BiLSTM, and the sequence modeling and generalization advantages of TCN. Experiments on five stock indices and 14 A-share stocks are conducted to evaluate both the modified Transformer and the contribution of the BiLSTM component. The proposed model achieves the best average performance in the stock-index experiments, attains the highest R 2 on 85.7% of the stock datasets, and yields the lowest RMSE on 78.6% of the stock datasets. Its performance also remains relatively stable across different time periods. These results indicate that BiLSTM-MTRANTCN offers strong accuracy, stability, and generalization ability for stock price prediction.