A BiLSTM–MLP–ReCast Method for Stock Price Forecasting

As financial market conditions continue to evolve, stock price sequences often exhibit strong nonlinear and nonstationary characteristics, reducing the effectiveness of traditional forecasting approaches. This paper proposes a stock price forecasting model based on the BiLSTM-MLP-ReCast framework. The model employs BiLSTM to capture bidirectional temporal dependencies in price sequences, integrates an MLP at the output stage for nonlinear transformation and introduces the ReCast mechanism to refine residual components, thereby mitigating lag effects and systematic bias. The framework is trained in an end-to-end manner, enabling joint optimization of temporal feature extraction, nonlinear mapping, and residual adjustment. Experimental results on three real-world stock datasets, including Google, Amazon, and Microsoft, demonstrate the effectiveness of the proposed framework. The complete BiLSTM-MLP-ReCast model achieved an R2 of 0.9068 and an RMSE of 0.0520 on the Microsoft dataset. In addition, the model achieved the lowest MAPE values of 5.20% on Microsoft and 8.88% on Amazon, indicating improved prediction accuracy and robustness compared with conventional baseline models.