A Study on Stock Price Forecasting in Transportation Industry Based on CNN-Attention Mechanism and SHAP-RFE Taking 18 Stocks in Different Areas of the Transportation Industry as an Example

The transportation industry plays a critical role in resource circulation, with stock prices highly sensitive to oil prices, macroeconomic shifts, policies, and emergencies. Traditional linear models are limited in capturing nonlinear and noisy dynamics, while deep learning, though accurate, often lacks interpretability and robustness. This paper develops a CNNMulti-Head Attention framework that models logarithmic returns with Huber loss, adaptive learning rate, and early stopping to mitigate outliers. SHAP combined with Recursive Feature Elimination enhances feature selection and interpretability, and posterior calibration corrects systematic bias before conversion to price space. Empirical results demonstrate lower RMSE and MAE, higher R, and improved generalization. The proposed framework provides a reproducible, interpretable, and robust approach to stock forecasting, offering practical implications for investment and regulatory supervision.