Enterprise logistics cost prediction model based on machine learning

Enterprise logistics cost forecasting faces challenges such as low prediction accuracy and difficulty quantifying the coupled relationships between multi-dimensional cost factors. This study proposes a hybrid model based on XGBoost and LSTM, integrating static feature interactions with dynamic time series modeling through a stacking integration framework. Feature engineering encompasses 12 dimensions, including transportation volume, fuel price fluctuations, and seasonality, and introduces sliding window statistics and Fourier series fitting to address time series dependencies. The model utilizes Bayesian optimization for parameter tuning and features are interpreted using SHAP metrics. The model achieves excellent performance in terms of RMSE (78.2 thousand), MAE (60.3 thousand), and R² (0.94). Ablation experiments demonstrate that historical cost statistics (RMSE increases by 23.8% after removal) and transportation volume (RMSE increases by 23.1% after removal) are key drivers. This model provides a quantitative decision-making basis for enterprise cost control.