The degree of accuracy of express packaging demand prediction is an important foundation for the construction of intelligent logistics systemization. The work aims to compare different machine learning methods, different feature selection methods, and different parameter optimization methods, to select a more accurate method of express packaging demand prediction, analyze the impact of the main features on demands of express packaging, and optimize the study of regional express packaging demand. Firstly, through comparison with different machine learning methods, different feature selection methods, and different parameter optimization methods, the effect of genetic algorithm on optimizing parameters of the random forest (RF) model was determined. Finally, in order to better explain the model, the method of SHAP analysis was introduced to analyze the importance of different features. The results showed that the improved random forest prediction model was the most effective with MAE value, MAPE value, RMSE value and R2 of 2 783, 5.1%, 4 343 and 0.99 respectively, and that female population and tertiary industry value were the most critical factors affecting the demand for express packaging. The results show that the proposed prediction method has better accuracy and interpretability, and can provide powerful decision support for express packaging demand prediction.