Regional logistics forecasting based on hybrid model of RFECV and wavelet denoising

ObjectiveAccurate forecasting of regional logistics demand serves as a crucial basis for formulating regional logistics development strategies. A hybrid model was established for predictive analysis on regional logistics data to improve the precision of regional logistics forecasting in low-frequency high-dimensional data environments.MethodFirst, an annual regional logistics feature set for Liaoning Province was constructed. RFECV method was employed to select indicators influencing regional logistics in Liaoning. A wavelet denoising approach was introduced to further clean the selected indicators, thereby establishing an optimal feature dataset. Second, building upon RFECV approach, the predictions were conducted by using GM(1, N), BP, XGBoost, LSTM, and GRU models on both pre-denoising and post-denoising datasets to compare the predictive accuracy of different models. Finally, SHAP model was introduced to visualize the regional logistics prediction models, while ArcGIS was employed to visualize the strength of regional economic and logistics connections.ResultThe prediction performance after data denoising generally outperformed that before denoising. MAPE values across models decreased from 6%-9% to 5%-8%. The denoised dataset could improve model prediction accuracy in subsequent forecasts. XGBoost model demonstrated exceptional stability and explainability in its predictions both before and after denoising, achieving MAPE of 5.897% post-denoising. SHAP visualization analysis on XGBoost indicated that regional GDP contributed most significantly to logistics forecasting in Liaoning Province.ConclusionApplying the proposed approach, i.e., RFECV+wavelet denoising, to feature sets would improve the accuracy of model prediction. The stability and explainability of XGBoost prediction model provide further reference for regional logistics development.