Research on an Integrated Model for Logistics Demand Forecasting Based on Random Forests and LSTM

Abstract

This paper proposes a logistics demand forecasting model integrating Support Vector Machines (SVM), Random Forests, and Long Short-Term Memory (LSTM) networks, focusing on the construction logic of this ensemble model and methods for integrating prediction results. First, SVM maps raw features into a high-dimensional space, combining sparse linear criteria for feature selection to preserve nonlinear representation capabilities while controlling model complexity. Second, enhanced features and lagged terms are fed into both the random forest and LSTM. The random forest enhances robustness through multi-tree voting/averaging, while the LSTM captures temporal dependencies via its gated structure, ensuring information consistency. Finally, prediction results are adaptively weighted and integrated using an inverse error criterion based on model performance on the validation set, followed by parameter tuning and cross-validation. This model supports logistics inventory management decisions by integrating the strengths of three algorithms. It effectively handles nonlinear and time-series data, enhances prediction accuracy and generalization capabilities, and demonstrates strong robustness against outliers and missing values.