Comparing Statistical and Machine Learning Methods for Time Series Forecasting in Data-Driven Logistics-A Simulation...

Area of Science:

• Operations Research
• Supply Chain Management
• Data Science

Background:

• Logistics and supply chain management heavily rely on accurate forecasting of time-dependent factors.
• The quality of planning and decision-making is directly linked to forecast accuracy.
• Existing research often focuses on case-specific forecasting, limiting generalizability.

Purpose of the Study:

• To compare the performance of various state-of-the-art forecasting methods.
• To provide general recommendations for forecasting in logistics and supply chain management.
• To evaluate methods across a broad set of simulated time series representing diverse scenarios.

Main Methods:

• Simulation of various linear and nonlinear time series.
• Comparison of different forecasting techniques, including machine learning and time series approaches.
• Evaluation of forecasting performance under different noise levels and complexities.

Main Results:

• Machine learning methods, especially Random Forests, demonstrated superior performance in complex scenarios.
• Differentiated time series training significantly enhanced the robustness of machine learning models.
• Traditional time series approaches remained competitive in low-noise environments.

Conclusions:

• Random Forests offer robust forecasting solutions for complex logistics and supply chain challenges.
• The choice of forecasting method should consider the complexity and noise level of the time series data.
• Generalizable recommendations for forecasting can be derived from broad simulation studies.