Machine Learning for Demand Forecasting in U.S. Third-Party Logistics: A Systematic Review and Meta-Analysis of Predictive Performance

Abstract

Demand forecasting in United States third-party logistics (3PL) remains difficult because volatile shipment volumes, intermittent orders, fragmented information, and inconsistent evaluation metrics complicate model selection. This study examines how machine learning approaches compare with traditional statistical methods and identifies the data and operating conditions shaping predictive performance. Guided by Organizational Information Processing Theory, the review considers whether access to integrated demand information matches uncertainty and service coordination requirements across logistics environments. It uses a cross-sectional, case-based systematic literature review guided by PRISMA 2020, with quantitative comparisons of published forecasting results and meta-analysis conditional on comparability. The review targets 100 studies published during 2005–2023, covering U.S. 3PL operations and distinguishable supporting freight, warehousing, distribution, and supply-chain cases; final eligibility counts remain pending. Variables include forecasting architecture, historical demand, external predictors, forecast horizon, demand volatility, operational complexity, and errors measured using root mean squared error and related indicators. The analysis classifies studies, assesses methodological comparability, synthesizes themes, and compares model performance against aligned benchmarks. In a published 3PL shipment case, hybrid CNN–LSTM attained RMSE 59.12 versus 62.32 for ARIMA, 59.96 for standalone CNN, 66.88 for LSTM, and 72.64 for a regression tree. Another transportation study covering 54 Brazilian distribution centers found LSTM better than statistical alternatives in approximately 94% of dispatching units. In an adjacent U.S. retail benchmark comprising 42,840 hierarchical series, the leading approach improved accuracy by 22.4% relative to its strongest comparator. Evidence indicates context-specific advantages for hybrid models, but standalone machine learning does not consistently outperform statistical baselines. Published cases also associate forecasting with capacity planning and warehouse staffing without proving downstream cost reductions. A pooled meta-analytic effect could not be estimated from the available comparable evidence. The findings underscore transparent error metrics, matched horizons, reliable data integration, and decision-oriented model selection for future U.S. 3PL implementation and evaluation.