When Do Foundation Models Pay Off for Retail Demand Forecasting? A Systematic Review and Cross-Benchmark Meta-Analysis

We present a PRISMA-informed systematic review of forecasting methods for short-term retail demand, combined with original gap-filling experiments and cross-benchmark extraction from fev-bench and GIFT-Eval. We screen approximately 150 papers from 2020 to 2026 and extract quantitative results into more than 800 rows spanning nine retail datasets. A model-level meta-regression on 543 paired comparisons across foundation models and conventional baselines finds a robust benchmark-level foundation-model advantage, strongest on distributional metrics and smaller on point-forecast metrics. The report includes sensitivity checks, benchmark-level caveats, and a practitioner-oriented cost-accuracy Pareto analysis.