ARTIFICIAL INTELLIGENCE-ENHANCED PREDICTIVE ANALYTICS FOR DEMAND FORECASTING IN U.S. RETAIL SUPPLY CHAINS

This systematic review synthesizes evidence on artificial intelligence enhanced predictive analytics for demand forecasting in U.S. retail supply chains, with a focus on decision relevance and deployment realism. Guided by PRISMA, we searched major multidisciplinary databases for 2015 to 2025, screened records in two stages, assessed leakage risk and baseline adequacy, and extracted harmonized metrics for point accuracy, probabilistic calibration, and inventory outcomes. The final analytic corpus comprises 95 peer-reviewed studies. Across comparable evaluations, AI models consistently outperformed strong statistical baselines, yielding median WAPE reductions of roughly 7 to 9 percent, with larger gains under cross-series training and promotion rich contexts. Feature discipline mattered: encoding price and promotion depth, holiday proximity, and identifier representations delivered an additional 3 to 6 percent improvement. Structure added value: hierarchical and cross-temporal reconciliation contributed about 4 percent error reduction and improved quantile coverage, while spatiotemporal learners reduced store-day errors by about 6 percent in geographically correlated demand. Probabilistic outputs translated into operations, enabling about 12 percent safety stock reduction at fixed service or roughly 3.5 percentage point fill rate gains at fixed inventory. Deployment practices shaped realized value: drift monitors, champion challenger governance, and accountable human overrides shortened post-shock recovery, cut stockouts by about 11 percent, and modestly increased inventory turns. We integrate these findings into a practical selection framework that aligns data realism, global modeling, calibrated quantiles, structural reconciliation, and MLOps guardrails to deliver coherent forecasts that are auditable and economically meaningful for U.S. retail planning. Implications for researchers and practitioners are discussed.