Federated Learning for Sustainable and Environmental, Social, and Governance-Aligned Supply Chain Forecasting in the Digital Economy: Prescriptive and Autonomous Analytics in Emerging Markets

Small and Medium-sized Enterprise (SME)-dominated emerging markets face structural barriers to AI-enabled supply chain analytics, including data fragmentation, privacy constraints, and limited governance infrastructure. This study develops and empirically evaluates a federated learning (FL)-enabled framework for Environmental, Social, and Governance (ESG)-aligned supply chain forecasting using multi-source observational datasets. The framework integrates three layers: (i) a predictive layer implementing Federated Averaging (FedAvg) across Long Short-Term Memory (LSTM), Extreme Gradient Boosting (XGBoost), and Transformer models; (ii) a prescriptive layer translating forecasts into carbon- and social-integrated Economic Order Quantity (EOQ) decisions; and (iii) an autonomous layer employing a Deep Q-Network trained via offline reinforcement learning on historical state–action data. The empirical analysis uses a mixed-source dataset of 61,470 observations combining directly observed public datasets (M5 Forecasting Competition, UCI Online Retail II, Food and Agriculture Organization Corporate Statistical Database FAOSTAT) with proxy-calibrated sector demand indicators for data-scarce industries (pharmaceuticals and consumer electronics). A retrospective backtesting framework evaluates counterfactual inventory policies derived from the ESG-integrated EOQ model against observed heuristic replenishment behavior. Results indicate that FL-based models achieve out-of-sample forecasting performance within 6–10% of centralized benchmarks under heterogeneous (Non-IID) conditions. Backtested policy evaluation shows reductions in inventory waste (− 413 units/month), stockout frequency (− 8.7 events/month), and carbon intensity (− 1.66 kg CO₂/order), alongside improvements in fill rate (+ 13.3% points). Robustness analysis using Federated Proximal Optimization (FedProx) and Stochastic Controlled Averaging for Federated Learning (SCAFFOLD) reduces the Non-IID performance differential to approximately 10.8–12.6%. These findings represent empirical performance differentials under historical conditions rather than causal estimates of real-world deployment effects. The study contributes an empirically grounded framework linking federated learning, ESG-aligned optimization, and privacy-preserving analytics in emerging market supply chains.