Uncertainty-Aware Hybrid AI Forecasting for Multi-Location Inventory Redistribution: A Forecast-then-Optimize Framework

Multi-location inventory optimization increasingly depends on machine-generated demand and anomaly signals, yet a persistent gap remains between these signals and the operator-ready redistribution plans that supply chain teams need to act on. This work presents a forecast-then-optimize (F&O) framework designed to close that gap.

The forecast stage combines ARIMA, XGBoost, and Long Short-Term Memory (LSTM) models through an adaptive weighting mechanism that adjusts model contributions based on observed performance, together with a conformal prediction layer that produces calibrated demand and anomaly signals. The optimize stage is a five-step LLM agent pipeline that converts these signals into concrete, operator-ready redistribution plans across locations. This two-stage architecture is grounded in the established predict-then-optimize literature, following Elmachtoub and Grigas (2022) and Bertsimas and Kallus (2020), and is explicitly positioned as an instance of that paradigm applied to last-mile inventory redistribution.

The framework has not yet been validated in a live deployment. To assess its potential, we conducted a synthetic evaluation across five network scenarios calibrated against published retail and logistics benchmarks. These synthetic estimates indicate a mean plan correctness rate of 92.4%, a mean MAPE of 8.3% under stable demand (rising to 14.7% under peak-season variability, consistent with published ML ensemble benchmarks), a simulated service-level improvement of 11.4 percentage points over a no-action baseline, and a mean decision-latency reduction of 132 minutes per incident relative to a manual workflow. These figures are indicative estimates from simulation, not empirically measured outcomes, and are presented as such.

This session will discuss the framework's design rationale, its grounding in the predict-then-optimize literature, and the simulation methodology used to generate these indicative results. Empirical validation through a live industry deployment is identified as the primary direction for future work, and the author is actively seeking industry partners for this next stage.