Abstract
Accurate demand forecasting alone is insufficient for inventory optimization in multi-channel supply chains, where replenishment decisions depend on both expected demand and predictive uncertainty. This study presents an end-to-end decision framework that combines machine learning-based forecasting with uncertainty modeling and interpretable, cost-based inventory optimization. The framework uses multi-channel retail demand signals and exogenous features as inputs, generates probabilistic forecasts using advanced deep learning architectures, and translates forecast uncertainty into operational inventory parameters such as safety stock and reorder decisions. Using a statistically representative synthetic dataset calibrated to real retail operating structures, we evaluate Temporal Fusion Transformers (TFT), Long Short-Term Memory networks (LSTM), and spatiotemporal Graph Neural Networks (ST-GNN) against classical statistical baselines. We introduce a quantile-based uncertainty-to-inventory translation mechanism that converts predictive distributions into safety stock and cost inputs for a transparent inventory cost minimization model. Results show that improvements in probabilistic forecasting and uncertainty calibration translate into tangible operational gains, including 22-35% improvements in forecast accuracy and 14-21% reductions in total inventory cost relative to statistical baselines, while maintaining target service levels. The study provides a modular and scalable pathway for deploying uncertainty-aware inventory optimization across multi-channel retail environments. • Presents an end-to-end framework: multi-channel inputs → ML forecasting → uncertainty modelling → cost-based inventory optimization. • Evaluates TFT, LSTM, and ST-GNN as probabilistic forecasting models for multi-channel retail demand. • Introduces a quantile-based mechanism translating predictive uncertainty into safety stock and reorder parameters. • Demonstrates 22-35% forecasting accuracy gains and 14-21% inventory cost reduction versus statistical baselines. • Delivers a transparent, modular decision-support architecture for uncertainty-aware inventory planning.
Showing the abstract — retrieve the full paper via the Exa API.