Uncertainty-Aware Forecasting and Inventory Optimization for ATM Cash Management in Vietnam

Huu-Thanh Phan, Xuan-Bach Le, Tho Quan

Abstract: ATM networks remain essential cash-distribution infrastructure in cash-intensive economies such as Vietnam. Effective replenishment must balance idle-cash holding cost against stockout risk. We evaluate a forecast-then-optimize framework that integrates probabilistic forecasting models with a periodic-review base-stock policy. We benchmark 30 forecasting configurations, including statistical baselines, tree-based ensembles, and global neural architectures, on daily withdrawals from 84 ATMs over 3.7 years. Performance is assessed using both forecasting quality and downstream inventory outcomes under realistic operations. Global neural quantile models substantially outperform classical alternatives, reducing simulated total cost by roughly half while maintaining near-perfect calibration and 99% fill rates. We further show that forecast accuracy alone is an unreliable proxy for operational value, as similarly accurate models can produce very different costs. ATM-level heterogeneity indicates no single model dominates all locations, supporting a monitoring-based deployment strategy with adaptive model assignment by realized cost.