Abstract
ATM networks remain essential cash-distribution infrastructure in cash-intensive economies such as Vietnam, where replenishment decisions must balance idle capital against stockout risk. We evaluate a forecast-then-optimize framework that combines probabilistic forecasting models with a periodic-review base-stock policy using daily withdrawal data from 84 ATMs over 3.7 years. Across 30 forecasting configurations, global neural quantile models substantially outperform classical baselines in downstream inventory performance. The results show that forecast accuracy alone is not a reliable proxy for business impact: models with similar prediction error can yield markedly different replenishment costs. The study also highlights ATM-level heterogeneity and supports a monitoring-based deployment strategy that adapts model assignments according to realized cost while maintaining high service levels.