Predicting ATM cash demand using machine learning

Efficient cash management in automated teller machine (ATM) networks requires accurate demand forecasting to minimize operational costs while preventing stockouts that reduce service quality. This paper presents a hybrid approach combining rigorous statistical analysis with deep learning for ATM cash demand prediction, analyzing 727,433 transactions from 10 ATMs over 24 months. We performed statistical preprocessing, including stationarity testing, Hurst exponent analysis, cross-correlation examination, and variance analysis across temporal scales. Partial autocorrelation analysis revealed multi-scale temporal dependencies from immediate to weekly patterns. Ablation studies demonstrated that calendar features provided the strongest improvements (19% error reduction), while lagged variables and aggregated statistics provided complementary gains. Combined features reduced error by 19.7% for LSTM and 16.6% for GRU versus the same architectures using only raw transaction amounts. Our optimized LSTM (1 layer, 96 units) achieved the lowest MAE, outperforming GRU by 1% and classical ML methods (e.g., Random Forest by 9%), while substantially exceeding naive baselines (48% improvement), validating the superiority of recurrent architectures for temporal dependency modeling in financial time series.