Daily ATM Cash Demand Forecasting with Deep Learning-Based Hybrid CNN-LSTM Architectures

Abstract

The predictability of daily cash demand at Automated Teller Machines (ATM) is a sensitive operational objective of financial institutions, which directly affects the efficiency of cash logistics, the availability of services, and the economy of costs. The resulting errors in forecasting can lead to the recurrence of cash shortages, therefore, customer dissatisfaction and reputational losses, or surplus idle cash, to the higher security and opportunity costs. The competition held in the NN5 forecasting predicted that there was an inherent complexity in predicting ATM cash withdrawals due to the strong seasonality in weekly, calendar effects, missing data, and the temporal heterogeneity in demand patterns across ATM locations. The paper provides a deep learning-based forecasting system of NN5 daily ATM cash demand prediction by incrementally scaling a stock market price prediction system first created to predict a financial time-series. The proposed framework introduces domain-specific enhancements, such as calendar-based feature engineering and hybrid convolutional-recurrent architectures, while retaining the fundamental ideas of sliding-window supervision, normalization, and recurrent neural modeling. Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU) as well as a hybrid model, Convolutional Neuron Network-LSTM (CNN-LSTM) model are deployed and tested within a single experimental framework. Decades of experiments on the NN5 data depict that the proposed CNN-LSTM setup is much better than the classical statistical models, such as ARIMA and exponential smoothing, and even standalone recurrent neural networks, and results in up to a 19 percent decrease in Root Mean Square Error (RMSE). The findings validate that the combination of the cross-domain architectural transfer, hybrid models of deep learning, and calendar-sensitive feature design offers a stable and scalable solution to the real-world prediction of the ATM cash demand.