The article analyzes a current problem of maintaining a service of automated deposit machines and ATMs of credit institutions. Solution of the forecasting problem will help in automating the process of cash collection management and organizing an efficient planned load of a self-service machine in order to minimize time when the machine is unavailable for service. To solve the forecasting problem, the comparative analysis of existing mathematical methods of time series forecasting (such as exponential smoothing models, seasonal autoregressive integrated models (moving average) and models based on the neural networks (single-layer neural network and neural-fuzzynetwork)) is obtained. The analysis is carried out based on the data of the amount of cash withdrawn from the ATM in a closed area of the business center. As a result, the most efficient forecasting method (neural-fuzzy network) is determined and the forecast of cash withdrawal for the selfservice machine is obtained. Management of a planned load of ATMs based on the obtained forecast minimizes time of technical delay of the machine in case of cash assets excess or shortage as well as allows optimizing the collector’s work.