A Long-Short-Term-Memory Based Model for Predicting ATM Replenishment Amount

Abstract

The advent of Automatic Teller Machines (ATMs) enable self-service, time-independent, easy to use, mechanism through which a financial institution supports large number of services to its users. Cash withdrawal from the ATM is still one of the major transactional loads for these networks. ATM cash replenishment is the process by which ATM machines are filled with the cash so that the users can withdraw it. The rapid adaptation and standardization of these network give rises to many challenging problems that requires intelligent management of these resources. ATM cash replenishment amount prediction is one such problem, predicting the right amount for everyday use such that the minimum amount of cash always available before the next replenishment. In this way there will be no customer dissatisfaction through empty ATM. The paper proposes a machine learning approach to ATM replenishment amount prediction by using a data driven approach for the estimation of right amount for each ATM or some group of ATMs. The data comprises of replenishment of 2241 ATMs for last 22 months from 6 different Banks of Pakistan. The Long Short-Term Memory (LSTM) based model produce Root Mean Squared Error (RMSE) of 132.53, which is quite encouraging for this problem.