Automated Teller Machine Cash Demand Prediction

Abstract

Physical cash is the most liquid asset and holds high usage rates despite rapid technological advancements in the digital cash industry, especially in the context of a developing country like India. As such, Automated Teller Machines (ATMs) form an integral component of any individual's cash needs and managing the amount of cash in any given ATM is of paramount importance for any bank. ATMs are periodically filled with cash by banks, predominantly with the help of cash management agencies. Through our research we want to determine the feasibility of adding mathematical support to the current system by using Data Analysis and Machine Learning (ML). The findings of this research and modelling approaches were applied to a procured dataset which consists of data of 5 ATMs of the same bank from 2011–2017. The year, month, day, festival religion, day of the week, holiday status of yesterday, today and tomorrow were the factors that were initially analyzed and later on, average withdrawal for a given window of time in the past, the weather status and the type of holiday were the other features engineered which proved to be useful. The algorithms being tested include Lasso, Ridge, Elasticnet, Random Forest & CatBoost which is a deep learning framework. Two approaches for testing were determined, one approach was to create one model for all years in the data and another approach was to develop a model per year. In both approaches, it was found that Lasso regression outperformed all the other algorithms in contention.