Multiple arima subsequences aggregate time series Model to forecast cash in ATM

Abstract

Presently, ATM is used to conveniently serve customers in financial banking for multiple services such as transfer, deposit, and withdrawal. The main concern of ATM is the lack of banknotes for customer withdrawing. Hence, one of the interesting ATM issues is to forecast cash demand on the individual ATM as a time series data. Typically, the customer cash usage behavior of ATM shows a sway and it often contains a hidden pattern due to multiple customer behaviors. Consequently, this research proposes a new forecasting model for predicting cash amount in ATM, called the multiple ARIMA subsequences aggregate time series model or the MASA model. In the algorithm process, the ATM data is split into the in-sample data for building the MASA model and the out-sample data for evaluating it. To build the MASA model, the aggregate technique is applied using the period of the in-sample data in order to reduce a sway of the time series which is subdivided into the training and the validating data. Then, subsequence patterns are generated and fitted using auto.arima function by the library forecast of the R programming language. They are evaluated using the symmetric mean absolute percentage error (SMAPE) on the validating data. In order to forecast a value of the next time step, the algorithm identifies the time step aggregate group and the best fitted ARIMA model from a subsequence containing that group. The algorithm then disaggregates the forecasted aggregate group value into individual future values using the ratio from the in-sample data. The MASA model is compared using SMAPE with the SARIMA and exponential smoothing model. The comparative analysis shows that the SMAPE of the MASA model is better than both models.