Potential Indicators Based Neural Networks for Cash Forecasting of an ATM

Abstract

Cash forecasting is one of the important tasks in the domain of computational finance. A number of tools have been developed by various groups of researchers and are being used by banks or corporate to identify future cash needs. However, due to the high degree of non-linearity of the problem and surrounded by many local optimal solutions, this paper propose a multi-layer locally tuned perceptron (MLTP) to forecast the future needs and at the same time reduce the users frustration. It uses a fine tuned MLTP to forecast a daily cash demand of an automated teller machine (ATM). Further, potential indicators are used to making the model robust in terms of its efficiency and accuracy. The accuracy is compared against a traditional time series method. Furthermore, it is validated using the past data collected from the SBI ATM of Bhadrak district of Odisha, India. The performance of the method is encouraging. This system can be scaled for all branches of a bank in an area by incorporating historical data from these branches.