Forecasting Cash Money Withdrawals Using Wavelet Analysis and Wavelet Neural Networks

The increasing demand for easily accessible cash drives banks to expand their Automatic Teller Machine networks. As the network increase it becomes more difficult to supervise it while the operating costs rise significantly. Cash demand needs to be forecasted accurately so that banks can avoid storing extra cash money and can profit by mobilizing the idle cash. This paper is motivated by the Neural Network Association and the NN5 competition. The objective of the paper is to describe a unique, non-supervising method for forecasting cash money withdrawals in different ATMs. More precisely, the data consists of 2 years of daily cash money demand at various ATMs at different randomly selected locations across England. The only available information is the total cash withdrawals in each ATM at the end of each day. Having limited domain knowledge and no information on the causal forces we use wavelet analysis to extract the dynamics of the underlying process of each ATM. Next wavelet neural networks were used in order to find the true generating process of each ATM and to forecast the cash money demand up to 56 day ahead. The performance of the proposed technique is evaluated using various error and fitting criteria.