Liquidity risk is one of the key risks faced by banks in their daily operations. Following the recent financial crisis, more stringent measures have been put in place to ensure that banks adequately cater for sufficient liquidity and stable funding. In liquidity planning a difficult component is the modelling of indeterminate maturity products, which from a liabilities point of view includes transactional and savings accounts (demand deposit accounts). Banks utilise the balances in these products to also partly supply the funds necessary for loans and other forms of credit, which generate most of their profits. The purpose of this study was to find a way to accurately forecast the daily bank balance of a demand deposit account portfolio across the period of a year. This would help the banks to more efficiently handle liquidity planning and also generate more profit by utilising their funds more effectively. In accomplishing this the study also presented the hypothesis that using a hybrid model which combined segmentation with a popular forecasting method such as autoregressive integrated moving average (ARIMA) models would do better than a single
time series forecasting model. The purposes of the segmentation was to identify customers with similar current account dynamics e.g. salaried individuals in comparison to a small business owner.
Segmentation was facilitated by extracting features from the time series that identified patterns of salaried individuals in comparison to other account holders. These features were used by the k-means algorithm to form the segments. ARIMA models were then implemented for each of the segments and forecasts obtained per segment. These segment level forecasts were then aggregated to obtain the portfolio level forecasts. The results were then compared to building a single model to forecast the portfolio daily balance. Results from the study suggest that the hybrid model statistically performs significantly better than the single model over shorter forecast horizons. This study also attempted to find a way to score customers into one of the identified segments using information available on enrolment. However, results suggested that there is not enough discriminative power available in the information collected at enrolment but rather it is better to include information regarding a customer’s first month’s bank balance which significantly improved the classification accuracy.
Dissertation (MEng)--University of Pretoria, 2018.