Forecasting vault cash with an extreme value long short-term memory network

Abstract

Effective cash management is key in banking operations and has implications for cost control, customer service, and risk management. As transactions become more diverse, manual forecasting methods have become inadequate for accurate vault cash forecasting, which involves extensive data analysis. To address this challenge, the banking industry has adopted FinTech tools based on big data and deep learning for various client services. These methods are generally accurate but perform poorly in cases with extreme events, for which data are scarce. In this study, we propose a time series prediction model with long short-term memory and an attention mechanism that effectively predicts the presence of extreme values. We applied extreme value theory to define the extreme value loss for extreme situations and use a sliding window to process time series data. The enhanced extreme value loss function in our model yields improved prediction accuracy for time series data.

We evaluated the proposed model against previous methods in evaluation experiments on data from three branches of a commercial bank in Taiwan, where the vault cash data of each exhibited extreme values. The proposed model was highly accurate: it had a lower mean absolute percentage error and higher trend accuracy than competing methods on a majority of time series, and it was also more accurate in predicting extreme values in time series data.