Cash demand prediction
Customer Issue: A bank in the US is facing challenges with cash management in its ATMs, leading to instances where machines frequently run out of cash or accumulate too much. To address this problem, the bank needs a model that can predict daily cash demand for each ATM.
Datamotus Methodology: We conducted extensive data cleaning and exploratory data analysis (EDA) to identify the factors influencing cash demand for ATMs. During the EDA, we examined various elements, including the "weekday effect" and "holiday effect." We employed two different strategies for time series forecasting, utilizing multiple methods such as ARIMA, ARIMAX, time series regression, exponential smoothing, and LSTM (a deep learning approach) for both grouped and individual datasets. We compared the forecasting accuracy of all models and selected the best performing one. This search for the optimal forecasting model was carried out in R, while the most effective model was implemented in Python for production use.
Results: The bank was able to notably decrease the amount of excess cash in its ATMs, lower the number of cash shortages, and enhance its overall cash management procedures.