Zusammenfassung:
Among the most important tasks of central banks is to ensure the availability of cash to credit institutions and retailers. Forecasting the demand for cash on a granular level is crucial in the process to keep logistics costs low, while being resilient to demand or supply shocks. Whereas to date, cash forecasts with central banks mostly comprise structural models to define banknote production for the coming years, our contribution is to combine features of macro level forecasting with more granular and short term regional forecasts methods. We show in an inventory simulation, that elaborate forecasting methods on granular level can substantially improve inventory performance for this use-case. To guide the implementation of a forecasting process at the Bundesbank, we benchmark statistical and machine learning methods on demand and supply of cash, using anonymized data on transactions of six regional branches of Deutsche Bundesbank. We use a pseudo out of sample predictive performance framework to evaluate the accuracy of our forecasts and perform an inventory cost simulation. We find that (i) DeepAR outperforms the other benchmarks substantially on all data sets. (ii) ETS, ARIMA, and DeepAR clearly outperform the naive benchmark in terms of accuracy across all data sets, and inventory performance.