Description

This study presents a system for forecasting cash demand in Chile, both for total currency in circulation and by denomination. The framework considers three families of models: time-series specifications; cointegrated VAR models with macroeconomic fundamentals grounded in a money demand framework (activity, inflation, and interest rates); and stationary conjunctural models based on expectations of those fundamentals. For each family, the mean, the median, and the best-performing model are evaluated, allowing for a comparison of predictive performance and the identification of more accurate methodologies, while controlling for parametric and functional uncertainty through forecast combinations. Forecasting results at one-, two-, and three-year horizons show that predictive accuracy differs across model families and denominations, and that forecast combinations outperform individual models in several cases. In particular, time-series models exhibit a robust and relatively superior performance, especially in the case of coins, while cointegrated VAR models perform well for several banknote denominations, albeit with a loss of accuracy during periods of higher volatility. Conjunctural models, while incorporating additional information, do not systematically outperform time-series specifications and display, on average, a similar level of performance. The system constitutes an operational input for the Central Bank of Chile and is potentially applicable to other central banks, as it supports treasury management and the planning of printing, minting, and logistics.