The big problem of forecasting small change1

Abstract

The United States Mint recently reviewed approaches to forecasting the demand for new coin. This paper reports on methods used to determine fundamental attributes of the data, and uses these to help better determine appropriate model specification in order to better plan coin production. In particular, the debate regarding trend versus difference stationarity in macroeconomic trending data is considered. The interest in the present paper is limited to applying a well known unit root test procedure to an untested macrodata set – changes in US Coin demand – to see whether the test is useful in guiding the specification to improved forecast performance. It is found that the forecast results are somewhat sensitive to the way in which the data are seasonally adjusted, and lessons learned from this ‘case study’ indicate that unit root tests are useful in guiding model specification.