A0821
Title: Time-series evidence on the influence of the choice of seasonal adjustment method on forecasting accuracy
Authors: Robert Kunst - Institute for Advanced Studies (Austria) [presenting]
Martin Ertl - Institute for Advanced Studies (Austria)
Adrian Wende - Institute for Advanced Studies (Austria)
Abstract: Seasonally adjusted data are routinely used in applied research, particularly in empirical economics. Mainly, two methods of seasonal adjustment are used: moving-average X-11 and the SEATS method, which is based on tentatively fitted ARIMA models. The aim is to study which of the two methods yields more accurate forecasts of annual targets and when it is better not to adjust seasonally. These issues are investigated empirically and with Monte Carlo simulations. For the simulations, data-driven time-series models are considered, both univariate and multivariate generating processes. In assessing the benefits of seasonal adjustment procedures, the basic challenge is that the true seasonally adjusted variable remains unknown. Whereas the literature uses criteria such as robustness to new information at the end of the sample and plausibility, this approach involves some arbitrariness. The comparison is subjected across methods to a quantitative criterion, and the accuracy of the final forecast is chosen for the annual variable that is non-seasonal and observed. For empirical applications, the focus is on quarterly EU and UK national accounts variables. Preliminary results show no large differences in performance between the two major adjustment approaches. However, the reaction to outliers, as they occurred during the Covid pandemic, can be a challenge. In simulations, this type of robustness is studied via heavy-tailed generating distributions and time-series outlier models.