Large Scale Fitting of ARIMA Models and Stylised Facts of Economic Time Series - Produits Statistical working papers

The Statistical Office of the European Communities (Eurostat) publishes information on the economies of the Member States using, for some units, some model-based procedures to treat several features of economic time series. The quality of the information published is thus related to the capacity of these models, namely univariate Arima models with exogenous regressors, to adequately describe a vast majority of economic time series. We evaluate that capacity on a set of 13238 monthly series. The results of our experiment give several messages: 1) the sensitivity of different economic indicators to calendar events can be quantifed; 2) the occurences and the typology of outliers found in practice are detailed; 3) information is obtained about the stationary behavior of the series; 4) the practical relevance of several model specifcations can be evaluated; 5) the type of the mis-specifications found is detailed, yielding for example an indication on nonlinear patterns actually encountered in monthly series.