Forecasting French GDP with Dynamic Factor Models : a pseudo-real time experiment using Factor-augmented Error Correction Models

Dynamic Factor Models (DFMs) allow to take advantage of the information provided by a large dataset, which is summarized by a small set of unobservable latent variables, and they have proved to be very useful for short-term forecasting. Since most of their properties rely on the stationarity of the series, these models have been mainly used on data which have been di_erenciated to achieve stationarity. However estimation procedures for DFMs with I(1) common factors have been proposed by Bai (2004) and Bai and Ng(2004). Further, Banerjee and Marcellino (2008) and Banerjee, Marcellino and Masten (2014) have proposed to extend stationary Factor Augmented VAR models to the non-stationary case, and introduced Factor augmented Error Correction Models (FECM). We rely on this approach and conduct a pseudoreal time forecasting experiment, in which we compare short term forecasts of French GDP based on stationary and non-stationary DFMs. We mimic the timeliness of data, and use in the non-stationary framework the 2-step estimator proposed by Doz, Giannone and Reichlin(2011). In our study, forecasts based on stationary or non-stationary DFMs have a similar precision.