First results series or last available series: which series to use? A real-time illustration for the forecasting of French quarterly GDP growth

Short-term analysts use many tools to forecast economic activity. Among these tools, estimating and, then, simulating univariate models is very common. Most of the time the series used for the variable of interest as well as for the regressors are last available releases. Except for the latter observation, data have thus usually been revised many times since they were first published. Using data as close as possible to the reality seems quite natural. From this point of view, last published data are the best candidates. However this study proves that models based on these series are not the most accurate when one aims at predicting the first published data (and not the definitive one published years after). It can be the case when the predictive accuracy is evaluated considering the forecast errors calculated in comparison to the first released values. In such a case, it is more efficient, under certain assumptions, to perform models based exclusively on the series of first published data (first results series also called "real time" data) instead of last available series. The interest of this approach is firstly theoretically assessed considering the framework of the French quarterly accounts and secondly empirically validated by comparing the results of both methods (series of last available results compared to series of first results) to forecast GDP growth in real time. It turns out that to forecast the first published value of GDP quarterly growth, using the series of the first results as the variable of interest improves significantly the quality of forecasts.