The importance of data revisions for statistical inference
The aim of the study is a quantitative analysis of revisions conducted by means of a new, real-time macroeconomic dataset for Poland, designed on the basis of the Statistical bulletin (Biuletyn statystyczny) published by Statistics Poland, covering the period from as early as 1995 until 2017. Polish data have positively verified a number of hypotheses concerning the impact of data revisions on the modelling process. Procedures assessing the properties of time series can yield widely discrepant results, depending on the extent to which the applied data have been revised. A comparison of the fitted ARIMA models for series of initial and final data demonstrates that the fitted models are similar for the majority of variables. In the cases where the form of the model is identical for both series, the coefficients retain their scale and sign. Most differences between coefficients result from a different structure of the fitted model, which causes differences in the autoregressive structure and can have a considerable impact on the ex ante inference. A prognostic experiment confirmed these observations. For a large number of variables, the total impact of revisions on the forecasting process exceeds 10%. Extreme cases, where the impact goes beyond 100%, or situations where data have a direct impact on the forecast sign, are also relatively frequent. Taking these results into account by forecasters could significantly improve the quality of their predictions. The forecast horizon has a minor impact on these conclusions. The article is a continuation of the author's work from 2017.