Gauging the Effect of Influential Observations on Measures of Relative Forecast Accuracy in a Post-COVID-19 Era: Application to Nowcasting Euro Area GDP Growth

The previous research already emphasised the importance of investigating the predictive ability of econometric models separately during expansions and recessions (Chauvet and Potter (2013), Siliverstovs (2020), Siliverstovs and Wochner (2020)). Using the data for the pre-COVID period, it has been shown that ignoring asymmetries in a model's forecasting accuracy across the business cycle phases typically leads to a biased judgement of the model's predictive ability in each phase. In this study, we discuss the implications of data challenges posed by the COVID-19 pandemic on econometric model estimates and forecasts. Given the dramatic swings in GDP growth rates across a wide range of countries during the coronavirus pandemic, one can expect that the asymmetries in the models' predictive ability observed during the pre-COVID period will be further exacerbated in the post-COVID era. In such situations, recursive measures that dissect the models' forecasting ability observation by observation allow to gain detailed insights into the underlying causes of one model's domination over the others. In this paper, we suggest a novel metric referred to as the recursive relative mean squared forecast error (based on rearranged observations) or R2MSFE(+R). We show how this new metric paired with the cumulated sum of squared forecast error difference (CSSFED) of Welch and Goyal (2008) highlights significant differences in the relative forecasting ability of the dynamic factor model and naive univariate benchmark models in expansions and recessions that are typically concealed when only point estimates of relative forecast accuracy are reported.