Rationality and Quality of Inflation Forecasts

In this paper individual forecasts from SPF-US and SPF-ECB were subject to rationality tests. Firstly, we classified all forecasters as “rational” or “non-rational” according to each type of the loss function: symmetric squared and linex for OLS estimation; symmetric squared, asymmetric squared, symmetric linear and asymmetric linear for GMM estimation. Then we calculated average forecasts (mean and median) for two subgroups and for all forecasters. After that three groups were pairwisely compared in terms of forecast’s accuracy using modified Diebold-Mariano statistics. Finally, we conducted rationality test for consensus forecast for both datasets.Our main findings could be summarized as follows:•Classification of the forecaster as “rational” or “non-rational” heavily relies on the assumption about loss function and employed estimation method. The share of rational forecasters fluctuates from 30% till 60% depending on the type of loss function and dataset. GMM is more prawn to classifying forecaster as “non-rational” and OLS as “rational. Moreover, these methods sometimes give opposite results for the same type of the loss function.•Rationality of the forecaster could not be seen as the criterion for forecast aggregation. For both datasets we did not find any evidence that rational forecasters make more accurate projections than non-rational ones or average forecast. That is why forecast combination based of the rationality concept does not seem to be an applicable task in spite of theoretical foundations.•Consensus forecasts could not be classified as “rational” for most of the cases. GMM estimation classified forecasts as “non-rational” or results were not stable with respect to forecast horizons for both datasets. OLS technique together with linex loss function classified SPF-US consensus forecast as rational. However, the use of OLS resulted in the rejection of rationality hypothesis for SPF-ECB consensus forecast for all types of loss functions. These results suggest that individual projections could not be represented by “aggregate forecaster” whose behavior could be described as rational according to some loss function.