A1665
Title: Bayesian evaluation of recursive multi-step-ahead path forecasts
Authors: Justyna Wroblewska - Krakow University of Economics (Poland)
Lukasz Kwiatkowski - Krakow University of Economics (Poland)
Anna Pajor - Krakow University of Economics (Poland) [presenting]
Abstract: The issue of ex-post evaluation of recursive multi-step-ahead path forecasts is analyzed. The approach adheres to the classical Bayesian paradigm hinged on the Bayes factors, which are here decomposed into the product of partial Bayes factors: the one for the entire k-step-ahead path, while the second reflecting the effect of updating the posterior odds ratio based on recursively updated data sets. The former factor reflects the relative k-step-ahead forecasting ability of models (with the whole k-period path being of interest, rather than only the final k-th observation), while the latter measures the updating effect as of a given time T to the beginning of the forecasting period. Next, a weighted approach is proposed, which amounts to using a weighted average of the logarithms of the Bayes factors, to hinge the recursive forecast assessment on the latest available data points, thus limiting the effect of a multiple incorporation of overlapping observations in the evaluation. The methodology is illustrated both with simulated as well as real-world data sets. In the latter, the predictive ability of vector error correction models featuring a variety of conditional heteroskedasticity specifications is investigated, for data sets representing the US and Polish economies. The results show that the model's forecasting performance depends on the weights, and forecast horizon as well as on taking into account the updating effect.
