CFE 2019: Start Registration
View Submission - CFE
A0166
Title: Quantile forecast evaluation: An application to growth-at-risk Authors:  Valentina Corradi - University of Surrey (United Kingdom) [presenting]
Abstract: A forecast evaluation procedure is introduced for multiple, possibly misspecified quantile models. The models are evaluated in terms of their relative unconditional coverage, that is we rank models in terms of the distance between actual and nominal coverage. The key novelty of our approach is that we do so uniformly over a compact set of quantile ranks, rather than at a single, pre-specified quantile level. In a final step, we then construct a model confidence set that contains all models which are `equally good' over a weighted average of quantile ranks. As all model parameters are estimated using a recursive scheme, the contribution of recursive quantile estimation error has to be taken into account. Inference is based on block bootstrap p-values, and we establish a bootstrap Bahadur representation which is valid uniformly over quantile ranks and in a recursive setting. Finally, we apply our procedure to compare out-of-sample predictions for growth-at-risk of different quantile models and of professional forecasters.