A0634
Title: Forecast evaluation of financial tail risk: Conditional MCS
Authors: Ekaterina Kazak - University of Birmingham (United Kingdom) [presenting]
Lukas Bauer - University of Freiburg, Statistics and Econometrics (Germany)
Abstract: The purpose is to address the evaluation of point forecasts for financial tail risk through the rationale for conditional model confidence set (CMCS). Financial regulations oblige financial institutions to perform stress testing, which involves evaluating risk forecasts conditional on a range of economic indicators, such as currency or monetary risk. Thus, evaluating global out-of-sample performance may obscure conditional differences that emerge under specific states, such as business cycles or volatility regimes. A prior study proposed testing for equal conditional predictive ability (ECPA) using instrumented moment conditions, while another study advanced this by approximating the conditional expected loss differential for uniform inference. These approaches, however, assume continuous loss functions over compact subsets, which may not be practical for discrete economic states or irregular loss differentials typical in tail risk evaluation. The CMCS concept diverges from traditional approaches by allowing for model selection conditional on specific states. It aims to identify sets of models for which ECPA cannot be rejected, providing a practical tool for decision-makers. Simulations and empirical applications indicate that CMCS effectively distinguishes between models with state-dependent performance, enhancing forecast reliability in stress scenarios.
