A1758
Title: Outlier-sensitive forecast evaluation of value-at-risk and expected shortfall
Authors: Takaaki Koike - Hitotsubashi University (Japan) [presenting]
Cathy W-S Chen - Feng Chia University (Taiwan)
Abstract: Backtesting via calibration tests is standard for financial risk measures such as Value-at-Risk and Expected Shortfall. However, for highly volatile assets, their risk forecasts can themselves become extremely large in absolute value, and existing calibration tests may not adequately penalize such outlier forecasts. An outlier-sensitive extension of VaRES calibration tests is proposed that reweights the identification functions so that extreme forecasts receive stronger penalties. This improves the ability of the tests to detect models that generate overly volatile or excessively conservative tail risk forecasts. At the same time, stronger penalization creates a trade-off since the validity of asymptotic inference becomes more demanding in terms of moment conditions. To address this issue, an admissible region is estimated in which these conditions are likely to remain satisfied and test statistics are developed that aggregate calibration evidence over that region. The proposed framework provides a practical diagnostic for assessing robustness and calibration under tail stress. An empirical application to cryptocurrency return forecasts shows that the method can distinguish models with stable tail risk forecasts from those whose apparent performance deteriorates once extreme forecasts are appropriately penalized.
