Journal of Risk Model Validation

Risk.net

The validation of filtered historical value-at-risk models

Pedro Gurrola-Perez

  • While backtesting can provide evidence of past coverage, it fails as a tool to assess the adequacy of FHS models.
  • The usual conditional coverage hypothesis tests used in backtesting tend to produce results which are biased in favour of overreacting models.
  • Incorporating information about the duration or the magnitude of the breaches does not significantly improve the power of backtesting for the calibration of FHS models.
  • The APL score test outperforms all the tests considered. This test uses information from the whole empirical distribution and not only from the set of breaches.

Recent value-at-risk (VaR) models based on historical simulation often incorporate approaches where the volatility of the historical sample is rescaled or filtered to better reflect current market conditions. These filtered historical simulation (FHS) VaR models are now widely used in the industry and, as is usually the case with VaR models, they are validated through backtesting. However, while backtesting is a natural way of testing a percentile forecast, it is not specifically designed to capture other features of the model, such as its efficiency in adapting to new volatility conditions. In this paper, we discuss the limitations of backtesting as a tool to assess the performance of FHS models and, using a Monte Carlo simulation framework, we examine whether incorporating information about the size of the breaches (through the use of score functions, for example) can improve the efficiency of these tests. The results show that, even when incorporating the size of the VaR violations, tests based solely on the breaches generally fail as a tool to discriminate between different calibrations of the decay factor; they also tend to be biased. Among the alternative tests considered, the asymmetric piecewise linear score performs best overall, followed by the dynamic quantile test. We conclude by considering some empirical examples.

To continue reading...

You need to sign in to use this feature. If you don’t have a Risk.net account, please register for a trial.

Sign in
You are currently on corporate access.

To use this feature you will need an individual account. If you have one already please sign in.

Sign in.

Alternatively you can request an indvidual account here: