Skip to content
All library documents

Why VaR Backtests Test Independence of Exceedances

Article Quant Q&A · Author: Oscar
Read the original

Summary

The document discusses the rationale for testing whether Value at Risk exceedances are independent over time, as in Christoffersen-style backtesting. The question challenges whether this test is meaningful when a risk model might not adjust quickly enough after a large loss to prevent another exceedance. It also distinguishes dependence in return magnitudes, which may arise from volatility clustering, from dependence in the sequence of VaR breaches.

The response argues that volatility forecasts can react to recent large returns or realized volatility with a one-period lag. It names GARCH and HAR models as examples: an elevated volatility estimate can raise the next period’s VaR and reduce the chance of another breach for a given return. This challenges the premise that consecutive exceedances are inherently unavoidable. The passage is brief and does not derive the independence test, compare forecasting models, or provide empirical results; it explains a rationale rather than establishing that any specific VaR model will pass the test.

Key ideas

  • VaR backtesting may examine whether breach indicators are independent across time.
  • Return magnitudes can exhibit volatility clustering, but that does not alone determine whether VaR exceedances are independent.
  • GARCH and HAR volatility forecasts can incorporate recent large returns or realized volatility with a period lag.
  • A higher forecast volatility can increase the next period’s VaR and reduce breach likelihood.
  • The passage offers a conceptual argument, not empirical proof that a model will pass independence testing.

Tags

BacktestingRisk managementVolatilityStatistics
Full text
# Rationale behind independence testing (e.g. Christoffersen’s test) for Value at Risk backtesting


# Rationale behind independence testing (e.g. Christoffersen’s test) for Value at Risk backtesting












I'm reading up on backtesting methodologies and having trouble understanding the rationale for independence testing using, for example, the Christoffersen method. (Christoffersen, Peter F: Evaluating Interval Forecasts, International Economic Review, 1998, link).

The idea is to test whether the exceedances are independent and that it is not more likely for there to be an exceedance on a given day if the previous day had an exceedance.

Now, one desirable feature of a Value at Risk model is that it is responsive to regime changes, but it seems infeasible for a model to react fast enough that, given two consecutive big losses for the portfolio, the second one would not trigger an exceedance. For example, if VaR is 15k and we observe two days of -20k losses, having the VaR jump to >20k would make the model appear erratic and likely deemed unusable.

With that in mind, what we are really testing is whether the magnitude of returns themselves is independent: something we already know is not the case due to volatility clustering.

Is there any actual merit to independence testing that I'm missing (aside from pleasing regulators)?

## Answer by Richard Hardy (score 1)

https://quant.stackexchange.com/a/81975

> it seems infeasible for a model to react fast enough that, given two consecutive big losses for the portfolio, the second one would not trigger an exceedance.

Why would it seem infeasible? GARCH, HAR and other volatility models based on some sort of ARMA patterns in volatility react with a single period's lag (a day for daily data) to the most recent squared return or realized volatility. If there is a large return or large realized volatility on day $t-1$, the model will produce large volatility on day $t$, lowering the probability on day $t$ relative to day $t-1$ of VaR exceedance by any given return. Thus, I do not think the following applies:

> With that in mind, what we are really testing is whether the magnitude of returns themselves is independent: something we already know is not the case due to volatility clustering.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.