Interpreting VaR Magnitudes Through Model Backtesting

Summary

The document addresses confusion about the sign and absolute value of Value at Risk outputs, using two negative figures as an example. The response redirects the question from which output is inherently better to whether the underlying risk model describes actual losses appropriately. A VaR figure is a model estimate at a chosen confidence level, so its magnitude alone does not show that a model is good or bad without context about how it is defined and evaluated.

The response describes two failure patterns: a model that overstates risk so much that realized losses do not exceed its VaR threshold, and one that understates risk so that exceedances occur too often. It says a reasonable model should produce an exceedance frequency consistent with its confidence level, and identifies backtesting as the tool used to assess this. The passage provides a conceptual explanation, not a detailed backtesting procedure or a resolution of the example’s sign convention. Interpretation still depends on defining the VaR output consistently and comparing it with realized outcomes.

Key ideas

• A VaR output cannot be judged as better or worse from its magnitude alone.
• Model quality depends on how well estimated risk aligns with observed losses.
• Too few or too many VaR exceedances can indicate poor calibration.
• Backtesting compares VaR estimates with realized losses to assess model performance.
• The discussion does not specify a detailed backtesting method or resolve every sign convention.

Tags

# Value at Risk Theory


# Value at Risk Theory












I am having a bit of trouble disseminating the true meaning behind VaR.

Say you have two V Values, prior to taking ABS value. Both values are negative, the first value being -10 and the second value being -20. After taking the ABS Value, the Var answer becomes 10 and 20, respectively. Both values have an 99% Confidence interval.

What value is better? Convention would say, prior to taking ABS, -20 is better, since its further from zero, hence less loss. But after you take the ABS Value, you are now have a higher degree of loss?

What am I missing?

## Answer by Nicholas (score 2)

https://quant.stackexchange.com/a/22287

I won't base my answer on your example as i couldn't understand what you mean.

Firstly, when you ask a question "what is better?" you should address this question to the model and not the output values.

Secondly, model is good only when it as accurately as possible explains the reality (with a degree of confidence). The model is useless if it overestimates the risk i.e. there are no overshoots - actual losses never exceed VaR levels. The model is still useless if it underestimates the risk i.e. too many overshoots - actually losses frequently exceed VaR levels. Reasonable model is expected to produce a limited number of overshoots (consistent with a degree of confidence).

Tool used to analyse the performance of VaR model is called backtesting.

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.