Skip to content
All library documents

VaR Alternatives, Tail Risk, and Better Backtesting

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

Summary

The document weighs alternatives to Value at Risk (VaR), especially Expected Shortfall and TailVaR, while discussing why VaR can give an incomplete picture of losses beyond a chosen threshold. One response notes growing attention to tail-sensitive measures and suggests heavier-tailed distributions such as Student-t. Another argues that poor VaR results can arise from simplistic implementation, and describes conditional volatility models such as GARCH paired with distributions that accommodate skewness and fat tails.

The discussion recommends evaluating risk models through backtests, including tests of exceedance frequency and clustering. It states that VaR is not subadditive, whereas Expected Shortfall is coherent, but claims VaR remains widely preferred in practice partly because of backtesting properties. These are practitioner opinions rather than a consensus survey or comparative study; the document does not settle which measure should be used for every reporting context. Model quality depends on implementation and the assumptions used for returns.

Key ideas

  • Expected Shortfall and TailVaR are discussed as alternatives that focus more directly on tail losses.
  • A heavier-tailed return distribution can represent extreme outcomes better than a normal distribution.
  • GARCH-style conditional volatility modeling can account for volatility clustering.
  • Backtests can assess whether VaR exceedances occur at the expected rate and whether they cluster.
  • VaR lacks subadditivity, while Expected Shortfall is described as a coherent risk measure.

Tags

Risk managementStatisticsBacktesting
Full text
# Is there a recognized alternative to Value at Risk for risk reporting?


# Is there a recognized alternative to Value at Risk for risk reporting?












In my research, VaR in risk reporting has regularly come under criticism for not capturing tail-risk adequately and for creating a false sense of confidence when taken literally as "the maximum value that can be erased at a 1% event".

Since I have been asked to provide that metric now, I was wondering: Has the finance community already converged on a best-practice alternative?

What are measurements that I could offer as a more sensible alternative?

EDIT:

I know about Expected Shortfall (https://en.wikipedia.org/wiki/Expected_shortfall) also called CVaR, so I would be grateful to hear about how widespread its usage is. I haven't looked at EVaR yet, but I'm generally looking for any opinions and a debate about what measure has emerged to be considered robust and where the intuitive interpretation corresponds to it's mathematical properties.

## Answer by arodrisa (score 1)

https://quant.stackexchange.com/a/32312

In the banking sector, people is getting more and more concern about this topic. We are listening a lot about Espected Shortfall and TailVaR.

But there are also methods to take more into account the tails. For example, you can change your distribution to a T-Student.

## Answer by PhD In Procrastination (score 1)

https://quant.stackexchange.com/a/32314

Regarding your "false sense of confidence" statement: most of it has nothing to do with VaR as a framework itself, it's about how you implement it. Naive HS-VaR fails you? Sure thing, but who's to blame in the situation except yourself?

Most of the pratitioners use VaR in its simplest form, which yields predictably poor results. For the sake of illustration let's consider one of the many ways of improvement: GARCH specification for conditional volatility paired with a distribution that allows for higher order (3rd and 4th) conditional moment dynamics (e.g. Normal Inverse Gaussian, Skewed Generalized t etc). This solves two problems: accounts both for heterosedastic structure of our series (volatility clustering) and fat tails. How great is the improvement? Time for backtesting.

VaR is flawed in a sense that it is not a coherent risk measure (it is not subadditive), but it is still massively preferred in practice over Expected Shortfall, which is a coherent measure, due to great backtesting properties, and that already tells a lot. I will not cover the tests themselves, you can easily look them up (unconditional coverage test, conditional coverage test, Berkowitz tail test etc). What you will notice (here, for example) when backtesting models based on normal distribution vs NIG, SGT etc, is that the former barely pass any tests while the latter perform fairly well (i.e. loss exceedances really occur at the specified confidence level, they do not cluster and fat tails are accounted for properly).

tl;dr: VaR is still one of the best practices out there when implemented properly.

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.