Why the Central Limit Theorem Does Not Justify Normal VaR

Summary

The document addresses whether a large dataset of dependent variables can be treated as multivariate normal for variance–covariance Value at Risk. It cautions that the central limit theorem does not establish normality of the raw observations simply because there are many data points. The theorem concerns averages of sufficiently well-behaved observations, and the response notes that dependence or serial correlation can invalidate assumptions behind a variance–covariance estimate and lead to understated risk.

It outlines several VaR approaches, including historical quantiles, variance–covariance, simulation, and copula methods. Bootstrapping is suggested as a way to estimate uncertainty in VaR itself, while empirical quantiles are offered as a direct alternative when the central limit theorem’s conditions do not apply. The discussion is conceptual and does not specify a confidence level, sampling design, or estimator. It also flags that VaR has limitations as a risk measure, so method choice and dependence assumptions matter.

Key ideas

• A large sample alone does not make observations multivariate normal under the central limit theorem.
• The central limit theorem applies to averages under suitable conditions, not arbitrary data vectors.
• Serial correlation can undermine variance–covariance VaR assumptions and understate risk.
• Historical quantiles and bootstrap estimation are alternatives, with bootstrap also expressing uncertainty in the VaR estimate.
• VaR has limitations as a risk measure, so its assumptions and interpretation require care.

Tags

# Central limit theorem and VaR


# Central limit theorem and VaR












If I have a lot of data points and number of different dependent variables, can I use central limit theorem to assume data is multivariate normal and compute my VaR? Is this the appropriate use of central limit theorem for VaR calculation?

## Answer by Ram Ahluwalia (score 5)

https://quant.stackexchange.com/a/3107

There are several methods to compute VaR: i) historical, ii) variance-covariance, and iii) monte carlo. iv) copula techniques. I assume you are asking about approach (ii).

If the data are not multivariate normal and i.i.d. then the variance-covariance approach will not reflect true risk. For example, if there is serial correlation then risk is understated.

Your intuition around the use of the central limit theorem can be applied by using a bootstrapping approach to estimating VaR. This approach treats VaR itself as a random variable which is estimated with confidence. VaR has some highly unusual properties since it is not a coherent risk measure - so good luck! There are various posts on this site on how to do bootstrapping.

## Answer by Quartz (score 0)

https://quant.stackexchange.com/a/3111

If you are not taking a mean of many values (with finite variance) then the central limit theorem does not apply. To calculate VaR anyway you can start taking the empirical quantile or use more sophisticated estimators, as the other answer mentioned.

Sorry but it's not clear to me the role of those different dependent variables.

PS: please distinguish between VaR (Value at Risk) and VAR (vector autoregression)

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.