Let us assume that each month of the year (up to November) we calculate a VaR (say 99%) with holding period to the end of the year. Thus the holding period starts with 12 months and goes down to 1 month in November.

Obviously the event of a breach of my June VaR is not independent of the event of a breach of my Augus VaR (similar to here).

So how can I calculate the significance of the number of breaches? How can I take the dependence into account?

PS: At first sight it looks a bit artificial but a lot of institutions want to measure their risk in an (accounting) year. The next step is VaR and its backtest.

If look at monthly log-returns then in October I estimate $$ P[R_1+ R_2 + R_3 \le VaR_1] $$ and in November I look at $$ P[R_2 + R_3 \le VaR_2]. $$

Another edit: Say I estimated my VaR in Jan, Feb, up to Novenber and in December the market drops by 50%. Then I can have a breach in all my VaR estimates.

In the usual setting I estimate VaR and check on the following period. Then I estimate VaR again and check on the next period. Thus there is no overlap and the above problem can not happen!

I could only form series of "estimate yearly loss in Month x" and e.g. look at all January VaRs for the year 2000, 2001, .. 2016. They would be (more or less) independent. Then I could look at Februaries ... this would take away the overlap. But I would only have a few observations.

  • $\begingroup$ if your model is correct the event of a breach of your June VaR should be independent of the event of a breach of your Augus VaR. $\endgroup$
    – Malick
    Commented Apr 22, 2016 at 10:30
  • 1
    $\begingroup$ No, because the indicator of the event "return from July to December less than x" is not indepenend from the event "return from August to December less than y" ... so I think there is something to be taken care of. Please see my edit., $\endgroup$
    – Richi Wa
    Commented Apr 22, 2016 at 10:47

2 Answers 2


If I'm correct Backtesting VaR usually boils down to two conditions:

  • The unconditional coverage hypothesis : the probability of an ex-post violation must be equal to the coverage rate. (ie : if 0.01 confidence level, you should get 1% violation). You can test it with the Kupiec Test .

  • The independence hypothesis, your VaR violations should be independent. (ie : you shouldn’t observe violations clustering- past violations should not influence current and future violations.) Engle and Manganelli (2004) develop a test for it.

So, normally if your model is correctly specified you shouldn’t observe dependence in your series of violations.

The point I think you underestimate is that every month you re-estimate your model with new data. It means that your $VaR_{2}$ forecast will implicitly take into account your past violations. Your $VaR_{2}$ already integrate your $VaR_{1}$ violations since your $VaR_{2}$ is based on a larger sample. So you can not consider a kind of square root of time method to extend your VaR forecast because you need to re-estimate VaR parameters every months.

For more details about the independence hypothesis : Christoffersen, Peter, Backtesting (November 19, 2008). Available at SSRN: http://ssrn.com/abstract=2044825 or http://dx.doi.org/10.2139/ssrn.2044825

Edit :

I’m not sure I really got you point but I just clarify my talk :

  • If you perform a rolling window VaR approach, ie every month you are estimating the next month VaR : then you perform 1 Month horizon VaR Forecast and not yearly VaR evaluation. There is no overlap. When you evaluate your VaR accuracy your are performing 11 VaR backtesting on a single period (one month) window. In doing so you can not obtain « yearly (calendar) loss estimate »
  • If you perform VaR estimation with different horizon but a fixed ending date (the calendar year), then to evaluate the accuracy of you VaR forecast, for each month the accuracy is evaluated also on a different (decreasing) window. That means that a large drop in December may be classified as a violation in the January VaR estimation, not classified as a violation in February VaR estimation and again as a violation in a March VaR estimation. The accuracy of your backtests should be conducted separately for each horizon , then your are performing 11 VaR backtesting on 11 diffferent periods. Edit : you cannot evaluate the VaR accuracy every month, you have to wait the end of the calendar year to evaluate the 11 VaR.
  • $\begingroup$ Thanks for your answer. The first part is the usual VaR backtesting. The second paragraph is clear. But e.g. if I have a suprising loss of 40% in December then all my VaR esimates from Jan - November will be breached at once ... on the other side in the standard setting I calculate VaR and check on the next periods return, then I loop this. The above can not happen. There is something different $\endgroup$
    – Richi Wa
    Commented Apr 22, 2016 at 11:05
  • $\begingroup$ In your rolling window approach you are computing VaR on a monthly basis, so for sure a drop in December won’t impact your previous VaR. But if you need to compute your VaR in January for a longer horizon (+ one month), the rolling window approach is not applicable. $\endgroup$
    – Malick
    Commented Apr 22, 2016 at 11:35
  • $\begingroup$ Please see my edit. $\endgroup$
    – Richi Wa
    Commented Apr 22, 2016 at 11:37
  • $\begingroup$ What is exactly your question? You should only compare your January (11month horizon) VaR with the corresponding 11 Month Period, next for February (10 month horizon) VaR you compare with 10 month period and so on.. Your "estimate yearly loss in Month x" are not directly comparable because the horizon is different ( ie your out-of-sample period decreases over time). $\endgroup$
    – Malick
    Commented Apr 22, 2016 at 12:12
  • $\begingroup$ The question is: can I use all VaR esimates during a year (in the above sense) to asses the quality of my VaR esimate. And no - I think that the usual tests can not (!) be applied because of the overlap ... $\endgroup$
    – Richi Wa
    Commented Apr 22, 2016 at 12:14

I know this is an old question, but I encountered this problem too, years ago, and the answer is that you can actually compute confidence intervals for VaR backtesting of overlapping returns.

From what I understand, banks usually derive them with heavy Monte Carlo simulations (so they generate the overlapping returns a large amount of times and look at the excesses compared to a percentile and take the confidence intervals based on those excesses' distribution). This is actually a quite fine approach given the ready available computation power nowadays.

I had a slightly better semi-analytical approach that I had written up here:



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.