Of course the issue here is dependence: can it be removed or accounted for (in independence tests too, which of course would be troublesome)? There's a lot of literature on regression in this setting, but little (and unconvincing) on VaR backtesting... We don't want to use near horizons for various (mostly obvious) reasons.

  • $\begingroup$ Is your model Gaussian or Non-gaussian? $\endgroup$
    – pyCthon
    Sep 28, 2012 at 16:10
  • $\begingroup$ Actually non-gaussian, but the intricacies there would be probablys significantly higher, so it'd be already nice to have something for the simpler case. I'm looking for perspectives on how to approach the issue, more than specific methods; and one can often devise atleast approximate extensions later anyway. $\endgroup$
    – Quartz
    Oct 1, 2012 at 9:07
  • $\begingroup$ why not just use a kupiec test? $\endgroup$
    – pyCthon
    Oct 2, 2012 at 19:58
  • $\begingroup$ Also Kupiec tests assume independence. Consider the extreme case of constant backtesting interval=perfect dependence and you get a bernoulli output which does not really allow for much inference. There are fixes, but what I saw was either too crude or too conservative; and after all the Kupiec test is not so good anyway. $\endgroup$
    – Quartz
    Oct 17, 2012 at 17:33


Your Answer

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

Browse other questions tagged or ask your own question.