Kupiec unconditional coverage test
Compares observed VaR exceedance frequency with its nominal probability.
How it works
The Kupiec test compares the number of VaR exceedances against the number implied by the target tail probability. Its binomial likelihood ratio tests unconditional calibration. Violation dates and their ordering are outside the statistic. The horizon and number of observations should be defined before testing.
Worked example
A 99% one-day VaR forecast expects 2.5 exceedances in 250 comparable daily forecasts. If there are five, the empirical breach rate is 2%. A formal coverage test weighs that observation against the variability of a binomial count. Five adjacent breaches give the same Kupiec statistic as five separated breaches.
Validation and limitations
Compare test p-values with an exact binomial calculation where the sample is small. Document any overlapping forecasts and changes in exposure or reporting. Consider Christoffersen coverage for clustering and ES diagnostics for the magnitude of tail losses.