kernels.cdr_risk_measures()
VaR and TVaR of the one-year LOSS, from simulated CDR draws.
Usage
kernels.cdr_risk_measures(
pred,
levels=(0.995,),
)The capital question is asked on the adverse side, so everything here is stated on the loss -CDR (the reserve strengthening): VaR_0.995 is the 99.5th percentile of that loss, the Solvency II reserve-risk basis, and TVaR_0.995 its mean beyond that point. A negative VaR means even the adverse tail at that level is still a release.
Quantiles are exact empirical order statistics of the draws, so the tail knots are as good as the draw count and no better - at 20k draws the 99.5th percentile rests on 100 observations. Raise n_draws before reading much into 99.9.
Needs the simulated distribution rather than the analytic msep: a closed form gives a second moment, and no second moment implies a quantile.