nonparametric credibility estimation
/ non-pair-uh-MET-rik /
Buhlmann credibility needs the two structural numbers, EPV and VHM, before it can set Z. One way to get them is to assume a full probability model — say, claim counts are Poisson and rates follow a gamma — and read the variances off that model. But what if you don't trust any particular distribution? Nonparametric credibility estimation gets EPV and VHM straight from the raw data using sample variances, without committing to a named distribution at all. It lets the data speak for themselves.
The method is essentially analysis-of-variance thinking. To estimate the expected process variance, you compute each risk's own within-group sample variance and average them — that captures the year-to-year wobble inside risks. To estimate the variance of the hypothetical means, you measure how far each risk's average sits from the grand average (the total between-group spread) and then subtract the portion that mere process noise would create, since some apparent between-risk difference is just luck. There is also a middle path, semiparametric estimation, where you assume a distribution for the noise part (often Poisson, so EPV is pinned by the mean) but estimate VHM nonparametrically — handy when one piece is well understood and the other is not.
Nonparametric estimation is the standard, robust way Buhlmann-Straub is implemented when you would rather not bet on a distributional form. Its strength is exactly that freedom from assumptions; its honest weaknesses are two. First, the VHM estimator subtracts two noisy quantities and can come out negative, in which case it is floored at zero (forcing Z = 0 — a hint the risks are not credibly distinguishable). Second, with few risks or short histories the sample variances are themselves shaky, so the estimated Z inherits real uncertainty that the tidy final formula hides.
Across 50 group plans, averaging each plan's own claim-rate variance gives an EPV estimate; measuring how the 50 plan averages scatter, minus the process-noise share, gives a VHM estimate. No distribution was assumed — only sample variances were used.
Within-group variance estimates EPV; between-group spread minus noise estimates VHM — pure data, no model.
Nonparametric estimators avoid distributional assumptions but not sampling error: the VHM estimate can be negative (floored at zero) and is unstable with few risks. 'No model' does not mean 'no uncertainty'.