expected value of the process variance

Take a single driver whose true accident rate is fixed. Even so, in some years she has no accidents, in others one or two — pure luck wobbling around her own true rate. That year-to-year wobble, the noise inside one risk whose true value is held fixed, is the process variance. The expected value of the process variance, EPV, is that wobble averaged over all the risks in the population — a measure of how noisy an individual risk's data are, on average.

Formally, for a risk with hidden parameter theta, the process variance is Var(claims given theta) — the spread of outcomes when the risk's type is known. We almost never know theta, so we average this over the distribution of risk types: EPV = E[ Var(claims | theta) ]. It is the 'within-risk' or 'noise' piece. A small numerical picture: if knowing a driver is high-risk still leaves a year's count varying with variance 0.5, and a low-risk driver's count varies with variance 0.3, and the two types are equally common, then EPV = (0.5 + 0.3)/2 = 0.4.

EPV is one of the two structural parameters of Buhlmann credibility (the other is the variance of the hypothetical means). It is the denominator's friend: large EPV means each risk's own data are very noisy, which pushes the Buhlmann constant k = EPV/VHM up and the credibility Z down — you trust the individual's data less because so much of what you see is luck. Estimating EPV from data (often as the average within-risk sample variance) is a core step in empirical-Bayes credibility.