Buhlmann credibility factor k

In Buhlmann credibility, the weight on your own data is Z = n/(n + k), where n is how many observations you have. The number k is the gatekeeper. It tells you, in units of observations, how much data you need before you start to trust the individual: when n equals k, Z is exactly one half. Think of k as 'the price of credibility' — a small k means cheap (your data win quickly), a large k means expensive (you stay tied to the benchmark for a long time).

k is not a free knob; it is fixed by the two structural parameters as k = EPV / VHM — the expected process variance divided by the variance of the hypothetical means. Read it as a signal-to-noise story turned upside down: the noise inside each risk sits on top, the real signal (how much risks differ) sits underneath. When risks barely differ (tiny VHM) or each one's data are very noisy (big EPV), k is large, Z is small, and you lean on the collective. When risks differ a lot relative to their internal noise, k is small, Z is large, and the individual's own data carry the rate. Example: EPV = 0.4, VHM = 0.05 gives k = 8, so a risk needs 8 years to reach Z = 0.5.

Because k bundles all the model's structure into one constant, estimating it well is the crux of applied Buhlmann credibility, normally via empirical Bayes. Two cautions: k has units (per observation, per exposure unit) so it only makes sense alongside how n is measured; and k is an estimate built on EPV and VHM, which themselves carry sampling error — so treat the resulting Z as a reasoned figure, not a precise constant of nature.