limited fluctuation credibility

The oldest credibility idea, born in workers'-compensation ratemaking around 1914, starts from a modest goal: keep the estimate from bouncing around too much. The thinking is simple and human — 'I will trust my own data as much as I safely can, and the test of safety is that the data don't fluctuate wildly from year to year.' Because it is built to limit the random fluctuation of the estimate, it is called limited fluctuation credibility, also known as classical credibility.

Concretely, it sets a target: we want the observed experience to be within a small percentage (say, plus or minus 5 percent) of its true mean with high probability (say, 90 percent). If we have enough data that this target is met, the data are 'fully credible' and get full weight (Z = 1). If we have less, we get partial weight, and the famous square-root rule sets exactly how much. The whole machinery turns on counting how many claims (or how much exposure) you need before random noise is small enough to trust. For a Poisson claim count, full credibility for the frequency typically requires on the order of a thousand-odd expected claims.

Limited fluctuation credibility is intuitive, easy to compute, and still embedded in many regulatory and rating formulas. But be honest about its weakness: it never asks how different one risk truly is from another. It treats the benchmark as exactly right and only worries about the noise in your own data. That is why the more modern Buhlmann (greatest-accuracy) approach, which weighs real between-risk differences against within-risk noise, is theoretically preferred — though classical credibility survives because it is simple and 'good enough' in many practical settings.