conjugate prior

Updating beliefs with Bayes' theorem can be messy: in general the posterior is some awkward shape that needs a computer to handle. But occasionally there is a magic pairing where the prior and the posterior belong to the same tidy family of distributions — observing data just shifts the prior's dial settings, never its shape. Such a prior is called a conjugate prior for that kind of data. It is the mathematical version of a key that fits a lock so cleanly the door swings open by hand.

Concretely, a prior distribution is conjugate to a likelihood if combining them via Bayes leaves a posterior in the same distributional family as the prior. The famous examples that drive credibility: a gamma prior is conjugate to Poisson claim counts (posterior is again gamma); a beta prior is conjugate to Bernoulli or binomial data (posterior again beta); a normal prior is conjugate to a normal mean (posterior again normal). In each, the posterior parameters are simple updates — for gamma-Poisson, you just add the observed claims to one parameter and the observed exposure to the other.

Conjugate priors are why credibility has such beautiful closed-form results. In exactly these pairings the Bayesian posterior mean turns out to be a clean linear blend — the Z-weighted credibility formula — so the Bayesian answer and the Buhlmann linear estimate coincide ('exact credibility'). The honest caveat: conjugacy is a mathematical convenience, not a fact about the world. Real claim data may not follow a gamma-Poisson story, and choosing a conjugate prior just because it is tractable can quietly bake in an assumption that does not hold. Modern computing (e.g. simulation) means we no longer need conjugacy to do Bayesian credibility — it is a clarifying special case, not a requirement.