the credibility problem

Imagine you run a small bakery and want to set the price for next year's insurance against fire. The insurer has two numbers in front of it. One is your own bakery's record: last year you had no fire, the year before a small one. The other is the average for all bakeries in the country. Your own record is about you, which is exactly what you want — but it is built from only a year or two of data, so it jumps around wildly. The big average is rock-steady, but it is about everyone, not specifically about you. Which number should the price lean on? That tension is the credibility problem.

Put precisely, the credibility problem asks: given a single risk (one policyholder, one class, one territory) for which we have some, but limited, experience, how much should we trust that risk's own data versus a broader benchmark when estimating its true expected cost? Pure own-data is responsive but noisy; pure benchmark is stable but ignores what makes this risk special. Credibility theory is the body of methods that answers 'how much weight to give my own data' in a principled, defensible way, rather than by gut feeling.

This is one of the oldest practical problems in actuarial work and it shows up everywhere: experience-rating a large employer's group health plan, setting territory factors in auto insurance, or blending a new product's thin loss data with a related product's. The danger to avoid is the two naive extremes — overreacting to one lucky or unlucky year (giving own-data full weight when it is too thin) or ignoring genuine differences between risks (giving it no weight at all). Credibility is the disciplined middle ground.