JOVANA
Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
All guides

The Credibility Problem

Next year's price for one small group: trust their own thin, jumpy experience, or lean on the big book's steady average? Credibility theory refuses to choose all-or-nothing and instead blends the two — and tells you exactly how much weight the small group has earned.

A real desk, a real dilemma

You price workers-compensation insurance for a single bakery. Across your whole book of thousands of employers, the average loss runs about $1,000 per worker per year — a number you trust, because it rests on an enormous pile of data. But this one bakery has been with you only three years, employing a handful of people, and over those three years its own losses have averaged just $400 per worker. So what should next year's price lean on: the bakery's own cheerful $400, or the book's sober $1,000? That single question is the whole of this rung, and it has a name — the credibility problem.

Two reflexes tempt you, and both are wrong. The first says: 'It's *their* data, charge them their own $400.' But three years from a handful of workers is painfully thin — recall from the estimation guides that an estimate is itself a random variable, and a small sample makes it jump around violently. That $400 could easily have been $1,300 in a slightly unluckier draw. The second reflex says: 'Ignore them, charge everyone the book's $1,000.' But that throws away a genuine signal — maybe this bakery really is safer than average, and pretending otherwise is both unfair and a quiet invitation for its better-run rivals to leave you.

Refuse the false choice: blend

The escape from the two bad reflexes is almost embarrassingly simple, and yet it is one of the most beautiful ideas an actuary owns. Do not choose. Blend. Take a weighted average of the group's own number and the broader average, and let a single weight — call it Z, a number between 0 and 1 — decide how the credit is split. This is the credibility-weighted estimate, and it is worth committing its shape to memory because every model in this rung is just a different recipe for Z.

Estimate = Z * (your own data) + (1 - Z) * (the prior/book average)

Bakery example, book = 1000, own data = 400:
  Z = 0    -> 0*400 + 1*1000 = 1000   (ignore the bakery entirely)
  Z = 1    -> 1*400 + 0*1000 =  400   (trust the bakery entirely)
  Z = 0.25 -> 0.25*400 + 0.75*1000 = 850   (a little credit, mostly the book)
  Z = 0.75 -> 0.75*400 + 0.25*1000 = 550   (mostly the bakery)
The one formula behind the whole rung. Z slides the answer between the book average and the group's own data; the only real question left is what Z should be.

Notice what the blend buys you. With Z = 0.25 the bakery pays $850 — below the book's $1,000, so its three good years are rewarded, but nowhere near the reckless $400 that one lucky stretch might not survive. The estimate is pulled toward the book average, or 'shrunk,' and the smaller and noisier the group's own data, the harder it gets pulled. That pull is not a fudge or a hedge; we will see it is the statistically optimal thing to do, the price you rationally pay for not yet knowing whether $400 was skill or luck. The whole credibility-weighted estimate is this disciplined act of partial trust.

So what should Z depend on?

Everything interesting now lives in Z, and before any formula, your intuition can already list what it ought to respond to. Three pressures push Z up toward 1 (trust the group) or down toward 0 (trust the book). Sit with each one — when the formulas arrive in the next guides, they are simply these instincts made precise.

  1. How much data does the group have? More exposure — more workers, more years, more claims — means the group's own average is a sharper, less jumpy estimate, so Z should rise toward 1. A bakery seen for thirty years deserves far more credit than one seen for three. This is the law of large numbers doing exactly what it always does: averages over more independent observations settle down.
  2. How noisy is a single observation? If losses within this kind of business swing wildly from year to year — a few huge claims among many tiny ones — then even a decent stretch of data is unreliable, and Z should be held down. The noisier each data point, the more of it you need before you believe the average. This is the same square-root-of-sample-size humility that governed the standard error back in the estimation rung.
  3. How different are the groups from each other? If every bakery in your book is genuinely alike, then this one's deviation to $400 is almost surely just noise, and you should lean hard on the common $1,000 — Z near 0. But if bakeries truly differ a lot in their underlying riskiness, then a low number more likely reflects a real, persistent difference worth charging for, and Z should rise. The book average is only as informative as the groups are similar.

The first two pressures are about the group's own data — its quantity and its inherent noise. The third is subtler and is what separates the two great schools you are about to meet. The oldest approach, limited-fluctuation credibility, looks only at the first two: it asks how much data the group needs before its own number is stable enough to lean on, and ignores the third pressure entirely. The deeper Bühlmann credibility that follows weaves all three together into a single, principled Z. That progression — from a practical rule of thumb to a theory that explains *why* — is the arc of this entire rung.

What the prior really is

One word in the formula deserves care: the (1 − Z) weight rides on the prior — your belief about this group *before* you looked at its own data. In our bakery it was the book-wide $1,000, but the prior is whatever you would honestly have charged knowing nothing about this particular bakery beyond its being a bakery. It might be the average for all bakeries specifically, or for all food-service employers, or the grand average of the entire book. Choosing the right comparison group is a real judgement, not a given, and a sloppy prior quietly poisons every blended answer.

There is also a graceful symmetry hiding in the blend. When you trust the group fully (Z = 1), the prior drops out and you are doing pure experience rating. When you trust it not at all (Z = 0), your own data drops out and you are doing pure manual rating off the prior. Every real answer lives on the continuum between these two extremes — credibility theory does not invent a new kind of estimate so much as it draws the honest line connecting two old ones, and tells you where on that line your data has earned the right to stand.

Honest limits, and what comes next

Guard, too, against two common misreadings. Z is not a probability that the group's number is 'correct' — it is a weight, an optimal mixing proportion, and nothing more. And a low Z is not a punishment or a verdict that the bakery is secretly risky; it is simply an admission that three years cannot yet outvote thousands of employers' worth of evidence. Used honestly, credibility is one of the rare tools that is *fair by construction*: it gives a group exactly as much say in its own price as its data has earned, no more and no less.

You now hold the whole rung in one sentence: the answer is Z times your data plus (1 − Z) times the prior, and the only craft left is choosing Z well. The next guide makes the first serious attempt — limited-fluctuation credibility, which asks the disarmingly direct question 'how much data is *enough* to be believed fully?' and builds the famous square-root rule around it. After that, Bühlmann's Z will give the deep, optimal answer that quietly contains the others as special cases. Keep the bakery in mind the whole way; every formula ahead is just a sharper tool for setting its price fairly.