The question credibility answers
You have two estimates of what a risk should cost. One is the account's own experience, which is directly relevant but may be noise. The other is the class rate, which is built on far more data but describes an average the account may not resemble. Credibility theory answers the only sensible question: how much weight should each get?
The answer is a single number Z between 0 and 1, and the estimate is Z × own experience + (1 − Z) × class rate. Because Z is bounded by 0 and 1, the blended figure always lies between the two rates you entered — credibility weighting can never produce an estimate outside the range of its inputs.
The intuition behind the formula is the standard error of a mean. Observe n Poisson claims and the standard deviation of the count is √n, so the relative error is √n ÷ n = 1 ÷ √n. Accuracy improves with the square root of the data, not with the data, and that square root is exactly what appears in the credibility formula. Quadrupling the claim count doubles the credibility.
This machinery is everywhere in insurance pricing even when it is not named. It is the weighting value in the workers comp experience modifier calculator, the blending of an account indication against a class rate built with the pure premium calculator, and the judgement about whether a single year's loss ratio means anything at all.
Two methods, and what each one actually assumes
Limited fluctuation credibility, also called classical credibility, starts from a testable statement: the observed loss rate should fall within ±k of the true rate with probability p. For Poisson claim counts that requires nfull = (z ÷ k)² claims, where z is the standard normal quantile at (1 + p) ÷ 2. At 90% confidence and 5% tolerance, z = 1.6449 and the standard is (1.6449 ÷ 0.05)² = 1,082 claims — the figure quoted in every credibility text. At 95% and 5% it is 1,537. Below the standard, credibility is assigned by the square-root rule Z = √(n ÷ nfull), which is chosen because it keeps the variance of the weighted estimate under control, and it is capped at 1.
The method's honest weakness is that the confidence and tolerance are arbitrary. Nothing in the mathematics says 90% and 5% are right; they are conventions. Changing the tolerance from 5% to 2.5% quadruples the required claim count, because the standard goes as 1 ÷ k².
Bühlmann credibility asks a different and better-posed question: which weight minimises the expected squared error of the estimate? The answer is Z = n ÷ (n + K), where K is the ratio of the expected process variance to the variance of hypothetical means. In words: K is large when individual risks are noisy but similar to one another, so their own data tells you little; K is small when risks are stable but genuinely different from each other, so their own data tells you a lot.
Two consequences distinguish the methods. Bühlmann's Z approaches 1 asymptotically and never reaches it, so there is no such thing as full credibility — a fact that makes actuaries prefer it and makes rating plans avoid it. And K has a clean interpretation this calculator reports directly: it is the claim count at which Z equals exactly one half.
Worked example: 350 claims against the 1,082 standard
An account has produced 350 claims over its experience period. Its own indicated loss rate is $480 per exposure unit; the class rate is $400. The underwriter uses limited fluctuation credibility at 90% confidence and 5% tolerance.
- Find the normal quantile. 90% confidence two-sided means the quantile at the 95th percentile: z = 1.644854.
- Full credibility standard. (1.644854 ÷ 0.05)² = 32.89707² = 1,082.2 claims.
- Credibility factor. √(350 ÷ 1,082.2) = √0.323407 = 0.5687.
- Blend the rates. 0.5687 × $480 = $272.97, and 0.4313 × $400 = $172.52. The credibility-weighted rate is $445.50.
Read what that did. The account's own experience says $480, which is 20% above the class rate. The blend accepts a little over half of that difference: $445.50 is $45.50 above the class rate, which is 56.87% of the $80 gap — exactly Z, as it must be, since the weighted rate is class + Z × (own − class).
Now note how hard full credibility is to reach. This account needs 1,082 claims and has 350. To raise its credibility to 1.0 it would need 1,082 claims, 3.09 times as many as it has — because Z goes as the square root, multiplying Z by 1 ÷ 0.5687 = 1.758 requires multiplying n by 1.758² = 3.09. That square-root relationship is why full credibility is out of reach for all but the largest accounts, and why almost every real rate is a blend.
How to read the credibility factor
Z below about 0.3 means the account's own experience is barely moving the answer: the estimate is essentially the class rate with a nudge. That is the correct treatment of a small account, and it is also why a small employer with one terrible year does not see its rate triple.
Z above about 0.75 means the account is largely being priced on itself, and the class rate has become a sanity check rather than a driver. At Z = 1 the class rate has no influence at all.
The weight-on-class-rate output is the same information from the other side, and it is the more useful framing in a negotiation with an insured who believes their own experience should count for everything. It rarely should, and the number says by how much.
Two cautions on interpretation. First, credibility says nothing about whether your own loss rate is right — it says how much of it to believe. If the underlying losses have not been developed to ultimate, a high Z confidently propagates an understated figure. Develop the losses first with the method in the IBNR reserve calculator, then weight them. Second, the class rate is only a valid complement if the account genuinely belongs to that class. Blending against the wrong class produces a precise-looking estimate of the wrong thing.
Full credibility standards at common confidence and tolerance settings
| Confidence | z | k = 10% | k = 5% | k = 2.5% | k = 1% |
|---|---|---|---|---|---|
| 90% | 1.644854 | 271 | 1,082 | 4,329 | 27,055 |
| 95% | 1.959964 | 384 | 1,537 | 6,146 | 38,415 |
| 99% | 2.575829 | 663 | 2,654 | 10,616 | 66,349 |
Each cell is (z ÷ k)² rounded to the nearest claim, using the same quantiles the calculator uses. Note the 1/k² relationship: halving the tolerance quadruples the claim count, which is why demanding tight accuracy from small accounts is hopeless.
The Poisson standard understates the requirement when severity varies
The (z ÷ k)² standard covers only the variance in the number of claims. Where you are estimating aggregate loss rather than frequency, the variability of claim size adds to the total, and the standard becomes (z ÷ k)² × (1 + CV²), where CV is the coefficient of variation of the severity distribution. For a line whose severity has a CV of 2 — not unusual for liability — that multiplies the requirement by 5, so the 1,082-claim standard becomes 5,411. This calculator reports the frequency-only standard, which is the one credibility tables normally quote. If you are weighting aggregate loss rates rather than frequencies, multiply the standard by (1 + CV²) yourself before comparing your claim count against it.
Assumptions and common mistakes
- Claim counts, not dollars. The standard is expressed in claims. Entering a loss amount where a claim count belongs produces an enormous Z that means nothing.
- Poisson claim counts. The square-root rule assumes the count is Poisson, so the variance equals the mean. Where claims cluster — one event producing many claims — the effective count is lower than the raw count and credibility is overstated.
- Undeveloped losses. Recent years are not fully reported. A high credibility factor applied to an undeveloped loss rate confidently prices the wrong number.
- The wrong complement. The class rate must be a genuine alternative estimate of the same quantity on the same exposure base. A countrywide rate blended against a state-specific own experience mixes two different things.
- Confidence and tolerance are conventions. Nothing in the mathematics selects 90% and 5%. They are the historical choice; state which you used, because the answer moves a great deal with them.
- Full credibility is not certainty. Z = 1 means the observed rate meets the stated accuracy standard, not that it is correct. At 90% confidence, one estimate in ten still falls outside the tolerance.
Where credibility sits in the wider ratemaking process
Credibility is the last step in a rate indication, not the first. The sequence runs: gather losses, develop them to ultimate, trend them to the period the rate will cover, adjust historical premium to current rate level, calculate the indicated rate or rate change, and only then weight that indication against a broader complement. Applying credibility to raw, undeveloped, untrended data is a common shortcut and it converts a defensible technique into a way of dressing up noise.
The choice of complement matters as much as the choice of Z. For an individual account the complement is usually the class rate. For a class it may be the rate for a broader group of classes, or the current rate adjusted for trend. For a state it may be countrywide experience. The general principle is that the complement should be an independent, less volatile estimate of the same quantity — and its own error enters the blend with weight (1 − Z), which is why a bad complement is worst precisely when Z is low.
Between the two methods, Bühlmann is theoretically superior and limited fluctuation is what filed rating plans mostly use, because the confidence-and-tolerance framing is easier to explain to a regulator and the full credibility standard gives a defensible bright line. In practice the two rarely disagree by enough to change a decision, and the more consequential judgements — loss development, trend selection, choice of complement — sit upstream of both. Use this calculator to make the weighting explicit; do not let it distract from the estimate it is weighting.