Diagnosing Segment-Level Calibration With Credibility

Why Segment-Level Calibration Needs Credibility

A portfolio can appear close to expectation overall while particular rating segments differ from their expected results. The reverse can also happen: a small segment can show a striking observed-to-expected ratio because a few outcomes have a large effect on its average. A useful diagnosis therefore asks two questions: how far is the segment from expectation, and how much information does its exposure provide?

For monitoring, compare losses and expected losses on a consistent basis: use the same observation period, exposure definition, rate level, and treatment of claim maturity and development. This tutorial focuses on interpreting the resulting segment comparisons. It does not prescribe changing a filed factor whenever a segment’s experience differs from expectation.

Definition: For a segment with observed losses $L$, expected losses $E$, and exposure $n$, the observed loss cost is $L/n$, the expected loss cost is $E/n$, and the observed-to-expected ratio is $L/E$, provided $E>0$. A ratio above 1 indicates observed losses above expected losses; a ratio below 1 indicates observed losses below expected losses.

The ratio is easy to interpret, but it does not describe how much random fluctuation to expect. Under a simple monitoring model, let $Y_j$ be the loss cost for exposure unit $j$. Assume that, conditional on a segment’s underlying mean loss cost $\theta$, the $Y_j$ are independent with mean $\theta$ and variance $v$. For a segment with $n$ exposure units, the observed loss cost is $\bar Y$, and its conditional variance is $v/n$. Thus, more exposure generally makes the observed loss cost more stable.

For a direct calibration comparison against a fixed expected loss cost $\mu$, the expected aggregate loss is $n\mu$. Under the model, the variance of aggregate loss is $nv$. A standardized residual is therefore

R=\frac{L-n\mu}{\sqrt{nv}}.

This calculation scales the dollar difference by the standard deviation expected at that exposure volume. It is a diagnostic under the stated model, not proof that a segment is or is not mispriced. In particular, a normal approximation for loss costs should be used only when it is a reasonable approximation for the data; skewed or sparse data may call for a more suitable predictive distribution.

Worked Example: Comparing Deviations at Different Exposure Volumes

A hypothetical insurer monitors two segments. Assume independent exposure-unit loss costs with conditional variance $v=\$90{,}000$ per exposure unit, and assume the expected loss cost is $\mu=\$400$ in both segments. Segment A has $n=4$ and observed losses of $\$2{,}600$. Segment B has $n=100$ and observed losses of $\$43{,}000$. Calculate each segment’s loss cost, observed-to-expected ratio, and standardized residual.

Method. Calculate expected aggregate losses as $n\mu$, then calculate the observed-to-expected ratio and divide the aggregate residual by $\sqrt{nv}$. This avoids ranking segments only by the size of their percentage difference.

For Segment A, expected losses are $4(\$400)=\$1{,}600$. Its observed loss cost is $\$2{,}600/4=\$650$, so its observed-to-expected ratio is $\$2{,}600/\$1{,}600=1.625$. The aggregate residual is $\$2{,}600-\$1{,}600=\$1{,}000$. Its standard deviation is $\sqrt{4(\$90{,}000)}=\$600$, giving $R=\$1{,}000/\$600=1.667$.

For Segment B, expected losses are $100(\$400)=\$40{,}000$. Its observed loss cost is $\$43{,}000/100=\$430$, and its observed-to-expected ratio is $\$43{,}000/\$40{,}000=1.075$. The aggregate residual is $\$3{,}000$. Its standard deviation is $\sqrt{100(\$90{,}000)}=\$3{,}000$, giving $R=\$3{,}000/\$3{,}000=1.000$.

Segment A has the larger ratio, but its residual is only 1.667 standard deviations under this model. Segment B’s smaller relative difference is 1.000 standard deviation. The comparison does not establish that either segment is miscalibrated; it shows why exposure volume belongs in the diagnosis.

Credibility-Adjusted Segment Estimates

A standardized residual measures the observed departure from a fixed expectation relative to process variation. Credibility asks a related but different question: how much should the segment’s observed experience influence an estimate of its underlying loss cost? The Bühlmann model provides a transparent way to answer that question by blending segment experience with a prior or portfolio-level expected value.

Model and assumptions: Suppose comparable segments have underlying means $\theta$ with $E[\theta]=\mu$ and $\operatorname{Var}(\theta)=a$. Conditional on $\theta$, assume the $n$ exposure-unit loss costs are independent, each with $E[Y_j\mid\theta]=\theta$ and $\operatorname{Var}(Y_j\mid\theta)=v$. Treat $a$ and $v$ as known for the calculation. Then the Bühlmann credibility estimate of the segment mean is $\hat\theta=\mu+Z(\bar Y-\mu)$, where $Z=na/(na+v)$.

Here $a$ describes variation in underlying means among comparable segments, while $v$ describes variation among individual exposure-unit results within a segment. The ratio $na/(na+v)$ is between 0 and 1 when $a$ and $v$ are positive. It increases with exposure $n$: more experience earns more weight. If $a=0$, there is no assumed variation in underlying means and the estimate remains at $\mu$. If $v=0$ and $n>0$, observed experience has no process noise under this model and receives full weight.

Why the Credibility Formula Holds

The credibility expression is the conditional mean under a normal prior and normal observation model. Specifically, add the assumption that $\theta\sim N(\mu,a)$ and that the observed average satisfies $\bar Y\mid\theta\sim N(\theta,v/n)$. The prior and likelihood are both normal. Their product gives a posterior mean that is a weighted average of $\mu$ and $\bar Y$.

E[\theta\mid\bar Y] = \frac{\mu/a+\bar Y/(v/n)}{1/a+1/(v/n)} = \frac{v\mu+na\bar Y}{v+na} = \mu+\frac{na}{na+v}(\bar Y-\mu).

The weight on $\bar Y$ is $Z=na/(na+v)$, as stated. The same weight is the Bühlmann credibility factor under the mean-variance assumptions even when the full normal distributional assumption is not made; in that setting it gives the linear credibility estimate, not necessarily an exact posterior mean. The normal version also assumes a continuous, unbounded parameter and can produce estimates outside a practical range. It is an approximation when that is inappropriate, not a reason to ignore parameter constraints or the data’s distribution.

Worked Example: Shrinking Two Segment Estimates

Use the segments in the previous example. Suppose comparable segments have between-segment variance $a=\$10{,}000$, within-segment variance $v=\$90{,}000$ per exposure unit, and prior expected loss cost $\mu=\$400$. Calculate each Bühlmann estimate. Use the same normal model and variance assumptions stated above.

Method. Find $Z$ from the segment’s exposure, then move the expected cost toward the observed cost by the fraction $Z$. This avoids using the raw segment cost as if it were equally reliable at every exposure volume.

For Segment A, $n=4$, so $Z=4(\$10{,}000)/[4(\$10{,}000)+\$90{,}000]=\$40{,}000/\$130{,}000=0.3077$, rounded to four decimals. Its observed cost is $\$650$, a difference of $\$650-\$400=\$250$ from expectation. The credibility estimate is $\$400+0.3077(\$250)=\$476.92$, rounded to the nearest cent.

For Segment B, $n=100$, so $Z=100(\$10{,}000)/[100(\$10{,}000)+\$90{,}000]=\$1{,}000{,}000/\$1{,}090{,}000=0.9174$, rounded to four decimals. Its observed cost is $\$430$, a difference of $\$30$ from expectation. The estimate is $\$400+0.9174(\$30)=\$427.52$, rounded to the nearest cent.

The smaller segment’s raw cost of $\$650$ is pulled substantially toward $\$400$; the larger segment’s estimate stays closer to its observed cost. The method avoids the common mistake of treating the larger raw ratio in the four-exposure segment as more reliable than the smaller deviation in the 100-exposure segment merely because the ratio looks dramatic.

Reading Several Segments Together

Credibility is useful when each segment has its own expected level. Compare observed and expected losses within each segment first; then assess the portfolio total separately. A portfolio total can conceal offsetting segment results, while the highest segment ratio can be dominated by a small denominator. The table below summarizes what the measures contribute.

Measure	What it describes	What it does not establish
Observed-to-expected ratio	Relative size and direction of the difference	Whether the difference is large compared with random variation
Standardized residual	Difference measured in process standard deviations under a specified model	That the expected level is wrong, or that a rate change is warranted
Credibility-adjusted estimate	A blend of segment experience and the expected level, weighted by exposure and variance assumptions	A guarantee that the blended estimate is the correct future rate

Worked Example: An Elevated Portfolio Ratio With Uneven Segment Evidence

A hypothetical book has three segments. Their expected loss costs are respectively $\$300$, $\$500$, and $\$800$. Exposures are $100$, $20$, and $5$, and observed losses are $\$31{,}000$, $\$12{,}000$, and $\$7{,}000$. Assume independent exposure-unit loss costs with variance $v=\$160{,}000$ within each segment. For credibility estimates, assume the between-segment variance of deviations from each segment’s expected level is $a=\$40{,}000$. Calculate the segment comparisons, portfolio ratio, and credibility estimates.

Method. Use each segment’s own expected cost as its baseline. Calculate the portfolio observed-to-expected ratio from totals, then calculate segment standardized residuals and credibility estimates. This prevents a different expected mix across segments from being mistaken for a within-segment calibration result.

Segment 1 has expected losses $100(\$300)=\$30{,}000$, observed cost $\$31{,}000/100=\$310$, and ratio $\$31{,}000/\$30{,}000=1.0333$. Its residual is $\$1{,}000$, and its standard deviation is $\sqrt{100(\$160{,}000)}=\$4{,}000$; therefore $R=\$1{,}000/\$4{,}000=0.2500$.

Segment 2 has expected losses $20(\$500)=\$10{,}000$, observed cost $\$12{,}000/20=\$600$, and ratio $1.2000$. Its residual is $\$2{,}000$, and its standard deviation is $\sqrt{20(\$160{,}000)}=\$1{,}788.85$; therefore $R=1.1180$, rounded to four decimals.

Segment 3 has expected losses $5(\$800)=\$4{,}000$, observed cost $\$7{,}000/5=\$1{,}400$, and ratio $1.7500$. Its residual is $\$3{,}000$, and its standard deviation is $\sqrt{5(\$160{,}000)}=\$894.43$; therefore $R=3.3541$, rounded to four decimals.

Across the portfolio, expected losses total $\$30{,}000+\$10{,}000+\$4{,}000=\$44{,}000$, and observed losses total $\$31{,}000+\$12{,}000+\$7{,}000=\$50{,}000$. The portfolio ratio is $\$50{,}000/\$44{,}000=1.1364$, rounded to four decimals. Its aggregate residual is $\$6{,}000$. Under the independence assumption, the aggregate variance is $(100+20+5)(\$160{,}000)=\$20{,}000{,}000$, so the standard deviation is $\$4{,}472.14$ and the standardized residual is $1.3416$.

For credibility, the weight for exposure $n$ is $Z=na/(na+v)$. Segment 1 has $Z=100(\$40{,}000)/[100(\$40{,}000)+\$160{,}000]=0.9615$, giving $\$300+0.9615(\$310-\$300)=\$309.62$. Segment 2 has $Z=20(\$40{,}000)/[20(\$40{,}000)+\$160{,}000)=0.8333$, giving $\$500+0.8333(\$600-\$500)=\$583.33$. Segment 3 has $Z=5(\$40{,}000)/[5(\$40{,}000)+\$160{,}000]=0.5556$, giving $\$800+0.5556(\$1{,}400-\$800)=\$1{,}133.33$. All estimates are rounded to the nearest cent.

Segment 3 has the largest ratio and residual relative to its modelled process variation, but its credibility estimate is still pulled toward its expected cost because it has only five exposures. The calculation is a diagnostic signal, not an automatic indication to change that segment’s rate. Investigation should consider the model assumptions, data quality, and whether the difference persists in relevant experience.

Common Mistakes and Practical Interpretation

Ranking only by observed-to-expected ratio. A ratio of 1.50 in a very small segment is not automatically stronger evidence than a ratio of 1.10 in a large one. Compare deviations with their exposure-scaled variation.
Using the wrong variance for an aggregate. If independent exposure-unit results each have variance $v$, aggregate variance is $nv$, not $v$ and not $n^2v$. The average has variance $v/n$.
Confusing the two questions. A standardized residual evaluates observed losses relative to a fixed expected level under a process model. A credibility estimate blends the observed cost with a prior expected cost under assumptions about between- and within-segment variation.
Over-interpreting a credibility estimate. The weight depends on the chosen $a$ and $v$, which should be estimated and reviewed appropriately. Credibility is not a substitute for validating exposure, expected loss, and claim adjustments.
Turning a statistical signal directly into a rate action. A segment departure can reflect random fluctuation, data issues, or a modelling limitation. A monitoring result merits investigation; it does not by itself justify a production change.

Key takeaway: Read segment calibration using the expected level, the size of the observed departure, and the exposure supporting it. Ratios show relative differences, standardized residuals scale differences for process variation, and credibility estimates shrink noisy segment experience toward an appropriate expected level.

Check Your Understanding

For each problem, use the stated assumptions and round monetary estimates to the nearest cent and ratios or standardized residuals to four decimals.

A segment has $n=12$, expected loss cost $\$250$, observed losses $\$3{,}600$, and process variance $v=\$14{,}400$ per exposure unit. Find its observed-to-expected ratio and standardized residual. (Answer: ratio $1.2000$; residual $1.4434$.)
A segment has expected cost $\mu=\$600$, observed cost $\bar Y=\$660$, exposure $n=25$, within-segment variance $v=\$40{,}000$, and between-segment variance $a=\$10{,}000$. Find its credibility factor and credibility estimate. (Answer: $Z=0.8621$; estimate $\$651.72$.)
Two segments have expected costs $\$400$ and $\$900$, respectively. Segment A has $n=64$ and observed cost $\$420$; Segment B has $n=4$ and observed cost $\$1{,}200$. With $v=\$90{,}000$ and $a=\$10{,}000$, calculate each credibility estimate and identify the segment with the larger raw observed-to-expected ratio. (Answer: A estimate $\$417.53$; B estimate $\$992.31$; Segment B has the larger raw ratio.)
A segment has $n=9$, expected cost $\$500$, and observed losses $\$5{,}850$. Given $v=\$40{,}000$ and $a=\$10{,}000$, calculate the observed-to-expected ratio and the credibility estimate. (Answer: ratio $1.3000$; estimate $\$603.85$.)
Two segments both have expected cost $\$300$, with $v=\$10{,}000$ and $a=\$5{,}000$. Segment P has $n=25$ and observed losses $\$8{,}000$; Segment Q has $n=1$ and observed losses $\$600$. Calculate both credibility estimates and state which has the larger raw ratio. (Answer: P estimate $\$318.52$; Q estimate $\$400.00$; Segment Q has the larger raw ratio.)

What You'll Learn