Assessing Credibility of Territory Experience

Why the Size of a Territory Difference Is Not Enough

Territory experience can vary substantially from one observation period to another, even when the underlying risk is unchanged. A small territory may show a high loss cost because a few claims happened to occur, or because one claim was unusually large. A larger territory with a similar indication may provide more stable evidence. The observed difference matters, but so does the amount and variability of the experience supporting it.

Assessing credibility at this stage means judging how much random fluctuation could plausibly affect an indication. It does not, by itself, decide how to adjust a territory factor or select a complement. Those decisions require additional considerations. The methods here are aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus.

Definition: Credibility assessment is the evaluation of how reliable an experience-based indication is as evidence about an underlying risk level. Exposure volume, claim frequency, and loss variability affect its reliability. A higher indication is not necessarily more dependable merely because it is higher.

The comparison should be made on a consistent basis. For example, territorial indications need to reflect comparable coverage, time periods, exposure definitions, and adjustments for risk mix. The earlier material on standardizing territory experience for mix and choosing a territory experience measure addresses those foundations. If the underlying bases differ, an apparent difference can reflect the comparison rather than a true territory effect.

Exposure Volume and Claim Frequency

Start with a frequency indication: claim count divided by exposure. Under a simple model, claims occur independently at a constant rate $\lambda$ per exposure unit. For $E$ comparable exposure units, the claim count $N$ is Poisson with mean and variance $\lambda E$. The estimated frequency is $\hat{\lambda}=N/E$.

N\sim\operatorname{Poisson}(\lambda E),\qquad \hat{\lambda}=\frac{N}{E},\qquad \operatorname{Var}(\hat{\lambda})=\frac{\lambda}{E}.

The variance formula follows because dividing a random variable by $E$ divides its variance by $E^2$: $\operatorname{Var}(N/E)=\lambda E/E^2=\lambda/E$. Since $\lambda$ is unknown, a common estimated standard error is $\sqrt{N}/E$. The standard error measures estimated random fluctuation in the frequency indication, not uncertainty from every possible source. As exposure grows, this measure generally falls, all else equal.

Worked Example: Comparing Frequency Indications

Suppose two hypothetical territories have comparable exposure units and claim definitions. Territory North has 120 claims in 6,000 exposures; Territory South has 30 claims in 1,000 exposures. Assume independent Poisson claim counts. Compare the estimated frequencies and their estimated standard errors.

Method. Divide claim count by exposure for each frequency estimate. Estimate its standard error by $\sqrt{N}/E$.

For North, $\hat{\lambda}_N=120/6{,}000=0.0200$ claims per exposure. Its standard error is $\sqrt{120}/6{,}000=10.9545/6{,}000=0.001826$. Relative to its estimate, this is $0.001826/0.0200=9.13\%$.

For South, $\hat{\lambda}_S=30/1{,}000=0.0300$ claims per exposure. Its standard error is $\sqrt{30}/1{,}000=5.4772/1{,}000=0.005477$. Relative to its estimate, this is $0.005477/0.0300=18.26\%$.

South’s estimated frequency is higher, but its estimated standard error is also larger, both in absolute terms and relative to its indication. For another perspective, if the two counts are independent, the estimated standard error of their difference is $\sqrt{0.001826^2+0.005477^2}=0.005774$. The observed difference is $0.0300-0.0200=0.0100$, or about $0.0100/0.005774=1.73$ estimated standard errors.

This calculation describes sampling variation under the stated model; it does not establish that the territories have identical or different underlying risks. Common mistake avoided: treating South’s higher raw frequency as equally dependable despite its much smaller exposure volume.

Loss Cost Variation Depends on Severity as Well

A loss cost per exposure combines frequency and severity. A territory can have a stable claim count but a volatile loss cost if its claims vary widely in size. To quantify both sources, consider a compound Poisson model: claim count $N$ is Poisson with mean $\lambda E$, and claim amounts $X_1,\ldots,X_N$ are independent and identically distributed, independent of $N$, with finite second moment. Let total loss be $S=\sum_{j=1}^{N}X_j$. This model excludes dependence between claims, such as a common event affecting many insureds.

E[S]=\lambda E\,E[X],\qquad \operatorname{Var}(S)=\lambda E\,E[X^2],\qquad \operatorname{Var}\left(\frac{S}{E}\right) =\frac{\lambda E[X^2]}{E}.

The expected loss cost per exposure is $\lambda E[X]$. To justify the variance result, condition on the claim count. Given $N$, the expected total is $N E[X]$ and its variance is $N\operatorname{Var}(X)$. The law of total variance gives

\begin{aligned} \operatorname{Var}(S) &=E[\operatorname{Var}(S\mid N)] +\operatorname{Var}(E[S\mid N])\\ &=E[N]\operatorname{Var}(X)+\operatorname{Var}(N)(E[X])^2\\ &=\lambda E\left(\operatorname{Var}(X)+(E[X])^2\right) =\lambda E\,E[X^2]. \end{aligned}

The final equality uses $E[X^2]=\operatorname{Var}(X)+(E[X])^2$. Dividing total loss by exposure divides its variance by $E^2$, yielding the loss-cost variance above. The formula shows why both claim frequency and the severity second moment matter: greater severity dispersion increases $E[X^2]$, even if average severity is unchanged.

Worked Example: Equal Expected Loss Costs, Different Variability

Consider two hypothetical territories, each with an expected claim frequency of 0.05 per exposure and an average claim amount of $4,000. North has 2,000 exposures and claim severity standard deviation $2,000. South has 500 exposures and severity standard deviation $8,000. Assume the compound Poisson model above and no adjustment differences. Compare their expected loss costs and standard errors.

Method. The expected loss cost is frequency times mean severity. For variability, calculate $E[X^2]=\operatorname{Var}(X)+(E[X])^2$, then use $\operatorname{SE}(S/E)=\sqrt{\lambda E[X^2]/E}$.

For both territories, the expected loss cost is $0.05\times \$4{,}000=\$200$ per exposure.

For North, $E[X^2]=\$2{,}000^2+\$4{,}000^2=\$4{,}000{,}000+\$16{,}000{,}000=\$20{,}000{,}000$. Its loss-cost variance is $0.05\times \$20{,}000{,}000/2{,}000=\$500$ squared, so the standard error is $\sqrt{\$500}=\$22.36$ per exposure. The standard error is about $22.36/200=11.18\%$ of the expected loss cost.

For South, $E[X^2]=\$8{,}000^2+\$4{,}000^2=\$64{,}000{,}000+\$16{,}000{,}000=\$80{,}000{,}000$. Its loss-cost variance is $0.05\times \$80{,}000{,}000/500=\$8{,}000$ squared, so the standard error is $\sqrt{\$8{,}000}=\$89.44$ per exposure. This is about $89.44/200=44.72\%$ of the expected loss cost.

The territories have the same modeled expected loss cost, but South’s indication is much more variable because it has less exposure and more variable claim severity. Common mistake avoided: judging reliability from claim frequency or average severity alone while ignoring the severity distribution and exposure volume.

Estimating Variability From Observed Losses

In practice, the underlying frequency and severity moments are not known. Under the same compound Poisson assumptions, an experience-based estimate of the variance of loss cost is the sum of squared claim amounts divided by exposure squared. If $x_1,\ldots,x_N$ are the observed claim amounts, then $\sum x_j^2/E^2$ estimates $\operatorname{Var}(S/E)$. This follows because $E[\sum X_j^2]=\lambda E\,E[X^2]$, the numerator in the theoretical variance formula.

Worked Example: Using Observed Claim Amounts to Assess Stability

A hypothetical territory has 1,000 exposures, 50 claims, total observed loss of $150,000, and sum of squared claim amounts of $1,000,000,000. Assume the claim amounts are on a common basis and the compound Poisson model is appropriate. Estimate its loss cost and the standard error attributable to process variation.

Method. The observed loss cost is total loss divided by exposure. Estimate its variance as the sum of squared claim amounts divided by exposure squared.

The observed loss cost is $\$150{,}000/1{,}000=\$150$ per exposure. The estimated variance is $\$1{,}000{,}000{,}000/1{,}000^2=\$1{,}000$ squared per exposure. Therefore the estimated standard error is $\sqrt{\$1{,}000}=\$31.62$ per exposure, or $31.62/150=21.08\%$ of the observed loss cost.

A moment-based check gives the same result. The estimated frequency is $50/1{,}000=0.05$, and the average claim is $\$150{,}000/50=\$3{,}000$. The estimated second severity moment is $\$1{,}000{,}000{,}000/50=\$20{,}000{,}000$. Thus the estimated variance is $0.05\times \$20{,}000{,}000/1{,}000=\$1{,}000$, as above. Common mistake avoided: dividing the sum of squared claims by exposure only once; estimating the variance of a per-exposure quantity requires division by $E^2$.

Interpreting an Assessment Carefully

A standard error is a measure of random process variation under a model, not a complete measure of uncertainty in a territorial indication. It does not automatically account for uncertainty in loss development, trend, coverage changes, exposure measurement, or risk-mix adjustments. Nor does a small standard error protect against systematic bias. If one territory’s exposures are not comparable with another’s, more data do not repair the comparison.

The compound Poisson calculation also assumes independent claims. A catastrophe or other shared event can cause claims to move together, making actual variation larger than an independence-based estimate. A few unusually large claims can also make a loss-cost indication unstable. Review the underlying experience and the conditions that generated it; do not rely on a summary measure without considering whether its assumptions fit.

Useful comparisons include the absolute standard error, the standard error relative to the indication, and the relationship between an observed difference and the uncertainty in that difference. These are diagnostic tools, not automatic rules for accepting or rejecting a territory factor. The following tutorial develops full-credibility standards; here, the focus is on understanding why the evidence behind territorial indications can differ in reliability.

Key takeaway: Assess territory experience by examining exposure volume, claim frequency, and loss severity variation together. A raw indication describes what was observed; an uncertainty assessment helps show how strongly that observation should inform a view of underlying risk.

Check Your Understanding

Use the assumptions stated in each problem. Round standard errors to the nearest cent where dollar amounts are requested.

A hypothetical territory has 80 claims in 4,000 exposure units. Under an independent Poisson model, calculate its frequency indication and estimated standard error. (Answer: 0.0200 claims per exposure; standard error 0.0022.)
Territory A has 40 claims in 2,000 exposures, and Territory B has 90 claims in 6,000 exposures. Assume independent Poisson counts. Calculate both frequencies and the estimated standard error of their difference. (Answer: A = 0.0200, B = 0.0150 claims per exposure; standard error of the difference = 0.0035.)
A portfolio has frequency 0.04 per exposure, mean claim amount $2,500, severity standard deviation $5,000, and 1,000 exposures. Under the compound Poisson model, calculate expected loss cost and its standard error. (Answer: expected loss cost = $100 per exposure; standard error = $35.36 per exposure.)
Two hypothetical territories each have frequency 0.05 per exposure. One has 25 claims in 500 exposures; the other has 100 claims in 2,000 exposures. Calculate their frequency standard errors and explain which estimate is more stable under the model. (Answer: 0.0100 and 0.0050 claims per exposure, respectively; the second is more stable.)
A territory’s claims arise largely from a single storm that affects many insureds at once. Explain why a standard error based on independent Poisson claims may understate the variability of its loss-cost indication. (Answer: Claims are dependent because of the shared event, so the independent-claims model omits common-event variation and can understate uncertainty.)

What You'll Learn