Why Frequency and Loss Credibility Can Differ
A class can produce a remarkably steady number of claims while its total losses move sharply from year to year. The reason is that frequency and claim size are different sources of variation. Frequent, predictable counts may support a relatively strong view of the class’s claim frequency, while a small number of highly variable claim amounts may make its average severity or aggregate loss much less stable.
That distinction matters when evaluating class experience for ratemaking. A frequency estimate describes claims per exposure unit; an aggregate-loss estimate combines how often claims occur with how much they cost. Treating the two estimates as equally informative simply because they come from the same class can overstate what the data support.
These judgments are not a substitute for checking whether the experience is complete, mature, and suitable for the intended use. The earlier material on completeness and maturity of experience and data quality checks for credibility addresses those questions. Here, assume that the experience has passed the necessary checks and focus on how count and size variation can lead to different credibility conclusions.
A Model That Separates Counts From Claim Sizes
A useful way to make the distinction precise is a compound Poisson model. Suppose \(E\) is exposure, \(N\) is the claim count, and claims occur independently at a constant rate \(\lambda\) per exposure unit, so that \(N\sim\operatorname{Poisson}(\lambda E)\). Let \(X_1,\ldots,X_N\) be claim severities that are identically distributed, mutually independent, and independent of \(N\), with finite mean \(\mu_X=E[X]\) and variance \(\sigma_X^2=\operatorname{Var}(X)\). Define aggregate loss \(S=\sum_{j=1}^{N}X_j\), with \(S=0\) if \(N=0\).
Under these assumptions, the expected aggregate loss and its variance are:
The variance has two components: variation in the claim sizes and variation in the number of claims. Since a Poisson count has variance equal to its mean, this can also be written:
This result depends on the stated model. If claims are dependent, the count is not Poisson, the claim-size distribution changes across risks or time, or counts and severities are related, the formula may not apply as written. In those situations, the sources of variation need an appropriate model rather than an automatic use of this expression.
Why the Aggregate-Loss Variance Has Two Parts
The decomposition follows from conditioning on the claim count. Given \(N=n\), aggregate loss is a sum of \(n\) independent severities. Its conditional mean is \(n\mu_X\), and its conditional variance is \(n\sigma_X^2\). The law of total variance states that the unconditional variance equals the expected conditional variance plus the variance of the conditional expectation:
For a Poisson count, \(E[N]=\operatorname{Var}(N)=\lambda E\). Substitution gives \(\operatorname{Var}(S)=\lambda E(\sigma_X^2+\mu_X^2)\). This derivation shows why steady counts alone do not settle aggregate-loss stability: the term \(E[N]\sigma_X^2\) remains even if count variation is relatively small. A volatile severity distribution can still make total losses uncertain.
Worked Example: Stable Counts, a Volatile Severity Sample
Worked Example: Stable Counts, a Volatile Severity Sample
A hypothetical class has 1,000 exposure units in each of four successive years. Its reported claim counts are 298, 304, 301, and 297. Separately, a small review sample of eight claims from the class has seven losses of $1,000 and one loss of $15,000. Compare the observed count pattern with the severity sample’s variability. Use the sample standard deviation for the severity calculation. Do not treat the eight-claim review sample as the complete experience for all 1,200 claims.
Method. Calculate each annual frequency and the overall frequency, then calculate the sample mean and sample coefficient of variation for the eight severities. The coefficient of variation is the sample standard deviation divided by the sample mean.
Frequency. The four annual frequencies are \(298/1{,}000=0.298\), \(304/1{,}000=0.304\), \(301/1{,}000=0.301\), and \(297/1{,}000=0.297\) claims per exposure unit. Total claims are \(298+304+301+297=1{,}200\), and total exposure is \(4(1{,}000)=4{,}000\). Thus the overall frequency is \(1{,}200/4{,}000=0.300\) claims per exposure unit.
Severity. The sample’s total loss is \(7(\$1{,}000)+\$15{,}000=\$22{,}000\), so its mean severity is \(\$22{,}000/8=\$2{,}750\). Seven observations are \(-\$1{,}750\) from the mean, and one is \(\$12{,}250\) above it. The sum of squared deviations is \(7(\$1{,}750)^2+(\$12{,}250)^2=\$171{,}500{,}000\). The sample variance is \(\$171{,}500{,}000/(8-1)=\$24{,}500{,}000\), giving a sample standard deviation of approximately \(\$4{,}949.75\). The sample coefficient of variation is \(\$4{,}949.75/\$2{,}750\approx1.800\).
The counts are clustered between 297 and 304 per 1,000 exposure units, while the small severity sample has a standard deviation about 1.8 times its mean. That contrast supports examining frequency and severity evidence separately. It does not, by itself, establish a numerical credibility weight: the severity sample is small, and its representativeness and the relevance of the large loss still need consideration.
Check. The annual claim counts sum to 1,200, the sample losses sum to $22,000, and the sample variance uses 7 degrees of freedom. A common mistake is to interpret the stable counts as proof that the class’s losses are equally stable; the severity calculation shows why that conclusion does not follow.
Worked Example: Equal Counts, Different Aggregate Losses
Worked Example: Equal Counts, Different Aggregate Losses
In a hypothetical class, two years each have 100 exposure units and 20 claims. In the first year, every claim is $1,000. In the second year, 19 claims are $1,000 and one claim is $20,000. Compare frequency, average severity, and aggregate loss. Assume the amounts are already on the same valuation and coverage basis.
Method. Calculate claims per exposure, total loss, and loss per claim for each year. This separates the count result from the claim-size result.
Both years have frequency \(20/100=0.20\) claims per exposure unit. In year one, aggregate loss is \(20(\$1{,}000)=\$20{,}000\), and average severity is \(\$20{,}000/20=\$1{,}000\). In year two, aggregate loss is \(19(\$1{,}000)+\$20{,}000=\$39{,}000\), and average severity is \(\$39{,}000/20=\$1{,}950\).
The frequency is identical, but year-two aggregate loss is \(\$39{,}000/\$20{,}000=1.95\) times year one’s, or 95% higher. The larger loss is relevant to the observed aggregate. Whether it should receive special treatment depends on the intended loss basis and the applicable approach to large claims; it should not be silently removed just because it changes the result.
Check. Each year has 20 claims over 100 exposure units, and the second-year total includes the $20,000 claim once. A common mistake is to infer similar loss experience from equal claim counts, overlooking how one claim can materially affect the total.
Worked Example: Separating Count and Severity Variation
Worked Example: Separating Count and Severity Variation
For a hypothetical class, suppose claim counts follow a Poisson model with rate \(\lambda=0.20\) claims per exposure unit. The class has \(E=500\) exposure units. Assume each claim severity is independently $1,000 or $3,000 with equal probability, and severities are independent of the count. Calculate the expected count, frequency estimate’s standard deviation, expected aggregate loss, and aggregate-loss standard deviation under the model.
Method. Use the Poisson count assumptions and the compound Poisson variance decomposition. All dollar amounts below refer to the stated hypothetical loss basis.
The expected count is \(\lambda E=0.20(500)=100\). The Poisson count variance is also 100, so its standard deviation is \(\sqrt{100}=10\). The count-per-exposure estimator has variance \(\lambda/E=0.20/500=0.0004\), and hence standard deviation \(\sqrt{0.0004}=0.0200\) claims per exposure unit.
For severity, the mean is \(\mu_X=0.5(\$1{,}000)+0.5(\$3{,}000)=\$2{,}000\). The variance is \(0.5(\$1{,}000-\$2{,}000)^2+0.5(\$3{,}000-\$2{,}000)^2=\$1{,}000{,}000\). Thus expected aggregate loss is \(100(\$2{,}000)=\$200{,}000\). Its variance is:
The aggregate-loss standard deviation is \(\sqrt{\$500{,}000{,}000}\approx\$22{,}360.68\). The first variance component is due to severity variation; the second is due to count variation. Here the severity component contributes $100,000,000 of variance and the count component $400,000,000.
Check. The expected count and count variance are both 100 under the Poisson assumption, and the two aggregate variance components sum to $500,000,000. A common mistake is to use only the variance of claim sizes and omit the count component, or to use the variance of the count without multiplying by the squared mean severity.
Making Separate Credibility Judgments
The examples point to a practical principle: evaluate the statistic that will actually inform the rate. Frequency experience may be plentiful and steady, yet a severity estimate may rely on relatively few claims or be highly affected by a large loss. Conversely, counts may be sparse or erratic even when claim sizes are comparatively consistent.
Where frequency and severity are modeled separately, distinct credibility weights can represent distinct assessments. For illustration, a frequency estimate might be blended with a relevant complement using a frequency weight \(Z_f\), while a severity estimate is blended using a severity weight \(Z_s\):
These expressions describe separate blends; they do not prescribe how either weight must be calculated. The weights should reflect the credibility framework being used and the suitability of each complement. Also, the product \(\widehat{f}\widehat{s}\) is not automatically the same as directly blending class and complement pure premiums. That depends on the model and blending method. Keep the rate component, experience basis, and complement aligned.
Frequency and aggregate-loss credibility are related but not interchangeable. Aggregate loss contains both count and severity variation, as the compound Poisson result demonstrates. A class can therefore have persuasive frequency experience and still have unstable aggregate losses. The appropriate response is to explain the difference and assess the relevant evidence, not to force one credibility judgment to stand in for another.
Common Traps
- Using count stability as a proxy for loss stability. Equal or steady counts say nothing conclusive about the variation in claim sizes.
- Giving frequency and severity the same weight by default. They are different statistics and may have different sample sizes, variability, or complements.
- Ignoring a large loss without a defined basis. First determine the loss measure the analysis is intended to use; then handle large claims consistently with that basis.
- Confusing a model illustration with a credibility standard. A variance decomposition explains sources of uncertainty. It does not by itself determine a credibility weight or establish that model assumptions are appropriate.
Check Your Understanding
For each problem, distinguish the count evidence from the loss evidence. Dollar amounts are hypothetical.
- A class has 800 exposure units in each of three years, with counts of 156, 164, and 160. Calculate total frequency. If a claim-size sample is small and dominated by one large claim, what conclusion can be drawn about frequency versus severity evidence? (Answer: Total frequency is \(480/2{,}400=0.200\) claims per exposure unit; the counts are fairly consistent, but this alone does not establish severity credibility.)
- Two years each have 100 exposure units and 12 claims. In year one each claim is $800. In year two, 11 claims are $800 and one is $10,000. Calculate each year’s frequency and aggregate loss, then state how the totals compare. (Answer: Both frequencies are 0.12; aggregate losses are $9,600 and $18,800, so the second is about 1.958 times the first.)
- Assume a compound Poisson model with \(\lambda=0.10\) claims per exposure unit and 1,000 exposure units. Claim severity is $500 or $2,500 with equal probability, independently of count. Calculate expected aggregate loss and aggregate-loss variance. (Answer: \(E[X]=\$1{,}500\), \(\operatorname{Var}(X)=\$1{,}000{,}000\), expected count \(=100\), expected aggregate loss \(=\$150{,}000\), and variance \(=100(\$1{,}000{,}000+\$1{,}500^2)=\$325{,}000{,}000\).)
- A class frequency is 0.16 and its complement frequency is 0.12, with a frequency weight of 0.80. Class severity is $2,400 and complement severity is $1,800, with a severity weight of 0.25. Calculate the two blended estimates and their product, assuming the analyst deliberately uses separate blends. (Answer: Blended frequency \(=0.80(0.16)+0.20(0.12)=0.152\); blended severity \(=0.25(\$2{,}400)+0.75(\$1{,}800)=\$1{,}950\); product \(=0.152(\$1{,}950)=\$296.40\) per exposure unit.)
- A class’s observed frequency is stable over several periods, but aggregate loss varies widely because each period has few claims and occasional large claims. Name one reason separate frequency and aggregate-loss judgments are appropriate, and one check needed before relying on the class experience. (Answer: Count variation and claim-size variation affect frequency and aggregate loss differently; before relying on the data, check that the experience is complete and mature enough for the intended evaluation.)