Frequency Data And Credibility

Why Frequency Data Need an Exposure Denominator

A claim count by itself does not describe how often claims occur unless the amount of insurance exposure is also known. Twenty claims could be a low frequency for a large portfolio and a high frequency for a small one. For ratemaking, frequency experience is therefore measured as claims per unit of exposure, with the claim definition and exposure basis held consistent.

This tutorial focuses on claim counts and exposure volume as inputs to judging the stability of observed frequency. It does not address claim severity or aggregate loss credibility. The scenarios below are hypothetical, and the calculations use no interest or mortality assumptions.

Definition: Let $N$ be the number of claims observed for a defined group and period, and let $E$ be the exposure for that group and period, expressed in the chosen exposure units. The observed claim frequency is $$ \hat{\lambda}=\frac{N}{E}. $$ It measures claims per exposure unit. If the result is reported per 100 exposure units, multiply $\hat{\lambda}$ by 100. The claim-count definition, coverage, period, and exposure unit must be specified and consistent for comparisons to be meaningful.

An exposure unit might be one earned vehicle-year, one insured location-year, or another unit appropriate to the coverage. Partial periods contribute partial exposure when the exposure measure supports that treatment: for example, six months of covered time contributes 0.5 vehicle-year. The numerator and denominator must refer to the same group, coverage, and period. A count of reported claims divided by earned exposure is not automatically comparable with a count of paid claims divided by written exposure.

What Claim Frequency Measures

Frequency describes the rate at which claims occur relative to exposure. It is distinct from the average size of claims. A group can have more claims per exposure unit but smaller average claims, or fewer claims but larger average claims. Frequency alone does not determine the expected cost of losses; it is one component of understanding claim experience.

Before calculating a frequency, decide what counts as one claim. For example, a claim count may be based on reported claims, closed claims, or another defined measure. The choice must be applied consistently. If one event can produce multiple claims, the count may also depend on whether the chosen unit is an individual claim or an occurrence. Those are different measures and should not be combined without a deliberate, consistent definition.

Worked Example: Converting Partial Coverage into Frequency

A hypothetical insurer observes 21 claims from a group of seasonal equipment policies. The group had 240 policies in force for a full year and another 120 policies in force for half a year. Treat the exposure as earned policy-years, and assume the 21 claims and all exposure refer to the same coverage and period. What is the observed frequency per policy-year and per 100 policy-years?

Method. Add the earned exposure, converting the half-year policies to policy-years. The stipulated model for assessing count variation is $N\sim\operatorname{Poisson}(\lambda E)$, with observed $N=21$ and total $E$ calculated below. No interest or mortality model is involved.

The first group contributes $240\times 1=240$ policy-years. The second contributes $120\times 0.5=60$ policy-years. Therefore, $$ E=240+60=300\text{ policy-years}. $$

The observed frequency per policy-year is $$ \hat{\lambda}=\frac{N}{E}=\frac{21}{300}=0.07. $$

Per 100 policy-years, this is $0.07\times 100=7.00$ claims. Thus the group had 0.07 claims per policy-year, or 7.00 claims per 100 policy-years. As a check, $300\times 0.07=21$, which reproduces the observed count.

Common mistake avoided. Dividing 21 by 360 policies would treat every policy as a full-year exposure. That ignores the shorter coverage period for 120 policies and understates the frequency per policy-year.

Frequency Variation Under a Poisson Model

A simple model for random claim counts assumes that, conditional on a fixed underlying frequency $\lambda$ and exposure $E$, the count has a Poisson distribution with mean $\lambda E$. In this model, claim occurrences are independent, exposure units are comparable, and the rate is constant across the group and observation period. These assumptions are useful for deriving a benchmark for random count variation; they are not guaranteed to hold in practice.

Poisson frequency model: If $N\sim\operatorname{Poisson}(\lambda E)$, then $E[N]=\lambda E$ and $\operatorname{Var}(N)=\lambda E$. For fixed exposure $E$, the frequency estimate $\hat{\lambda}=N/E$ has $$ E[\hat{\lambda}]=\lambda,\qquad \operatorname{Var}(\hat{\lambda})=\frac{\lambda}{E},\qquad \operatorname{SE}(\hat{\lambda})=\sqrt{\frac{\lambda}{E}}. $$ Because $\lambda$ is usually unknown, an estimated standard error under this model is $\sqrt{N}/E$, obtained by substituting $\hat{\lambda}=N/E$.

The variance result follows directly from the scaling rule for variance. Since $E$ is fixed, $\hat{\lambda}=N/E$, so $$ \operatorname{Var}(\hat{\lambda}) =\operatorname{Var}\left(\frac{N}{E}\right) =\frac{\operatorname{Var}(N)}{E^2} =\frac{\lambda E}{E^2} =\frac{\lambda}{E}. $$ The square root gives the standard error. This explains why, at a fixed underlying frequency, increasing exposure reduces the random variation in the estimated frequency. The relative standard error under the model is $1/\sqrt{\lambda E}$, because the expected count is $\lambda E$. In practice, substituting the observed count gives the approximate relative standard error $1/\sqrt{N}$, when $N>0$.

The count matters because more claims generally provide more information about a frequency than fewer claims. Exposure matters because it provides the denominator and, under the model, determines the expected count at a given frequency. Two groups with equal exposure can have different estimated precision if their frequencies differ; two groups with equal observed counts can have different estimated frequencies if their exposures differ. Neither the raw count nor the raw frequency alone tells the whole story.

Worked Example: Comparing Two Frequency Estimates

Two hypothetical segments have consistent claim definitions and earned exposure measured in policy-years. Segment A has 64 claims in 1,600 policy-years; Segment B has 18 claims in 300 policy-years. Assume counts follow the Poisson model with mean $\lambda E$ within each segment, using each segment’s own underlying frequency. Compare their observed frequencies and estimated standard errors. No interest or mortality assumptions are involved.

Segment A. Its observed frequency is $$ \hat{\lambda}_A=\frac{64}{1600}=0.04 $$ claims per policy-year, or 4.00 per 100 policy-years. The estimated standard error is $$ \widehat{\operatorname{SE}}(\hat{\lambda}_A) =\frac{\sqrt{64}}{1600} =\frac{8}{1600} =0.005 $$ claims per policy-year, or 0.50 per 100 policy-years. Its estimated relative standard error is $1/\sqrt{64}=0.125$, or 12.50%.

Segment B. Its observed frequency is $$ \hat{\lambda}_B=\frac{18}{300}=0.06 $$ claims per policy-year, or 6.00 per 100 policy-years. Its estimated standard error is $$ \widehat{\operatorname{SE}}(\hat{\lambda}_B) =\frac{\sqrt{18}}{300} \approx 0.01414 $$ claims per policy-year, or approximately 1.41 per 100 policy-years. Its estimated relative standard error is $1/\sqrt{18}\approx 0.2357$, or 23.57%.

Interpretation. Segment B’s observed frequency is higher: 6.00 rather than 4.00 claims per 100 policy-years. Under the stated model, however, its estimate is less stable: its estimated standard error is larger, and its estimated relative standard error is nearly twice Segment A’s. These are measures of sampling variation, not proof that the segments have the same underlying frequency or that their difference is random.

Check. Multiplying each estimated frequency by its exposure returns the observed count: $0.04\times 1600=64$ and $0.06\times 300=18$. A common mistake is to conclude that Segment B’s estimate is more reliable because its frequency is higher. The count and exposure, not the size of the frequency by itself, inform the model-based precision.

From a Precision Goal to an Exposure Standard

A frequency credibility standard can be expressed as a requirement for the observed count to be close to its expected value with a specified probability. One common approximate approach uses the normal approximation to a Poisson count when the expected count is sufficiently large. It sets a target relative error $k$ and a central probability $p$, then finds the expected count needed to meet that target approximately.

Approximate full-credibility count standard: Under the Poisson model, let the expected count be $m=\lambda E$. For a two-sided central probability $p$, let $z_{(1+p)/2}$ be the corresponding standard normal quantile. Using the normal approximation, a target relative error $k$ requires approximately $$ m\geq \left(\frac{z_{(1+p)/2}}{k}\right)^2. $$ This is an approximation, not an exact Poisson probability statement. It assumes a stable frequency and comparable independent exposures; it may be inadequate for small expected counts or overdispersed, heterogeneous data.

The logic is that a Poisson count with mean $m$ has variance $m$, and its approximate standard deviation is $\sqrt{m}$. Relative to its mean, the approximate standard deviation is $1/\sqrt{m}$. Requiring a $z$-multiple of that relative standard deviation to be no greater than $k$ gives $z/\sqrt{m}\leq k$. Squaring and rearranging yields $m\geq(z/k)^2$. If a benchmark frequency is available, expected count can then be translated into exposure by dividing by that frequency.

Worked Example: Converting a Precision Goal to Exposure

A hypothetical rating segment has benchmark frequency $\lambda=0.08$ claims per exposure unit. The analyst wants an approximate 90% central probability that the observed frequency is within 5% of the benchmark, using the Poisson normal-approximation standard. What expected claim count and exposure are required? Assume $N\sim\operatorname{Poisson}(\lambda E)$. No interest or mortality assumptions are involved.

Step 1: Select the normal quantile. For a two-sided central probability of 90%, $z_{(1+0.90)/2}=z_{0.95}\approx 1.645$.

Step 2: Calculate the expected-count target. Here $k=0.05$, so $$ m\geq\left(\frac{1.645}{0.05}\right)^2 =(32.9)^2 =1082.41. $$ Because the expected count must meet or exceed this threshold, the count target is at least 1,082.41 expected claims. If the operational standard is stated in whole claims, round the target up to 1,083.

Step 3: Convert expected claims to exposure. Using the stated benchmark frequency, $$ E\geq\frac{1082.41}{0.08}=13{,}530.125 $$ exposure units. If the standard uses the rounded whole-claim target of 1,083, the requirement is $$ E\geq\frac{1083}{0.08}=13{,}537.5, $$ or 13,538 whole exposure units when exposure must be an integer. At 13,538 units, expected claims are $0.08\times 13{,}538=1{,}083.04$, which exceeds the rounded target.

Check. The unrounded formula gives 1,082.41 expected claims; the whole-claim convention raises that to 1,083, and the exposure is rounded upward rather than downward. The normal approximation does not guarantee that an exact Poisson probability meets the target. A common mistake is to use $z/k$ as the count requirement without squaring; the variance-to-mean relationship leads to $(z/k)^2$.

Conditions, Interpretation, and Common Traps

The Poisson calculation is a benchmark, not a substitute for examining how the data were generated. If risks differ substantially in their underlying frequencies, a single constant $\lambda$ may not describe the segment. Dependence among claims, concentration in time, inconsistent reporting, or changes in coverage can also produce variation that the simple model does not capture. In such cases, the Poisson standard error may understate uncertainty.

A credible frequency comparison also requires an appropriate exposure base. For example, policy-years are useful only if they represent reasonably comparable opportunities for claims within the coverage being studied. If exposures differ in duration or definition, first make the units consistent or use an exposure measure suited to the comparison. A rate can be arithmetically correct and still be misleading if its numerator and denominator do not correspond.

Do not compare counts without exposure. A larger portfolio can have more claims simply because it insures more exposure.
Do not confuse frequency with severity. Frequency is claims per exposure unit; it does not describe the dollar size of claims.
Do not treat a higher frequency as more stable. Under the Poisson benchmark, the count and expected count drive relative precision.
Do not use inconsistent claim units. Changes in how claims, reports, or occurrences are counted can create apparent frequency differences.
Do not treat an approximate count standard as exact. The normal approximation and its assumptions should be stated when applying it.

Key takeaway: Frequency experience is a claim count divided by corresponding exposure. A larger count generally reduces random relative variation under a homogeneous Poisson model, but a sound credibility judgment also depends on exposure quality, consistent definitions, and whether the model’s assumptions are reasonable.

Check Your Understanding

For each problem, use the stated claim-count definition and exposure basis. Unless a problem says otherwise, assume a homogeneous Poisson model for any requested model-based precision calculation.

A hypothetical book has 30 claims over 750 earned exposure units. Calculate the observed frequency per exposure unit and per 100 exposure units. (Answer: 0.04 per exposure unit; 4.00 per 100 exposure units.)
Two hypothetical groups have 14 claims in 100 exposure units and 26 claims in 400 exposure units. Calculate the combined frequency using total claims divided by total exposure. (Answer: 40/500 = 0.08 claims per exposure unit.)
Group A has 9 claims in 150 exposure units, while Group B has 36 claims in 1,200 exposure units. Find each observed frequency and its estimated Poisson standard error, then identify which estimate has the smaller standard error. (Answer: A: 0.06, standard error 0.02; B: 0.03, standard error 0.005; Group B has the smaller standard error.)
A segment targets a 95% central probability and a 10% relative error under the normal-approximation full-credibility standard. Use $z_{0.975}=1.96$. If its benchmark frequency is 0.05 claims per exposure unit, find the rounded-up expected-claim target and the corresponding exposure target. (Answer: $(1.96/0.10)^2=384.16$, rounded up to 385 expected claims; $385/0.05=7{,}700$ exposure units.)

What You'll Learn