Frequency Credibility with Unequal Risk Exposure

Why Unequal Risk Exposure Changes the Calculation

Frequency credibility measures how much claim-count experience supports an estimate of claim frequency. When every exposure unit has the same expected frequency, the volume measure is straightforward: it is the expected number of claims in the group. When units have different expected frequencies, the exposure count alone is not enough. One policy-year with an expected frequency of 0.20 contributes more expected claims than one with an expected frequency of 0.02.

The appropriate volume measure is therefore the sum of expected claim counts across the exposure groups. This is the quantity used to assess whether the experience meets a full-credibility standard and, if not, to calculate partial credibility. It is not the unweighted average of the group frequencies, nor is it automatically the number of policies.

Definition: For exposure group \(i\), let \(\lambda_i\) denote its expected claim count during the observation period. For independent Poisson claim counts, the aggregate expected claim count is \(\Lambda=\sum_i\lambda_i\). Frequency credibility is determined from \(\Lambda\), using the applicable full-credibility standard. Unequal frequencies are allowed; the expected counts must be calculated on a consistent basis and then added.

A group’s expected count may be specified directly. If instead the group has \(e_i\) exposure units and expected frequency \(f_i\) claims per exposure unit, then \(\lambda_i=e_i f_i\). For example, 300 policy-years at 0.08 expected claims per policy-year contribute \(300(0.08)=24\) expected claims. A half-year policy exposure should be represented as half a policy-year if the frequency is stated per policy-year.

The Full-Credibility Standard for Claim Counts

A common limited-fluctuation standard asks for the probability that the observed claim count is within a specified relative error of its expected value. Write \(N\) for the aggregate observed claim count and \(\Lambda\) for the aggregate expected claim count. Let \(p\) be the target confidence level and \(k\) the maximum relative error. The criterion is \(P(|N-\Lambda|\le k\Lambda)\ge p\). The usual claim-count standard uses a Poisson model and a normal approximation when the expected count is sufficiently large.

Full-credibility count standard: Under the Poisson model \(N\sim\operatorname{Poisson}(\Lambda)\), the normal approximation gives the required expected claim count \(\Lambda_0=(z_{(1+p)/2}/k)^2\), where \(z_{(1+p)/2}\) is the standard normal quantile with cumulative probability \((1+p)/2\). The experience is fully credible when \(\Lambda\ge\Lambda_0\). This is an approximation-based standard, not an exact Poisson probability calculation.

For instance, a two-sided confidence level of 90% uses \(z_{0.95}\approx1.645\). If the tolerated relative error is 5%, then the full-credibility count is \((1.645/0.05)^2=1082.41\) expected claims. The fractional result is a threshold from the approximation; it does not mean that a fractional claim must be observed.

Calculate each group’s expected count.
Use the supplied expected counts, or multiply each group’s exposure by its expected frequency on the same exposure basis.

Aggregate the expected counts.
Set \(\Lambda=\sum_i\lambda_i\). Do not substitute the number of policies or average group frequencies without appropriate exposure weights.

Compare with the full standard.
Calculate \(\Lambda_0\) from the specified confidence level and error tolerance. Full credibility applies if \(\Lambda\ge\Lambda_0\).

If needed, calculate partial credibility.
When \(\Lambda<\Lambda_0\), use the square-root rule \(Z=\sqrt{\Lambda/\Lambda_0}\). Cap the result at 1.

Why Aggregation and the Square-Root Rule Work

Suppose the claim count \(N_i\) in each group is independent and has a Poisson distribution with mean \(\lambda_i\). The sum \(N=\sum_i N_i\) is then Poisson with mean \(\Lambda=\sum_i\lambda_i\). In particular, \(E[N]=\Lambda\) and \(\operatorname{Var}(N)=\Lambda\). Thus, the variability of the aggregate count is determined by the total expected count, not by how many groups were used to form it.

For a sufficiently large \(\Lambda\), the normal approximation gives a standard deviation of approximately \(\sqrt{\Lambda}\) for \(N\). Relative to the expected count, the standard deviation is approximately \(\sqrt{\Lambda}/\Lambda=1/\sqrt{\Lambda}\). Therefore the approximate probability of staying within relative error \(k\) is

P\left(\left|\frac{N-\Lambda}{\Lambda}\right|\le k\right) \approx 2\Phi(k\sqrt{\Lambda})-1.

To achieve confidence \(p\), set \(k\sqrt{\Lambda_0}=z_{(1+p)/2}\) and solve for \(\Lambda_0\). This gives \(\Lambda_0=(z_{(1+p)/2}/k)^2\). The same reasoning explains the partial-credibility factor: relative sampling error varies approximately as \(1/\sqrt{\Lambda}\), so credibility relative to the full standard scales as \(\sqrt{\Lambda/\Lambda_0}\).

Partial frequency credibility: For a full-credibility expected count \(\Lambda_0\), the limited-fluctuation credibility factor is \(Z=\min\{1,\sqrt{\Lambda/\Lambda_0}\}\). If \(O\) is the observed aggregate count, the observed-to-expected frequency ratio is \(R=O/\Lambda\). When the baseline expected frequency is represented by an index of 1, the partial-credibility indication is \(1+Z(R-1)\). This adjustment applies the credibility factor to the experience’s deviation from the baseline.

This square-root rule relies on the Poisson variance assumption and the stated limited-fluctuation standard. It is not automatically appropriate when counts are materially overdispersed, dependent, or subject to unmodeled variation. In those situations, the variance assumption underlying the standard needs to be reconsidered rather than silently using the same threshold.

Worked Example: Combining Expected Counts to Test Full Credibility

A hypothetical insurer evaluates three independent segments under a 90% confidence standard with a 5% relative error tolerance. The expected claim counts are 240, 360, and 500. Observed claims are 258, 345, and 527, respectively. Determine whether the experience is fully credible and calculate its observed-to-expected frequency ratio. If it is fully credible, use \(Z=1\).

Method. Aggregate expected claim counts, compare their total with the Poisson limited-fluctuation standard, and then calculate the aggregate ratio. This avoids treating the three segments as three equal exposure units.

Step 1: Find the full-credibility count. For 90% confidence, \(z_{0.95}=1.645\), and \(k=0.05\):

\Lambda_0=\left(\frac{1.645}{0.05}\right)^2 =(32.9)^2=1082.41.

Step 2: Aggregate expected and observed claims.

\Lambda=240+360+500=1100, \qquad O=258+345+527=1130.

Step 3: Determine credibility. Since \(1100\ge1082.41\), the experience meets the full standard, so \(Z=1\).

Step 4: Calculate the frequency ratio.

R=\frac{O}{\Lambda} =\frac{1130}{1100} =1.0272727\ldots\approx1.0273.

The experience is fully credible under the specified approximate standard, and its aggregate observed frequency is about 2.73% above expected. As a check, \(1+Z(R-1)=1+1(1.0273-1)=1.0273\), so full credibility leaves the experience ratio unchanged. The common mistake of averaging the three segment counts would not measure total claim-count volume.

Worked Example: Partial Credibility With Unequal Expected Counts

A hypothetical program has three independent groups with expected counts of 90, 150, and 240. It records 510 claims in total. Using 90% confidence and a 5% relative error tolerance, calculate the partial credibility factor and the partial-credibility frequency indication.

Method. Use the sum of the group expected counts as the frequency volume, then apply the square-root rule to the observed-to-expected ratio.

Step 1: Aggregate expected claims and find the standard.

\Lambda=90+150+240=480, \qquad \Lambda_0=\left(\frac{1.645}{0.05}\right)^2=1082.41.

Since \(480<1082.41\), the experience is not fully credible.

Step 2: Calculate partial credibility.

Z=\sqrt{\frac{480}{1082.41}} =\sqrt{0.4434549\ldots} =0.6659.

Step 3: Calculate the observed-to-expected ratio.

R=\frac{510}{480}=1.0625.

Step 4: Apply partial credibility to the deviation from 1.

1+Z(R-1) =1+0.665924(1.0625-1) =1+0.665924(0.0625) =1.041620\ldots\approx1.0416.

The credibility factor is 0.6659, and the adjusted indication is 1.0416, or about 4.16% above the baseline. The unadjusted experience ratio is 1.0625; the indication moves only 66.59% of the way from 1 toward that ratio. Rounding is to four decimals at the final step.

Worked Example: Converting Unequal Frequencies Into Expected Claims

A hypothetical portfolio contains 400 policy-years in a group with expected frequency 0.12 claims per policy-year, 800 policy-years at 0.06, and 1,200 policy-years at 0.03. The groups report 57, 44, and 39 claims, respectively. Under a 95% confidence standard with a 10% relative error tolerance, calculate total expected claims, partial credibility, and the partial-credibility frequency indication.

Method. Convert each group’s exposure and frequency into an expected count before aggregating. This prevents a high-frequency policy-year and a low-frequency policy-year from being treated as equivalent expected claim volume.

Step 1: Calculate expected claims by group.

\lambda_1=400(0.12)=48,\qquad \lambda_2=800(0.06)=48,\qquad \lambda_3=1200(0.03)=36.

Thus, \(\Lambda=48+48+36=132\). The total exposure is 2,400 policy-years, but that is not the expected claim count.

Step 2: Find the full-credibility standard and partial factor. For 95% confidence, \(z_{0.975}=1.96\), and \(k=0.10\):

\Lambda_0=\left(\frac{1.96}{0.10}\right)^2=384.16, \qquad Z=\sqrt{\frac{132}{384.16}}=0.5862.

Since \(132<384.16\), the portfolio has partial rather than full credibility.

Step 3: Calculate the aggregate observed frequency ratio.

O=57+44+39=140, \qquad R=\frac{140}{132}=1.060606\ldots\approx1.0606.

Step 4: Adjust the deviation using \(Z\).

1+0.5862(1.060606-1) =1+0.5862(0.060606) \approx1.0355.

The partial-credibility indication is 1.0355, about 3.55% above the baseline. The trap is to compare 2,400 policy-years with the threshold of 384.16 and declare full credibility. That threshold is expressed in expected claims; the correct volume is 132.

Common Mistakes and Model Checks

Using the number of exposure units as the claim-count volume. Policy-years, vehicle-years, and similar units must first be translated into expected claims using their applicable frequencies.
Taking an unweighted average of frequencies. If a combined rate is needed, the groups’ exposures matter. For credibility volume, calculate each expected count and sum them.
Comparing unlike quantities. The full standard \(\Lambda_0\) is an expected-claim threshold. Compare it with \(\Lambda\), not with a raw policy count or an observed-to-expected ratio.
Forgetting the cap. If \(\Lambda\) exceeds \(\Lambda_0\), the formula \(\sqrt{\Lambda/\Lambda_0}\) exceeds 1; the credibility factor is capped at 1.
Applying the Poisson standard without checking its assumptions. Independent Poisson group counts with known expected counts aggregate to a Poisson count. Dependence or extra-Poisson variation can invalidate the variance calculation used by the standard.

Unequal expected frequencies do not, by themselves, prevent aggregation. Under independent Poisson counts, the expected counts add and the aggregate count retains the Poisson variance-to-mean relationship. The practical task is to construct each \(\lambda_i\) correctly, on consistent exposure and time bases, before applying the standard. This method is aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus.

Check Your Understanding

For each problem, aggregate expected claims before deciding credibility. Unless a problem says otherwise, use the Poisson limited-fluctuation approximation and round credibility factors to four decimals.

A hypothetical book has group expected counts of 75, 125, and 200. Under 90% confidence and a 10% relative error tolerance, determine whether it is fully credible and give \(Z\). (Answer: \(\Lambda=400\), \(\Lambda_0=270.6025\), fully credible, \(Z=1\).)
Three groups have expected counts of 40, 80, and 100. Find \(Z\) under 95% confidence and a 10% relative error tolerance. (Answer: \(\Lambda=220\), \(\Lambda_0=384.16\), \(Z=0.7568\).)
Expected claims in three segments are 50, 75, and 125; observed claims are 54, 70, and 132. Under 90% confidence and a 5% relative error tolerance, calculate \(Z\) and the adjusted frequency indication. (Answer: \(\Lambda=250\), \(O=256\), \(R=1.0240\), \(Z=0.4806\), adjusted indication \(=1.0115\).)
A portfolio has 300 policy-years at frequency 0.04, 500 at 0.10, and 1,000 at 0.02. Under 95% confidence and a 15% relative error tolerance, calculate expected claims and \(Z\). (Answer: expected counts \(12+50+20=82\), \(\Lambda_0=(1.96/0.15)^2=170.7378\), \(Z=0.6930\).)
A hypothetical portfolio consists of two groups with expected claim counts of 700 and 500. Under 90% confidence and a 5% relative error tolerance, determine the credibility status and explain which volume measure controls the decision. (Answer: \(\Lambda=1200\), \(\Lambda_0=1082.41\), fully credible with \(Z=1\); the expected claim count, not the number of exposure units, controls.)

What You'll Learn