Worked Claim-Count Credibility Calculations

From a Credibility Standard to a Claim Threshold

A claim-count credibility calculation translates a desired degree of stability into a required volume of claims. The main inputs are the confidence level, the allowable relative deviation from the expected count, and—when exposure is requested—the expected claim frequency per exposure unit. Keeping the claim threshold separate from the exposure calculation prevents a common error: treating a number of claims as though it were a number of policies or policy-years.

The calculations below use the standard limited-fluctuation approach for a Poisson claim count. It provides an approximate full-credibility threshold; it is not an exact probability calculation for every discrete Poisson distribution.

Definition: Let \(N\) be the claim count for a specified block of exposure, with expected count \(\lambda=E[N]\). A two-sided limited-fluctuation standard with confidence level \(p\) and relative tolerance \(k\) requires \(P(|N-\lambda|\le k\lambda)\ge p\). The claim-count threshold is the minimum expected claim count prescribed by the standard. If the expected claim frequency is \(\nu\) claims per exposure unit and the block has \(e\) exposure units, then \(\lambda=\nu e\).

The relative tolerance is a proportion of the expected count: for example, \(k=0.08\) allows a deviation of 8% of \(\lambda\). The confidence level is two-sided, so the standard asks that the count be within that range on either side of its mean. A 95% standard therefore leaves 2.5% in each tail under the normal approximation.

Why the Poisson Threshold Has This Form

Assume claims occur independently at a stable rate over the exposure being studied, so that \(N\) has a Poisson distribution with mean and variance \(\lambda\). For a sufficiently large expected count, approximate the standardized count by a standard normal random variable \(Z\). This is the approximation used in the limited-fluctuation claim-count standard.

P(|N-\lambda|\le k\lambda) \approx P\left(\left|\frac{N-\lambda}{\sqrt{\lambda}}\right|\le k\sqrt{\lambda}\right) \approx P(|Z|\le k\sqrt{\lambda}).

Let \(z_{(1+p)/2}\) denote the standard normal quantile with cumulative probability \((1+p)/2\). The central probability \(P(|Z|\le z_{(1+p)/2})\) equals \(p\). Thus the approximate confidence requirement is met when \(k\sqrt{\lambda}\) is at least that quantile. Squaring both nonnegative sides gives the threshold.

\lambda \ge \left(\frac{z_{(1+p)/2}}{k}\right)^2, \qquad N_0=\left(\frac{z_{(1+p)/2}}{k}\right)^2.

Here \(N_0\) is the unrounded threshold. For a practical claim-count threshold, round \(N_0\) up to the next whole claim count. Common two-sided normal quantiles are \(z_{0.95}=1.645\) for 90% confidence, \(z_{0.975}=1.960\) for 95%, and \(z_{0.995}=2.576\) for 99%.

Important condition: The formula is a normal-approximation standard for a Poisson count with a stable rate and independent claims. It does not model claim severity. If the count is small, the Poisson distribution’s discreteness can make the normal approximation inaccurate; the threshold should then be understood as the prescribed approximation, not an exact Poisson guarantee.

Worked Claim-Count and Exposure Calculations

Worked Example: A 90% Standard for a Hypothetical Fleet

A hypothetical fleet program wants its claim count to be within 8% of its expected count with 90% confidence. Its expected frequency is 0.12 claims per vehicle-year. Calculate the whole-claim threshold and the required whole vehicle-years.

Method. First calculate the count threshold from confidence and tolerance; then convert that rounded threshold to exposure. The calculation assumes a Poisson count with the stated constant frequency.

For 90% two-sided confidence, \(z_{(1+p)/2}=z_{0.95}=1.645\). The relative tolerance is \(k=0.08\), so

N_0=\left(\frac{1.645}{0.08}\right)^2 =(20.5625)^2 =422.8164.

Round up to a threshold of 423 expected claims. At a frequency of 0.12 claims per vehicle-year, the required exposure is

e=\frac{423\text{ claims}}{0.12\text{ claims per vehicle-year}} =3{,}525\text{ vehicle-years}.

Check: \(3{,}525(0.12)=423\) expected claims, matching the rounded threshold. The method avoids the common mistake of reporting 423 as the required number of vehicle-years; it is a claim-count threshold, and the frequency is needed to convert it to exposure.

Worked Example: A 95% Standard With a Tighter Tolerance

A hypothetical book of small commercial policies requires a claim count within 5% of its mean with 95% confidence. The expected frequency is 0.25 claims per policy-year. Find the minimum whole-claim threshold and the required whole policy-years.

For 95% confidence, use the central normal quantile \(z_{0.975}=1.960\). With \(k=0.05\),

N_0=\left(\frac{1.960}{0.05}\right)^2 =(39.2)^2 =1{,}536.64.

Rounding the expected-claim threshold up gives 1,537 claims. Convert this count using the frequency in claims per policy-year:

e=\frac{1{,}537\text{ claims}}{0.25\text{ claims per policy-year}} =6{,}148\text{ policy-years}.

Check: \(6{,}148(0.25)=1{,}537\) expected claims. The stricter tolerance and higher confidence both contribute to the larger threshold. A common error is to use 1.96 as though it were the confidence probability; it is the normal quantile associated with 95% central confidence.

Worked Example: Testing Whether a Hypothetical Portfolio Meets the Standard

A hypothetical specialty portfolio has 12,000 exposure units and an expected frequency of 0.04 claims per unit. Management specifies 99% confidence and a 10% relative tolerance. Calculate the claim threshold and required exposure, then determine whether the portfolio meets the standard.

For 99% two-sided confidence, \(z_{0.995}=2.576\), and \(k=0.10\). Therefore,

N_0=\left(\frac{2.576}{0.10}\right)^2 =(25.76)^2 =663.5776.

The whole-claim threshold is 664. The exposure required at 0.04 claims per unit is

e_{\mathrm{required}}=\frac{664}{0.04} =16{,}600\text{ exposure units}.

The portfolio’s expected claim count is

\lambda_{\mathrm{portfolio}}=12{,}000(0.04)=480\text{ claims}.

Since \(480<664\), the portfolio does not meet the specified full-credibility standard. It is short by \(16{,}600-12{,}000=4{,}600\) exposure units at the stated frequency. The check compares expected claims with the claim threshold, or actual exposure with required exposure; it does not compare the exposure count directly with 664 claims.

Converting Frequency and Rounding Carefully

The exposure conversion is a units calculation. If \(\nu\) is expressed as claims per policy-year, divide the required claim count by claims per policy-year to obtain policy-years. If frequency is supplied per a different unit, convert it before dividing. For example, a frequency of 60 claims per 1,000 exposure units is \(0.06\) claims per exposure unit, not 60 claims per unit.

A conservative practical sequence is to calculate the unrounded threshold, round the claim threshold up to a whole number, divide by the expected frequency, and round required exposure up to the next feasible whole exposure unit. This can be slightly conservative compared with retaining all fractional expected claims through the final calculation, but it makes the stated whole-claim requirement explicit. State the rounding convention so that the result can be checked.

Identify the standard.
Record the two-sided confidence level \(p\) and relative tolerance \(k\), using the tolerance as a decimal.

Find the normal quantile.
Use \(z_{(1+p)/2}\), not a one-sided quantile, for the central limited-fluctuation requirement.

Calculate the claim threshold.
Evaluate \((z_{(1+p)/2}/k)^2\), then apply the stated whole-claim rounding convention.

Convert to exposure or test the portfolio.
Divide the count threshold by claims per exposure unit, or calculate portfolio expected claims as exposure times frequency and compare.

Common Exam Traps

Using a one-sided quantile. The stated standard is two-sided: deviations above and below the expected count both matter. A 95% standard uses 1.960, not the one-sided 95th percentile of 1.645.
Entering a percentage instead of a decimal. A 5% tolerance is \(k=0.05\). Using 5 in the denominator changes the result by a factor of 10,000.
Forgetting to square the ratio. The threshold is \((z/k)^2\), because the Poisson standard deviation is \(\sqrt{\lambda}\).
Mixing claims and exposure. A frequency has units. Check that the claim units cancel when dividing the count threshold by claims per exposure unit.
Rounding down. A required minimum is rounded upward under the whole-count and whole-exposure convention used here.

The central reasoning is compact: confidence determines the normal quantile, tolerance determines how much deviation is allowed, and the Poisson variance links the two to a required expected count. Frequency then determines how much exposure is needed to generate that count. These results are aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

Check Your Understanding

Use the two-sided Poisson limited-fluctuation standard and the upward-rounding convention stated above.

A hypothetical vehicle block requires 90% confidence and a 10% relative tolerance. Its frequency is 0.20 claims per vehicle-year. Find the whole-claim threshold and required whole vehicle-years. (Answer: 271 claims; 1,355 vehicle-years)
For a hypothetical count standard with 95% confidence and a 4% relative tolerance, calculate the whole-claim threshold. (Answer: 2,401 claims)
A hypothetical account requires 99% confidence and an 8% relative tolerance. Its expected frequency is 0.05 claims per exposure unit. Calculate the whole-claim threshold and required exposure. (Answer: 1,037 claims; 20,740 exposure units)
A hypothetical portfolio requires 95% confidence and a 6% tolerance. Its frequency is 0.12 claims per exposure unit. Calculate the required whole-claim threshold and exposure. (Answer: 1,068 claims; 8,900 exposure units)
A hypothetical block has 5,000 exposure units and an expected frequency of 0.10 claims per unit. Under a 90% confidence, 5% tolerance standard, calculate the threshold and determine whether the block meets it. (Answer: 1,083 claims; expected count is 500, so it does not meet the standard)

What You'll Learn