Partial Credibility for Claim Frequency

Introduction

An observed claim frequency can be informative without meeting the standard required for full credibility. In that situation, limited-fluctuation credibility assigns part of the weight to the observed experience and the remainder to a manual frequency. The square-root rule supplies the weight from the amount of exposure available relative to the full-credibility standard.

For claim frequency, the key quantities are the number of claims, the exposure used to calculate the observed rate, and the full-credibility requirement. Keep the units consistent: the observed frequency and manual frequency must both be claims per the same exposure unit.

Definition: Let \(C\) be the number of claims observed over exposure \(E\), measured in a stated unit such as policy-years. The observed claim frequency is \(\hat{\lambda}=C/E\). If \(m\) is the manual claim frequency on the same basis and \(Z\) is the partial-credibility factor, the credibility-weighted frequency is \(Z\hat{\lambda}+(1-Z)m\). Under the square-root rule, \(Z=\min(1,\sqrt{E/E_0})\), where \(E_0\) is the exposure required for full credibility under the stated standard.

From the Full Standard to a Frequency Estimate

The full-credibility standard may be given as an exposure amount or as a required number of expected claims. If it is given as an exposure \(E_0\), compare the available exposure directly with that amount. If it is given as a full-credibility expected-claim threshold \(K\), and the appropriate manual frequency is \(m\), the corresponding full-credibility exposure is \(E_0=K/m\). Thus the rule can also be written \(Z=\min(1,\sqrt{mE/K})\). The quantity \(mE\) is the expected number of claims at the manual rate, not the actual number observed.

Z=\min\left(1,\sqrt{\frac{E}{E_0}}\right) =\min\left(1,\sqrt{\frac{mE}{K}}\right) \qquad\text{when } E_0=\frac{K}{m}.

These forms express the same calculation when \(K\), \(m\), and \(E_0\) describe the same standard. Do not use the expected-claim form unless the threshold and manual rate refer to compatible risks and exposure units. The square-root rule is a limited-fluctuation convention for assigning partial credibility; it is not a claim that the credibility-weighted frequency is an exact posterior estimate under every model.

Why the Square Root Appears

For a simple justification, suppose claims follow a homogeneous Poisson process over exposure: \(C\sim\operatorname{Poisson}(\lambda E)\), where \(\lambda\) is the underlying frequency and exposure is measured consistently. Then \(E[C]=\lambda E\) and \(\operatorname{Var}(C)=\lambda E\). Since the observed frequency is \(C/E\), its variance is \(\lambda/E\), so its standard deviation is \(\sqrt{\lambda/E}\). Increasing exposure therefore reduces the standard deviation in proportion to \(1/\sqrt{E}\).

At exposure \(E_0\), the same model gives standard deviation \(\sqrt{\lambda/E_0}\). The ratio of the standard deviation at full-standard exposure to that at available exposure is

\frac{\sqrt{\lambda/E_0}}{\sqrt{\lambda/E}} =\sqrt{\frac{E}{E_0}}.

This relative precision scaling motivates the square-root factor. When the available exposure reaches or exceeds \(E_0\), the factor is capped at one: the experience receives full credibility, not more than full credibility. The formula presumes exposure is a suitable measure of the amount of comparable experience. Material changes in risk mix, coverage, claim reporting, or the definition of exposure can make a direct comparison inappropriate.

Worked Claim-Frequency Calculations

The calculation has two distinct stages: determine the weight from exposure, then apply that weight to the two rates. The claim count determines the observed rate, but it does not replace exposure in the square-root credibility factor.

Worked Example: A Small Portfolio With a Known Exposure Standard

A hypothetical insurer observes 28 claims over 100 policy-years. Its manual frequency is 0.20 claims per policy-year. The full-credibility standard is 80 expected claims at that manual frequency. Calculate the credibility-weighted frequency.

Method and standard. Use the expected-claim version of the square-root rule. The frequency is assumed comparable across the observed and manual experience, and exposure is measured in policy-years.

Step 1: Find the full-credibility exposure. At 0.20 claims per policy-year, 80 expected claims correspond to

E_0=\frac{80}{0.20}=400\text{ policy-years}.

Step 2: Calculate the observed frequency.

\hat{\lambda}=\frac{28}{100}=0.28\text{ claims per policy-year}.

Step 3: Calculate the credibility factor. The observed exposure is one-quarter of the full-credibility exposure, so

Z=\sqrt{\frac{100}{400}}=\sqrt{0.25}=0.50.

Step 4: Blend the rates.

\hat{\lambda}_{\mathrm{cred}} =0.50(0.28)+0.50(0.20) =0.14+0.10 =0.24.

The indicated frequency is 0.2400 claims per policy-year. As a check, it is halfway between 0.28 and 0.20, as required by \(Z=0.50\). The common mistake this method avoids is using the observed 28 claims in place of expected claims to set \(Z\). For example, \(\sqrt{28/80}\) would make the weight depend on a random fluctuation in the claim count, rather than on the amount of exposure.

Worked Example: Converting an Expected-Claim Standard

A hypothetical equipment insurer has 225 policy-years of experience and observes 108 claims. The manual frequency is 0.40 claims per policy-year, and full credibility requires 360 expected claims. Find the partial-credibility estimate.

Method and standard. First translate the expected-claim threshold into policy-years, then apply the capped square-root rule and blend frequencies.

Step 1: Convert the standard.

E_0=\frac{360}{0.40}=900\text{ policy-years}.

Step 2: Calculate the observed rate and credibility factor.

\hat{\lambda}=\frac{108}{225}=0.48, \qquad Z=\min\left(1,\sqrt{\frac{225}{900}}\right) =\sqrt{0.25}=0.50.

The expected number of claims at the manual rate is \(0.40(225)=90\), and \(90/360=0.25\), which gives the same credibility factor: \(\sqrt{0.25}=0.50\).

Step 3: Blend.

\hat{\lambda}_{\mathrm{cred}} =0.50(0.48)+0.50(0.40) =0.24+0.20 =0.44.

The result is 0.4400 claims per policy-year. It lies between the manual and observed frequencies. A unit check confirms that both are claims per policy-year; the expected-claim threshold itself is a count and must be converted before it is compared with policy-years.

Worked Example: Credibility From Expected Claims Without Reaching Full Credibility

A hypothetical roadside-assistance portfolio has 1,500 covered vehicle-years. It experiences 141 claims. The manual frequency is 0.08 claims per vehicle-year, and the full-credibility standard is 200 expected claims. Calculate the indicated frequency, rounding the final rate to four decimal places.

Method and standard. Use \(mE\) to find expected claims at the manual rate, then compare that quantity with the full-credibility threshold. The observed and manual rates use vehicle-years as their common exposure basis.

Step 1: Calculate expected claims and the credibility factor.

mE=0.08(1{,}500)=120, \qquad Z=\sqrt{\frac{120}{200}}=\sqrt{0.60}=0.7745967\ldots.

The factor is below one, so no cap changes it.

Step 2: Calculate the observed rate.

\hat{\lambda}=\frac{141}{1{,}500}=0.094.

Step 3: Blend the rates. Using the manual rate as the starting point makes the adjustment easy to verify:

\begin{aligned} \hat{\lambda}_{\mathrm{cred}} &=0.7745967(0.094)+\left(1-0.7745967\right)(0.08)\\ &=0.08+0.7745967(0.094-0.08)\\ &=0.08+0.7745967(0.014)\\ &=0.0908443\ldots. \end{aligned}

Rounded to four decimal places, the indicated frequency is 0.0908 claims per vehicle-year. It is closer to the observed rate than to the manual rate because \(Z\) exceeds one-half, but it remains between the two. The method avoids treating the 141 actual claims as the numerator of the credibility ratio: the numerator in this standard is the 120 expected claims at the manual frequency.

Calculation Process and Checks

A consistent sequence prevents most errors in frequency credibility problems. In particular, calculate the observed rate and credibility factor separately before combining them.

Identify the exposure basis.
Record the unit used for \(E\), such as policy-years or vehicle-years. Confirm that both the observed and manual frequencies are expressed per that same unit.

Put the full standard on the same basis.
If the standard is \(E_0\), use it directly. If it is \(K\) expected claims, use the applicable manual frequency \(m\) to calculate \(E_0=K/m\).

Calculate \(Z\).
Use \(Z=\min(1,\sqrt{E/E_0})\), or the equivalent expected-claim form when its inputs are compatible.

Calculate the observed rate and blend.
Find \(\hat{\lambda}=C/E\), then use \(Z\hat{\lambda}+(1-Z)m\). Retain adequate precision until the final rate.

As a final check, \(0\leq Z\leq1\), and the weighted frequency must lie between the observed and manual frequencies. If \(E=0\), no observed frequency \(C/E\) can be calculated; this blending formula cannot be applied from zero exposure. When \(E\geq E_0\), \(Z=1\) and the indicated frequency equals the observed frequency. These checks do not establish that the manual rate is appropriate; that remains a modeling judgment.

Common Exam Traps and Interpretation

Using claims instead of exposure to determine \(Z\). The random observed count \(C\) belongs in the observed rate \(C/E\). The square-root factor is based on exposure, or on expected claims under the standard's specified basis.
Forgetting the square root. If \(E/E_0=0.25\), then \(Z=0.50\), not 0.25. The ratio is squared relative to the credibility factor.
Mixing exposure units. Policy-years cannot be compared directly with vehicle-years unless the problem establishes that they represent equivalent exposure.
Failing to cap the factor. Exposure above the full standard does not produce \(Z>1\). Use the cap, so full credibility gives the observed rate its full weight.
Blending inconsistent rates. A frequency per month and one per policy-year cannot be averaged until they are converted to the same time and exposure basis.

This technique is useful when experience is too limited to meet a full-credibility standard but still deserves some weight. Its interpretation is deliberately narrow: \(Z\) is a prescribed partial-credibility weight under the stated standard, not a probability that the observed frequency is correct. The calculation is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Key takeaway: For claim frequency, determine partial credibility from comparable exposure relative to the full-credibility standard. Calculate the observed rate as claims divided by exposure, then blend it with the manual rate using \(Z\hat{\lambda}+(1-Z)m\).

Check Your Understanding

For each problem, calculate the square-root credibility factor and the indicated frequency. Frequencies are claims per the exposure unit stated.

A hypothetical book has 64 policy-years, 12 claims, and a manual frequency of 0.125 per policy-year. Its full-credibility exposure is 256 policy-years. Find \(Z\) and the indicated frequency, rounding the rate to four decimals. (Answer: \(Z=0.5000\); indicated frequency \(=0.1563\) claims per policy-year.)
A hypothetical portfolio has 1,000 vehicle-years and 315 claims. Its manual frequency is 0.30 per vehicle-year, and full credibility requires 150 expected claims. Find the indicated frequency. (Answer: \(Z=1.0000\); indicated frequency \(=0.3150\) claims per vehicle-year.)
A hypothetical service plan has 360 covered units-years, 108 claims, and a manual frequency of 0.25 per unit-year. Full credibility requires 300 expected claims. Find the indicated frequency to four decimals. (Answer: \(Z=\sqrt{0.30}=0.5477\); indicated frequency \(=0.2774\) claims per unit-year.)
A hypothetical insurer has 500 policy-years, 65 claims, and a manual frequency of 0.10 per policy-year. Its full-credibility standard is 100 expected claims. Find \(Z\) and the indicated frequency to four decimals. (Answer: \(Z=\sqrt{0.50}=0.7071\); indicated frequency \(=0.1212\) claims per policy-year.)
Over three hypothetical periods, a portfolio has exposures of 40, 70, and 90 policy-years, with 7, 14, and 25 claims, respectively. The manual frequency is 0.20 per policy-year and full credibility requires 160 expected claims. Pool the experience and calculate the indicated frequency. (Answer: total exposure \(=200\) policy-years, total claims \(=46\), \(Z=0.5000\), and indicated frequency \(=0.2150\) claims per policy-year.)

What You'll Learn