Scaling Credibility Standards by Exposure

Why Credibility Standards Must Be Scaled

A full-credibility standard is often expressed as a required expected number of claims. That count by itself does not say how many policy-years, member-months, or other exposure units are needed. The conversion depends on the expected claim frequency per exposure unit. A relatively high-frequency group can reach the required count with less exposure; a lower-frequency group needs more.

This tutorial uses the Poisson normal-approximation standard for claim frequency. The key operational step is to match the units: multiply exposure by claim frequency to obtain expected claims, then solve for the exposure that produces the required expected count. This is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Definition: Let \(E\) be the amount of exposure, measured in a stated unit, and let \(\lambda\) be the expected claim frequency per that unit. Under a homogeneous Poisson model, the claim count \(N\) has mean \(m=E\lambda\). For a specified normal multiplier \(z\) and relative error tolerance \(k\), the usual full-credibility standard requires an expected count of at least \(m_0=(z/k)^2\). Therefore, the exposure requirement is \(E_0=m_0/\lambda=(z/k)^2/\lambda\).

From Required Claim Count to Required Exposure

Here, \(k\) is the maximum relative error expressed as a decimal: for example, a 10% tolerance means \(k=0.10\). The multiplier \(z\) is chosen for the stated probability criterion. For a central 95% normal criterion, for instance, \(z\) is approximately 1.96. The count standard is applied to the expected count, not to the number of claims that happen to be observed in one particular period.

If the expected frequency is \(\lambda\) claims per policy-year and the portfolio has \(E\) policy-years, then its expected count is \(E\lambda\). Setting this expected count equal to the required count gives \(E\lambda=m_0\). Dividing by \(\lambda\) gives the exposure requirement. If frequency is instead stated per member-month, the answer initially comes out in member-months; convert it only after completing the calculation.

m_0=\left(\frac{z}{k}\right)^2, \qquad E_0=\frac{m_0}{\lambda} =\frac{z^2}{k^2\lambda}.

The result \(E_0\) is a threshold under the stated model and approximation. If exposure can be measured continuously, the calculated value is the minimum exposure at the threshold. If the problem requires a whole number of complete policies or exposure units, round up to the next whole unit; rounding down could leave the expected count below the standard.

Identify the claim-count standard.
Record the specified \(z\) and relative tolerance \(k\), and calculate \(m_0=(z/k)^2\).

Identify the frequency unit.
Write \(\lambda\) as claims per exposure unit, such as claims per policy-year or claims per member-month.

Scale the count into exposure.
Calculate \(E_0=m_0/\lambda\). The units of \(\lambda\) determine the units of \(E_0\).

Check the result.
Multiply the exposure by the frequency and confirm that the expected count is at least \(m_0\).

Why the Scaling Formula Holds

The exposure scaling follows from the mean and variance of a Poisson count. Under the model, \(N\sim\operatorname{Poisson}(E\lambda)\), so \(E[N]=\operatorname{Var}(N)=E\lambda\). The estimated claim frequency is \(N/E\). Its variance is therefore \(\lambda/E\), and its standard deviation relative to the true frequency \(\lambda\) is \(1/\sqrt{E\lambda}\). The normal-approximation margin, expressed relative to \(\lambda\), is \(z/\sqrt{E\lambda}\). Requiring this margin to be no more than \(k\) gives:

\frac{z}{\sqrt{E\lambda}}\leq k \quad\Longleftrightarrow\quad E\lambda\geq\left(\frac{z}{k}\right)^2 \quad\Longleftrightarrow\quad E\geq\frac{(z/k)^2}{\lambda}.

Thus the required expected count is determined by the precision criterion, while the frequency converts that count into exposure. The count requirement does not change merely because the unit of exposure changes; the numerical value of \(\lambda\) and the numerical value of \(E\) change together so their product remains an expected claim count.

Worked Example: Required Policy-Years at a Known Frequency

A hypothetical portfolio has an expected claim frequency of 0.08 claims per policy-year. Using a central 95% normal criterion with \(z=1.96\), determine the policy-years needed for a relative error tolerance of 10%. State the expected count at the threshold and check the result.

Method. First calculate the full-credibility expected count; then divide by the frequency in claims per policy-year.

Step 1: Convert the tolerance. Ten percent is \(k=0.10\).

Step 2: Calculate the required expected count.

m_0=\left(\frac{1.96}{0.10}\right)^2 =19.6^2 =384.16\text{ claims}.

Step 3: Convert to exposure.

E_0=\frac{384.16\text{ claims}}{0.08\text{ claims per policy-year}} =4{,}802\text{ policy-years}.

Check. At 4,802 policy-years, the expected count is \(4{,}802(0.08)=384.16\) claims. The relative normal margin is \(1.96/\sqrt{384.16}=1.96/19.6=0.10\), exactly the specified 10% tolerance. The threshold calculation is consistent.

Common mistake avoided: Do not report 384.16 as the required policy-years. It is the required expected claim count. Dividing by the frequency is what converts claims into policy-years.

Using Existing Exposure to Assess the Standard

The same relationship can be rearranged when exposure and frequency are known but the relative tolerance is not. First calculate the expected count \(m=E\lambda\). The smallest tolerance supported by the normal-approximation standard is then \(k_{\min}=z/\sqrt{m}\). A group meets a proposed tolerance \(k\) when \(k_{\min}\leq k\).

Worked Example: Finding the Minimum Tolerance for Existing Exposure

A hypothetical group has 1,800 policy-years of exposure and an expected frequency of 0.075 claims per policy-year. Using \(z=1.96\), find the smallest relative error tolerance for which the expected count meets the normal-approximation standard. Give the result as a percentage to two decimal places.

Method. Calculate expected claims from exposure and frequency before using the tolerance formula.

Step 1: Find the expected claim count.

m=1{,}800(0.075)=135\text{ claims}.

Step 2: Find the minimum tolerance.

k_{\min}=\frac{1.96}{\sqrt{135}} =\frac{1.96}{11.61895\ldots} =0.16869\ldots.

As a percentage to two decimal places, the smallest qualifying tolerance is 16.87%. The standard is met for tolerances at least this large; a smaller tolerance would require more expected claims.

Check and common mistake avoided. The expected count is 135, not 1,800. Using exposure directly in the denominator gives \(1.96/\sqrt{1{,}800}=0.04620\), or 4.62%, but that calculation incorrectly treats every policy-year as one expected claim. At a 4.62% tolerance, the correct margin is \(1.96/\sqrt{135}=16.87\%\), which exceeds 4.62%; the standard is not met.

Matching Frequency and Exposure Units

A frequent source of error is mixing a rate stated per month with exposure stated in years, or vice versa. The units must cancel correctly in \(E\lambda\). If the frequency is 0.015 claims per member-month, multiplying it by member-months gives expected claims. If exposure is stated in member-years, either convert the frequency to claims per member-year or convert the exposure to member-months before multiplying.

Worked Example: Converting Member-Months to Member-Years

A hypothetical coverage block has an expected frequency of 0.015 claims per member-month. Find the exposure needed for a central 90% normal criterion, with \(z=1.645\), and a 10% relative error tolerance. Report the result in member-months and in whole member-years.

Method. Calculate the required expected count using the specified criterion, divide by the monthly frequency, and then convert months to years.

Step 1: Calculate the expected count standard. Here \(k=0.10\), so:

m_0=\left(\frac{1.645}{0.10}\right)^2 =16.45^2 =270.6025\text{ claims}.

Step 2: Calculate member-months.

E_0=\frac{270.6025}{0.015} =18{,}040.1667\text{ member-months}.

Step 3: Convert to member-years. There are 12 member-months in one member-year:

\frac{18{,}040.1667}{12} =1{,}503.3472\text{ member-years}.

If only whole member-years are allowed, round up to 1,504 member-years. The corresponding expected count is \(1{,}504(12)(0.015)=270.72\), which is at least 270.6025. Rounding down to 1,503 member-years would produce \(1{,}503(12)(0.015)=270.54\), just below the required expected count.

Common mistake avoided: Dividing the claim standard by 0.015 and labeling the answer as member-years would be a unit error. That division produces member-months; the conversion by 12 is still required.

Assumptions, Rounding, and Practical Interpretation

The scaling formula is appropriate when claim counts follow a Poisson model with a constant expected frequency per exposure unit, and when the full-credibility standard is the stated normal-approximation standard. Exposure is assumed to accumulate proportionally: twice as much exposure has twice the expected count. The formula is not a general guarantee for every count process. If the frequency varies materially across risks, or count variability differs from the Poisson assumption, the mean-count scaling may not provide the stated precision. The model and the standard must therefore be checked before applying the calculation.

Keep three quantities separate:

Exposure \(E\): the amount of observation, such as policy-years.
Frequency \(\lambda\): expected claims per exposure unit.
Expected count \(E\lambda\): the product used in the count standard.

When the question asks for a continuous exposure requirement, give the calculated threshold in the stated unit. When it asks for a whole number of policies or complete periods, round upward and check the expected count after rounding. Do not round intermediate values unnecessarily: a small rounding change can matter when the result is close to the threshold.

Key takeaway: First find the required expected claim count, \(m_0=(z/k)^2\). Then divide by the expected frequency per exposure unit: \(E_0=m_0/\lambda\). Always verify that exposure and frequency use matching units and that their product reaches the required count.

Check Your Understanding

For each problem, use the stated normal-approximation standard and show the expected-count step before scaling by frequency.

A hypothetical book has frequency 0.04 claims per policy-year. For a central 95% criterion with \(z=1.96\) and relative tolerance 20%, find the required policy-years. (Answer: 2,401 policy-years.)
A group has 2,500 policy-years and expected frequency 0.06 claims per policy-year. With \(z=1.645\), find the smallest relative tolerance, as a percentage to two decimal places, supported by the expected count. (Answer: 13.43%.)
A hypothetical insured population has frequency 0.02 claims per insured-year. For \(z=2.576\) and relative tolerance 5%, find the required insured-years, then state the number of whole insured-years needed. (Answer: 132,715.52 insured-years; 132,716 whole insured-years.)
A monthly coverage block has frequency 0.03 claims per member-quarter. For \(z=1.96\) and relative tolerance 15%, find the required member-quarters and the minimum whole number of member-quarters. (Answer: 5,691.26 member-quarters; 5,692 whole member-quarters.)
A portfolio has 1,000 policy-years at frequency 0.12 claims per policy-year. With \(z=1.96\), calculate the smallest relative tolerance to two decimal places. Explain why using 1,000 instead of the expected count gives the wrong result. (Answer: 17.89%; the expected count is 120, not 1,000.)

What You'll Learn