Common Errors in Credibility Threshold Calculations

Why Credibility Thresholds Are Easy to Miscalculate

A limited-fluctuation credibility standard asks how much experience is needed for an observed result to fall sufficiently close to its expected value with a specified probability. The arithmetic may be short, but the setup matters: a one-sided probability uses a different normal quantile from a two-sided probability; a required claim count is not the same as a required number of policy-years; and an absolute dollar tolerance is not a percentage tolerance.

This tutorial focuses on diagnosing those setup errors. The formulas below use normal approximations to the relevant sampling distributions. They are useful standards for sufficiently large volumes, but they are not exact finite-sample probability guarantees. In particular, claim counts are discrete, and a normal approximation can be inaccurate when the expected count is small.

Definition: A credibility threshold is the minimum amount of experience, expressed in the unit required by the model, that meets a stated probability and tolerance standard. The threshold may first be calculated in claims or exposure units; these units must not be interchanged without a frequency conversion.

Start With the Probability Statement

Let $M$ denote the expected quantity being estimated and $\widehat{M}$ its estimate. A two-sided relative-tolerance standard requires the estimate to be within a proportion $k$ of the expected value:

P\bigl(|\widehat{M}-M|\leq kM\bigr)\geq p.

A one-sided standard instead limits error in just one direction. For example, it may require the estimate not to fall more than $kM$ below $M$. These are different probability statements and generally produce different thresholds. For a standard normal random variable $Z$, use $z_q$ to denote the $q$-quantile, so $P(Z\leq z_q)=q$. A central two-sided probability $p$ uses $z_{(1+p)/2}$; a one-sided probability $p$ uses $z_p$.

Key rule: Write the probability event before selecting a normal quantile. For a central two-sided standard use $z_{(1+p)/2}$; for a one-sided standard use $z_p$. Do not choose a quantile from the confidence percentage alone.

Claim Counts: Threshold in Claims, Then Convert to Exposure

Suppose claims follow a Poisson model with frequency $\lambda$ claims per exposure unit, over $e$ exposure units. Then the claim count $N$ has mean and variance $\lambda e$. Under the normal approximation, standardizing the count gives a standard deviation of $\sqrt{\lambda e}$. For a two-sided relative tolerance $k$, the approximate probability of being within $k$ of the expected count is:

P\bigl(|N-\lambda e|\leq k\lambda e\bigr) \approx 2\Phi\bigl(k\sqrt{\lambda e}\bigr)-1.

Requiring this probability to be at least $p$ gives $k\sqrt{\lambda e}\geq z_{(1+p)/2}$. Therefore, the required expected number of claims is approximately $\bigl(z_{(1+p)/2}/k\bigr)^2$. If exposure is requested, divide that count threshold by the frequency $\lambda$.

N_{\mathrm{req}}\approx\left(\frac{z_{(1+p)/2}}{k}\right)^2, \qquad e_{\mathrm{req}}\approx\frac{N_{\mathrm{req}}}{\lambda}.

The one-sided version uses $z_p$ in place of $z_{(1+p)/2}$. These formulas assume a Poisson count model and a normal approximation to the count. If the problem instead states a fixed claim count, use that count directly; do not treat it as an exposure requirement.

Worked Example: Two-Sided Claim-Count Standard

A hypothetical portfolio has a Poisson claim frequency of 4 claims per policy-year. Determine the approximate number of policy-years required for the observed count to be within 10% of its expected value with 90% two-sided confidence. Use $z_{0.95}=1.645$, and round the final exposure requirement up to a whole policy-year.

Method. First calculate the required expected claim count, then convert it to policy-years. The central 90% standard uses the 95th percentile because $(1+0.90)/2=0.95$.

The required expected claim count is:

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

The frequency is 4 claims per policy-year, so the exposure threshold is:

e_{\mathrm{req}} = \frac{270.6025\text{ claims}}{4\text{ claims per policy-year}} = 67.650625\text{ policy-years}.

Rounding the exposure requirement up gives 68 policy-years. At 68 policy-years the expected claim count is $68(4)=272$, which exceeds 270.6025. The calculation is approximate because it uses a normal approximation to a discrete Poisson count.

Common mistake avoided: Reporting 270.6025 as the number of policy-years confuses claims with exposure. Dividing by the stated frequency is essential.

Worked Example: One-Sided Standard and Unit Conversion

A hypothetical coverage has a Poisson frequency of 0.8 claims per exposure unit. Find the approximate exposure required for a one-sided 95% standard that limits the downward deviation of the claim count from its expectation to 20% of that expectation. Use $z_{0.95}=1.645$, and round exposure up to a whole unit.

Method. This is a one-sided standard, so use $z_{0.95}$, not the 97.5th percentile. The threshold is first obtained in expected claims:

N_{\mathrm{req}} = \left(\frac{1.645}{0.20}\right)^2 = 8.225^2 = 67.650625\text{ claims}.

Convert claims to exposure using 0.8 claims per exposure unit:

e_{\mathrm{req}} = \frac{67.650625}{0.8} = 84.56328125\text{ exposure units}.

Round up to 85 exposure units. At that volume, the expected count is $85(0.8)=68$ claims, above the threshold of 67.650625.

Common mistake avoided: Using the central 95% quantile $z_{0.975}=1.96$ would answer a different, two-sided question. Also, the result 67.650625 is a claim threshold, not the exposure requirement.

Relative and Absolute Tolerances Are Not Interchangeable

A relative tolerance is a proportion of the quantity being estimated. For example, “within 5% of the mean” has tolerance $kM$, so the permissible dollar error changes when $M$ changes. An absolute tolerance specifies a fixed amount, such as “within $100,” regardless of the size of the mean. Substituting a percentage for a fixed dollar amount, or vice versa, changes the problem.

For example, suppose independent exposure units each produce a random aggregate loss with mean $m$ dollars and variance $v$ dollars squared. For $e$ such units, the sample average has variance $v/e$. Under a normal approximation, a one-sided probability $p$ that the average is no more than $T$ dollars below $m$ requires $z_p\sqrt{v/e}\leq T$. Solving gives $e\geq z_p^2v/T^2$. The tolerance $T$ here is absolute, not a percentage of $m$.

Worked Example: Absolute Dollar Tolerance

For a hypothetical book, the aggregate loss per policy-year has mean $2,400 and variance 320,000 dollars squared. Assume policy-year losses are independent and identically distributed. How many policy-years are needed so that the average aggregate loss is no more than $100 below its expected value with one-sided 95% probability? Use $z_{0.95}=1.645$, and round up to a whole policy-year.

Method. The tolerance is an absolute $100. The variance of the average over $e$ independent policy-years is $320{,}000/e$. Thus the requirement is $1.645\sqrt{320{,}000/e}\leq100$, or equivalently:

e \geq \frac{z_{0.95}^{\,2}v}{T^2} = \frac{2.706025(320{,}000)}{(100)^2} = \frac{865{,}928}{10{,}000} = 86.5928\text{ policy-years}.

Rounding up gives 87 policy-years. As a check, at 87 policy-years the one-sided normal margin is $1.645\sqrt{320{,}000/87}\approx \$99.77$, which is below $100. At 86 policy-years it is approximately $1.645\sqrt{320{,}000/86}\approx \$100.35$, which exceeds $100.

Common mistake avoided: Replacing the $100 tolerance with a percentage of the $2,400 mean would impose a different standard. The given tolerance is fixed in dollars, so use $T=100$ directly.

Severity Variance: Use the Quantity the Model Requires

For a fixed number $n$ of independent claim severities with mean $\mu$ and variance $\sigma^2$, the sample average severity has variance $\sigma^2/n$. Under a normal approximation, requiring the average to be within a relative tolerance $k\mu$ with two-sided probability $p$ gives:

n \geq \frac{z_{(1+p)/2}^{\,2}\sigma^2}{k^2\mu^2}.

The input is the variance $\sigma^2$, not the standard deviation $\sigma$. If a problem supplies a standard deviation, square it before using this formula. Also distinguish a fixed number of claims from a random claim count: for a compound Poisson aggregate, the variance of aggregate loss per exposure depends on the severity second moment $E[X^2]=\sigma^2+\mu^2$, not on severity variance alone.

Derivation: For fixed $n$, independence gives $\operatorname{Var}(\bar{X})=\sigma^2/n$. Standardizing the two-sided event $|\bar{X}-\mu|\leq k\mu$ yields an approximate probability $2\Phi(k\mu\sqrt{n}/\sigma)-1$. Requiring this to be at least $p$ means $k\mu\sqrt{n}/\sigma\geq z_{(1+p)/2}$. Squaring both sides and rearranging gives $n\geq z_{(1+p)/2}^{\,2}\sigma^2/(k^2\mu^2)$. The result relies on independence and a normal approximation to the average.

For a random Poisson claim count with frequency $\lambda$ per exposure unit and independent severities $X$, the aggregate loss $S$ over $e$ units has mean $e\lambda\mu$ and variance $e\lambda(\sigma^2+\mu^2)$. The additional $\mu^2$ term reflects variation in the number of claims. If the target is a two-sided relative tolerance around the expected aggregate loss, the corresponding normal-approximation threshold is:

e \geq \frac{z_{(1+p)/2}^{\,2}(\sigma^2+\mu^2)} {k^2\lambda\mu^2}.

This aggregate-loss formula has specific assumptions: a Poisson count, independent claim severities with finite second moment, and a normal approximation at the volume considered. It should not be substituted for the fixed-claim-count severity formula.

Worked Example: Severity Variance Versus Standard Deviation

A hypothetical portfolio has a fixed number of independent claims with severity mean $1,200 and severity standard deviation $600. Determine the approximate number of claims required for the average severity to be within 5% of its mean with 90% two-sided probability. Use $z_{0.95}=1.645$, and round up to a whole claim.

Method. The formula requires the severity variance. Convert the stated standard deviation before substituting:

\sigma^2=(600)^2=360{,}000, \qquad k\mu=0.05(1{,}200)=\$60.

The required claim count is:

n \geq \frac{(1.645)^2(360{,}000)}{(60)^2} = \frac{2.706025(360{,}000)}{3{,}600} = \frac{974{,}169}{3{,}600} = 270.6025.

Rounding up gives 271 claims. For 271 claims, the approximate standard error of average severity is $600/\sqrt{271}\approx \$36.45$, and the 1.645-standard-error margin is approximately $59.96, just below the $60 tolerance. With 270 claims, the corresponding margin is approximately $60.07, just above the tolerance.

Common mistake avoided: Using 600 as the variance would omit the required squaring and produce a result with incorrect units. The variance is 360,000 dollars squared.

A Reliable Threshold-Calculation Check

Before finalizing a result, check the probability direction, the model, the tolerance, the input variance, the units, and the rounding. A useful sequence is to translate the words of the standard into an event, select the corresponding quantile, solve for the model's natural volume measure, then convert only if the requested answer uses another unit.

Write the event.
State whether the condition is one-sided or two-sided, and whether the allowed error is relative or absolute.

Match the model and quantile.
Use the variance appropriate to fixed counts, random counts, or an average, and choose the normal quantile from the stated probability event.

Track units.
Keep claims, policy-years, dollars, and dollars squared distinct. Convert from claims to exposure using the claim frequency when needed.

Check and round.
Substitute the result back into the inequality and round a minimum requirement upward when the unit must be whole.

These checks matter because a numerically plausible answer can still represent the wrong standard. In particular, choosing a two-sided quantile for a one-sided event, reporting expected claims as policy-years, or using variance where a second moment is required can materially distort the indicated volume.

Check Your Understanding

For each problem, identify the requested unit and apply the stated approximation. Round required counts or exposure units upward where specified.

A Poisson claim count must meet a one-sided 90% standard with relative tolerance 8%. Use $z_{0.90}=1.282$. What is the approximate required expected claim count? (Answer: 257 claims)
A hypothetical portfolio has a frequency of 2.5 claims per exposure unit. Using the threshold from the preceding problem, what whole number of exposure units is required? (Answer: 103 exposure units)
For a fixed number of independent claims, severity has mean $800 and variance 40,000 dollars squared. Find the number of claims required for the average severity to be within 10% of its mean with two-sided 95% probability, using $z_{0.975}=1.96$. (Answer: 25 claims)
Claim counts are Poisson with frequency 1.5 per exposure unit. Severity has mean $500 and variance 90,000 dollars squared. Using a two-sided 90% standard, $z_{0.95}=1.645$, and a 10% relative tolerance, find the required exposure using the compound Poisson aggregate-loss standard. (Answer: 246 exposure units)
Independent policy-year aggregate losses have variance 125,000 dollars squared. How many policy-years are required for the average to be within an absolute $50 of its expected value with two-sided 90% probability? Use $z_{0.95}=1.645$. (Answer: 136 policy-years)

What You'll Learn