Full Credibility for Claim Frequency Rates

Why Frequency-Rate Precision Depends on Exposure

A claim count describes how many events occurred; a claim frequency rate describes how many events occurred per unit of exposure. For example, a hypothetical portfolio might record claims per policy-year or per vehicle-year. When exposure differs between portfolios, comparing raw claim counts alone can be misleading. The frequency-rate estimate puts the count on a common exposure basis, while its precision depends on both the count variation and the amount of exposure observed.

This tutorial uses the Poisson model and the normal approximation underlying the standard limited-fluctuation calculation. The resulting full-credibility standard is stated as a required expected number of claims. It is not a guarantee that a particular observed count will fall within a specified error band.

Definition: Let $e>0$ be total exposure, measured in a specified exposure unit, and let $N$ be the number of claims observed over that exposure. Assume claims follow a homogeneous Poisson model with frequency rate $\lambda>0$ per exposure unit. Then $N\sim\operatorname{Poisson}(\lambda e)$, so $E[N]=\lambda e$ and $\operatorname{Var}(N)=\lambda e$. The estimated frequency rate is $\hat{\lambda}=N/e$.

The model assumes a constant claim rate over the exposure being combined, and that the count has Poisson variation: its mean and variance are both $\lambda e$. The exposure units must be consistent. If $e$ is in policy-years, then $\lambda$ is claims per policy-year; if exposure is in vehicle-years, the rate is claims per vehicle-year. This model does not allow extra-Poisson variation, changing rates across the observations, or dependence between claim events.

Deriving the Precision of the Rate Estimate

The count is random, but $e$ is treated as fixed. Dividing $N$ by $e$ therefore divides its mean by $e$ and its variance by $e^2$. This gives an unbiased rate estimate and its variance:

E[\hat{\lambda}] =E\left[\frac{N}{e}\right] =\frac{\lambda e}{e} =\lambda, \qquad \operatorname{Var}(\hat{\lambda}) =\operatorname{Var}\left(\frac{N}{e}\right) =\frac{\lambda e}{e^2} =\frac{\lambda}{e}.

The standard error is consequently $\sqrt{\lambda/e}$. Since the true rate $\lambda$ is the quantity being estimated, the standard error relative to that rate is $\sqrt{\lambda/e}/\lambda=1/\sqrt{\lambda e}$. Write $m=\lambda e$ for the expected number of claims. The relative standard error is then $1/\sqrt{m}$: precision improves as the expected claim count increases.

Result: Under the Poisson model, the estimated rate has mean $\lambda$, variance $\lambda/e$, and relative standard error $1/\sqrt{\lambda e}$. Here $m=\lambda e$ is the expected claim count over the exposure.

For a full-credibility standard, a relative error limit $r$ means that the estimated rate is required to lie within $r\lambda$ of the true rate. A two-sided confidence requirement includes deviations in both directions. Using the normal approximation to the Poisson count, the standardized rate estimate is approximately standard normal:

\frac{\hat{\lambda}-\lambda}{\sqrt{\lambda/e}} \ \approx\ N(0,1).

Let $p$ be the desired two-sided probability and let $z_{(1+p)/2}$ be the standard normal quantile with cumulative probability $(1+p)/2$. The approximate probability of meeting the relative-error requirement is $P(|\hat{\lambda}-\lambda|\le r\lambda)\approx P(|Z|\le r\sqrt{\lambda e})$. Requiring this probability to be at least $p$ leads directly to the full-credibility standard.

Set the error band.
The desired estimate is within $r\lambda$ of $\lambda$, so the standardized error band has half-width $r\lambda/\sqrt{\lambda/e}=r\sqrt{\lambda e}=r\sqrt{m}$.

Apply the normal quantile.
For a two-sided probability $p$, the corresponding standard-normal band is $[-z_{(1+p)/2},z_{(1+p)/2}]$. Thus the error band must satisfy $r\sqrt{m}\ge z_{(1+p)/2}$.

Solve for expected claims.
Squaring both nonnegative sides gives $m\ge \bigl(z_{(1+p)/2}/r\bigr)^2$. The required expected-claim standard is the smallest whole number at least this large.

Theorem: For a two-sided relative-error requirement $r$ and confidence level $p$, the Poisson normal-approximation full-credibility standard is $$ m_0=\left(\frac{z_{(1+p)/2}}{r}\right)^2 $$ expected claims, rounded up to a whole claim. This standard applies when the Poisson model and normal approximation are appropriate; it is not an exact finite-count probability calculation.

For a one-sided requirement, the quantile changes because the probability is placed in one tail rather than split between two tails. For example, if the requirement is that the estimate not exceed the true rate by more than $r\lambda$, the approximate condition is $P(\hat{\lambda}-\lambda\le r\lambda)\ge p$. With $z_p$ the standard-normal quantile at probability $p$, this gives $m_0=(z_p/r)^2$. The direction of the one-sided condition must be stated; it is not the same criterion as being within a band on both sides.

Common two-sided quantiles are $z_{0.95}\approx1.645$ for 90% confidence, $z_{0.975}\approx1.960$ for 95%, and $z_{0.995}\approx2.576$ for 99%. Use the quantile specified by the confidence criterion, not a memorized value detached from whether the requirement is one-sided or two-sided.

Worked Applications

Worked Example: Finding a Two-Sided Claim Standard

A hypothetical insurer wants its estimated claim frequency per policy-year to be within 10% of the true rate with approximate 95% confidence. Find the required full-credibility standard in expected claims.

Method. Use the two-sided Poisson normal-approximation standard. Here $r=0.10$, $p=0.95$, and $z_{(1+p)/2}=z_{0.975}=1.96$.

Substitute into the formula:

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

Expected claims must be at least 384.16. Since the standard is expressed as a whole number of claims, round up: the full-credibility standard is 385 expected claims. The square root of 384.16 is 19.6, and $0.10(19.6)=1.96$, verifying that the selected quantile and error limit match. Rounding down to 384 would not meet the calculated threshold. This avoids the common mistake of rounding a minimum claim standard to the nearest integer instead of rounding up.

Worked Example: Checking an Estimated Rate Against the Standard

A hypothetical block has 6,000 policy-years of exposure and records 540 claims. Estimate its frequency rate and assess whether the observed experience meets a two-sided 95% relative-error standard of 10%, using the observed rate as a plug-in estimate of the expected rate.

Method. First calculate the rate, then calculate its estimated standard error under the Poisson model. The estimated rate is $\hat{\lambda}=540/6{,}000=0.09$ claims per policy-year. Using this as the plug-in value gives an estimated expected count of $0.09(6{,}000)=540$.

The plug-in estimated variance and standard error are

\widehat{\operatorname{Var}}(\hat{\lambda}) =\frac{0.09}{6{,}000} =0.000015, \qquad \widehat{\operatorname{SE}}(\hat{\lambda}) =\sqrt{0.000015} \approx0.003873.

The estimated relative standard error is $0.003873/0.09\approx0.04303$, or 4.303%. The approximate 95% relative-error margin is $1.96(0.04303)\approx0.08434$, or 8.434%, which is less than 10%. Equivalently, the observed plug-in expected count of 540 exceeds the standard of 384.16 expected claims. Thus the portfolio meets this approximate standard. As a rate interval check, $0.09\pm1.96(0.003873)\approx[0.0824,0.0976]$; its half-width is about $0.0076$, or 8.434% of $0.09$. Rates are rounded to four decimal places in this check.

The comparison uses the estimated rate as a plug-in value because the true rate is unknown. The full-credibility formula itself is stated in terms of the expected claim count under the assumed rate. Do not confuse a plug-in assessment based on observed data with a guarantee about the next observed count.

Worked Example: A One-Sided Standard

A hypothetical administrator requires at least 95% approximate probability that the estimated frequency rate does not exceed the true rate by more than 8% of that rate. Find the one-sided expected-claim standard.

Method. This is an upper-deviation, one-sided requirement, so use $z_{0.95}=1.645$, not the two-sided 95% quantile $1.96$. The relative error limit is $r=0.08$.

m_0=\left(\frac{1.645}{0.08}\right)^2 =20.5625^2 =422.8164.

Rounding upward gives 423 expected claims. Substitution checks the boundary: $0.08\sqrt{422.8164}=0.08(20.5625)=1.645$. Using 1.96 instead would impose a two-sided criterion and produce a larger standard; it would answer a different question.

Interpretation and Common Traps

The standard is about the expected count $m=\lambda e$, not the numerical value of the rate by itself. It is therefore reported in expected claims, even though the quantity being estimated is claims per exposure unit. The rate’s unit follows from the exposure unit, while the claim-count threshold is dimensionless.

Using a count as if it were a rate. The estimator is $N/e$, not $N$. Dividing by exposure is essential when estimating frequency per exposure unit.
Using the wrong quantile. A two-sided confidence criterion uses $z_{(1+p)/2}$; a one-sided criterion uses $z_p$ for the specified tail.
Rounding the threshold down. A minimum standard must be rounded up to the next whole expected claim.
Mixing exposure units. The rate and exposure must use matching units for $\lambda e$ to be the expected claim count.
Treating the approximation as exact. The standard comes from a normal approximation to Poisson variation. It is less dependable when expected counts are small or when the Poisson assumptions do not describe the data.

A frequent incorrect calculation is to compare a rate estimate directly with the expected-claim threshold, such as comparing $0.09$ claims per policy-year with 384.16 claims. Those quantities have different meanings and units. The correction is to calculate the expected count $m=\lambda e$, or its stated plug-in estimate, and compare that count with the expected-claim standard. This tutorial does not replace the assumed Poisson model with a different count model; if the data are overdispersed or rates vary materially, the Poisson standard may not be suitable.

Key takeaway: Under a Poisson frequency model, the rate estimate’s relative standard error is $1/\sqrt{m}$, where $m$ is expected claims. A specified relative error and confidence level therefore translate into a minimum expected-claim count through the appropriate normal quantile.

Check Your Understanding

Use the Poisson normal-approximation standards and round each minimum expected-claim standard upward.

A two-sided 95% standard requires frequency-rate error within 5% of the true rate. Find the expected-claim standard. (Answer: 1,537 expected claims.)
For a two-sided 90% standard with a 12% relative-error limit, use $z_{0.95}=1.645$. Find the expected-claim standard. (Answer: 188 expected claims.)
A hypothetical portfolio has 9,000 exposure units and 225 claims. Using the observed rate as a plug-in estimate, does it meet a two-sided 95% standard with a 10% relative-error limit? (Answer: No; the estimated expected count is 225, below the standard of 385.)
A one-sided 95% upper-deviation criterion has a 8% relative-error limit. Use $z_{0.95}=1.645$. Find the expected-claim standard. (Answer: 423 expected claims.)
A hypothetical portfolio has 16,000 vehicle-years and 640 claims. Using the observed rate as a plug-in estimate, does it meet a two-sided 99% standard with a 10% relative-error limit? Use $z_{0.995}=2.576$. (Answer: No; the estimated expected count is 640, below the standard of 664.)

What You'll Learn