Deriving the Square-Root Credibility Rule

Why Credibility Grows More Slowly Than Exposure

More experience generally makes an estimate more reliable, but reliability does not increase in direct proportion to the amount of data. If the number of independent observations is multiplied by four, the variance of their average is divided by four, while its standard deviation is divided only by two. This relationship explains why the limited-fluctuation partial credibility factor is a square root of the ratio of current experience to the full-credibility standard.

The derivation starts with a full-credibility standard: the amount of experience needed for an estimator to meet a specified error tolerance at a specified confidence level. Below that standard, the square-root rule expresses current experience as a fraction of the standard while respecting how sampling variability changes with volume. The method is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Definition: Let $n$ denote the amount of homogeneous, independent experience, and let $n_0$ denote the full-credibility standard under a specified precision criterion. When $n\leq n_0$, the square-root partial credibility factor is $$ Z=\sqrt{\frac{n}{n_0}}. $$ Here, $Z$ is a weight describing the relative credibility assigned to the current experience under the limited-fluctuation rule. It is not, by itself, the probability that the estimate meets the precision criterion.

Assumptions Behind the Derivation

The rule applies when the estimator’s sampling variance is inversely proportional to the amount of experience. For example, if $n$ independent observations have a common finite variance $\sigma^2$, their sample mean $\bar{X}$ has variance $\sigma^2/n$. A similar inverse-volume relationship holds for a Poisson frequency estimate based on exposure. The observations or exposure units must be sufficiently comparable for a single variance relationship to be appropriate.

A full-credibility standard also requires a stated error tolerance and confidence level. Under a normal model, or a normal approximation when it is suitable, a two-sided confidence level $p$ corresponds to a standard normal quantile $z$ satisfying $P(|N(0,1)|\leq z)=p$. If the permitted absolute error is $a$, a general estimator $T_n$ with mean $\theta$ and variance $C/n$ meets the full standard when $z\sqrt{C/n_0}=a$. Solving gives $n_0=z^2C/a^2$. The same reasoning works for a relative tolerance by setting $a$ equal to the allowed fraction of the target mean.

This relationship does not automatically apply if observations are dependent, their variances differ in a way the model does not account for, or the estimator’s variance does not decrease in proportion to $1/n$. Also, a normal-approximation confidence calculation is not an exact probability statement for every discrete or skewed model.

Deriving the Square-Root Rule

Write the variance of the estimator at volume $n$ as $C/n$, where $C$ is constant for the model and measure being studied. At the full standard $n_0$, the variance is $C/n_0$. The ratio of the standard deviation at the full standard to the standard deviation at current volume is

\frac{\sqrt{C/n_0}}{\sqrt{C/n}} =\sqrt{\frac{n}{n_0}}.

The constant $C$ cancels. Thus the square root appears because credibility is related to the scale of random error, measured by standard deviation, rather than to variance itself. The derivation can also be understood by comparing the precision criterion at the two volumes: a standard error shrinks as $1/\sqrt{n}$, so the corresponding relative credibility factor grows as $\sqrt{n}$.

State the estimator’s variance.
Express it in the form $C/n$, or establish the corresponding inverse-exposure relationship for the model.

Find the full standard.
Use the specified error tolerance and confidence level to determine $n_0$.

Compare standard deviations.
The ratio of the full-standard standard error to the current standard error is $\sqrt{n/n_0}$.

Interpret the result.
For experience below the full standard, $Z=\sqrt{n/n_0}$ is the partial credibility factor under this rule.

Worked Applications

Worked Example: Poisson Claim Frequency

A hypothetical insurer models claim counts as Poisson with rate $\lambda=0.05$ claims per policy-year. It seeks a two-sided 95.45% confidence standard for estimating $\lambda$ with relative error no greater than 10%. Use the normal approximation with $z=2$. If current experience is 2,000 policy-years, calculate the full standard and the square-root partial credibility factor. Check the current standard error against the tolerance.

Method. For exposure $E$, let $N\sim\operatorname{Poisson}(\lambda E)$ and estimate the frequency by $\hat{\lambda}=N/E$. Since $\operatorname{Var}(N)=\lambda E$, $$ \operatorname{Var}(\hat{\lambda}) =\frac{\lambda E}{E^2} =\frac{\lambda}{E}. $$ The permitted absolute error is $0.10(0.05)=0.005$ claims per policy-year.

Full-credibility exposure. At the full standard, the tolerance equals two standard errors: $$ 0.005=2\sqrt{\frac{0.05}{E_0}}. $$ Therefore, $$ \sqrt{\frac{0.05}{E_0}}=0.0025,\qquad \frac{0.05}{E_0}=0.00000625,\qquad E_0=\frac{0.05}{0.00000625}=8{,}000\text{ policy-years}. $$ This corresponds to $0.05(8{,}000)=400$ expected claims.

Current credibility. The exposure ratio is $2{,}000/8{,}000=0.25$, so $$ Z=\sqrt{0.25}=0.50. $$ For a direct check, the current estimate has standard error $$ \sqrt{\frac{0.05}{2{,}000}}=\sqrt{0.000025}=0.005. $$ The tolerance is therefore one current standard error, rather than two. This agrees with $Z=0.50$: the current exposure provides half the full-standard number of standard errors relative to the same tolerance.

Check and common mistake. The full standard produces a tolerance of $2(0.0025)=0.005$, as required. A common error is to use $Z=E/E_0=0.25$. That uses the exposure ratio directly, although standard errors scale with the square root of inverse exposure. The corrected factor is $0.50$.

Worked Example: Credibility for an Average Loss

A hypothetical portfolio has independent annual loss observations with mean $\$120$ and standard deviation $\$240$. Use a normal model and a two-standard-error criterion for an absolute error no greater than $\$24$. If 100 observations are available, find the full-credibility standard and the partial credibility factor. State the approximate probability that the current sample mean is within $\$24$ of its mean.

Method. For a sample mean based on $n$ independent observations, $$ \operatorname{Var}(\bar{X})=\frac{240^2}{n}=\frac{57{,}600}{n}. $$ At the full standard, two standard errors must fit within the $\$24$ tolerance: $$ 2\sqrt{\frac{57{,}600}{n_0}}=24. $$ Thus the full-standard standard error is $24/2=\$12$, and $$ \frac{57{,}600}{n_0}=12^2=144,\qquad n_0=\frac{57{,}600}{144}=400. $$

Partial credibility. With $n=100$, $$ Z=\sqrt{\frac{100}{400}}=\sqrt{0.25}=0.50. $$ The current sample mean has variance $57{,}600/100=576$ and standard error $\sqrt{576}=\$24$. Under the stated normal model, the probability of being within $\$24$ is $P(|N(0,1)|\leq 1)$, approximately $0.6827$.

Check and common mistake. At $n_0=400$, the standard error is $\sqrt{57{,}600/400}=\$12$, so $\$24$ is exactly two standard errors. At $n=100$, the tolerance is one standard error, giving the stated probability. The credibility factor $0.50$ is not that probability: it is a volume-based factor, while $0.6827$ is the approximate probability of meeting this particular tolerance.

Worked Example: Recovering the Full Standard From a Variance Model

For a hypothetical estimator $T_n$, suppose $E[T_n]=\theta$ and $\operatorname{Var}(T_n)=2{,}500/n$, where $n$ is the number of independent observations. A normal approximation is appropriate. The required criterion is a two-sided 95.45% confidence level, using $z=2$, with absolute error at most 10 units. Find $n_0$, then calculate the partial credibility factor at $n=64$. Verify the result using the standard errors.

Method. At full credibility, the tolerance of 10 must equal two standard errors: $$ 10=2\sqrt{\frac{2{,}500}{n_0}}. $$ Dividing by two gives a full-standard standard error of 5. Therefore, $$ \frac{2{,}500}{n_0}=5^2=25,\qquad n_0=\frac{2{,}500}{25}=100. $$ At $n=64$, the square-root factor is $$ Z=\sqrt{\frac{64}{100}}=\sqrt{0.64}=0.80. $$

Check. The current standard error is $\sqrt{2{,}500/64}=\sqrt{39.0625}=6.25$. The ratio of the full-standard standard error to the current standard error is $5/6.25=0.80$, confirming the factor. The current tolerance of 10 is $10/6.25=1.6$ standard errors, which is also $2Z=1.6$.

Common mistake. Using the variance ratio $64/100=0.64$ as the credibility factor confuses variance scaling with standard-deviation scaling. The variance ratio is $0.64$, but the standard-deviation ratio—and hence the square-root factor—is $0.80$.

Interpreting the Rule Correctly

A useful check is to compare the current and full-standard error thresholds. If $n$ is one-quarter of $n_0$, then $Z=0.5$, and the current standard error is twice the full-standard standard error. If $n$ is one-half of $n_0$, then $Z=\sqrt{0.5}$, not $0.5$. These comparisons help catch arithmetic errors without recomputing the entire standard.

The credibility factor should not be confused with a probability or with a statement that the estimate is unbiased. It summarizes the amount of experience relative to a particular full-credibility standard. Changing the confidence level or the permitted error changes $n_0$, and therefore changes $Z$, even if the current exposure remains fixed. A different model or variance structure may require a different relationship between experience and precision.

The main exam trap is to apply the exposure ratio linearly. The variance is inversely proportional to exposure, but the factor is based on standard deviation, the square root of variance. Always identify the variance model first, determine the full standard from the stated precision criterion, and only then take the square root of the experience ratio.

Key takeaway: When estimator variance is proportional to $1/n$, standard error is proportional to $1/\sqrt{n}$. Comparing current experience $n$ with full-standard experience $n_0$ therefore gives the limited-fluctuation partial credibility factor $Z=\sqrt{n/n_0}$ for experience below the full standard.

Check Your Understanding

For each problem, identify the relevant volume ratio and use the square-root rule. Assume the stated normal criteria are appropriate.

A full-credibility standard is 900 independent observations. Current experience is 225 observations. Find the partial credibility factor. (Answer: $Z=0.50$.)
An estimator has variance $144/n$. The criterion is a two-standard-error tolerance of 0.8 units. Find the full standard and $Z$ when $n=225$. (Answer: $n_0=900$, $Z=0.50$.)
A Poisson claim frequency is $0.04$ per policy-year. The target is 20% relative error at a two-sided 95% confidence level, using $z=1.96$. Find the full exposure and $Z$ at 600.25 policy-years. (Answer: $E_0=2{,}401$ policy-years, $Z=0.50$.)
An estimator has variance $2{,}500/n$, and the required confidence criterion uses $z=2$ with an absolute tolerance of 10. Find the full standard and the factor at $n=64$. (Answer: $n_0=100$, $Z=0.80$.)
A sample mean has variance $900/n$. Under a two-standard-error criterion, the full standard is determined by an absolute tolerance of 6. If current experience is 100 observations, find the full standard and the partial factor. (Answer: $n_0=100$, $Z=1.00$.)

What You'll Learn