Confidence And Tolerance In Credibility Standards

Why Confidence and Tolerance Are Design Choices

A credibility standard turns a goal for stability into a target amount of experience. Two choices define that goal: how often the result should fall within a specified range and how wide that range may be. A higher confidence level demands greater assurance. A smaller allowable error demands a closer result. Either choice can require substantially more experience.

In a limited-fluctuation standard, these are design criteria, not conclusions that the observed experience is free of bias or fully representative. The standard describes random fluctuation under a specified model. Its interpretation depends on the claim measure, exposure basis, and assumptions used in the calculation.

Definition: The confidence level \(p\) is the target probability that the aggregate result lies within the chosen bounds around its expected value. The allowable relative error \(k\) is the permitted distance from that expected value, expressed as a proportion. For example, \(k=0.10\) means a range of 10% below to 10% above the expected value. A two-sided confidence level \(p\) leaves total probability \(1-p\) outside the bounds, split equally between the two tails under a symmetric normal approximation.

Let \(z_p\) be the positive standard-normal critical value satisfying \(P(-z_p \leq Z \leq z_p)=p\) for \(Z\sim N(0,1)\). Common approximate values are \(z_{0.90}=1.645\) and \(z_{0.95}=1.96\). The critical value is not the confidence level itself: a 90% standard uses 1.645, not 0.90, in the normal-approximation formula.

From the Design Choices to an Experience Target

Consider first a claim-count standard. Assume claim counts over the selected exposure base follow a Poisson model, with independent claim occurrences and a constant expected frequency \(q\) per exposure unit. If there are \(E\) exposure units, the expected count is \(M=Eq\), and the variance of the count is also \(M\). For a sufficiently large expected count, approximate the count by a normal distribution. The same relative-error criterion applies to aggregate loss if every claim has the same fixed payment, because aggregate loss is then proportional to claim count.

P\left(\left|\frac{N-M}{M}\right|\leq k\right)\approx p \quad\Longrightarrow\quad k \approx z_p\frac{\sqrt{M}}{M} =\frac{z_p}{\sqrt{M}}.

Solving for the expected claim count gives \(M\approx z_p^2/k^2\). Since \(M=Eq\), the approximate exposure target is:

E\approx\frac{z_p^2}{k^2q}.

This relationship explains the central sensitivity: the target is proportional to \(z_p^2\), proportional to \(1/k^2\), and inversely proportional to the expected claim frequency \(q\). In particular, halving the allowable error multiplies the required experience by four, all else equal. Raising the confidence level also increases the target because it raises \(z_p\).

For variable claim severities, use a compound Poisson model: the claim count is Poisson with mean \(M\), individual payments are independent and identically distributed as \(X\), and claim count is independent of severities. Let aggregate loss be \(S=\sum_{j=1}^{N}X_j\). Under these assumptions, \(E[S]=M E[X]\) and \(\operatorname{Var}(S)=M E[X^2]\). Define the severity coefficient of variation by \(\operatorname{CV}_X=\sqrt{\operatorname{Var}(X)}/E[X]\). Since \(E[X^2]/E[X]^2=1+\operatorname{CV}_X^2\), the approximate relative standard deviation of aggregate loss is \(\sqrt{1+\operatorname{CV}_X^2}/\sqrt{M}\). Therefore:

M\approx\frac{z_p^2(1+\operatorname{CV}_X^2)}{k^2}, \qquad E\approx\frac{z_p^2(1+\operatorname{CV}_X^2)}{k^2q}.

This severity adjustment is relevant only when the stated model and inputs are appropriate. The normal approximation can be unreliable when expected counts are small or the loss distribution is highly skewed or dominated by a few large claims. A numerical standard does not remove those limitations.

Worked Example: A 90% Standard for Claim Counts

A hypothetical small-equipment portfolio has an expected claim frequency of \(q=0.06\) claims per exposure unit. The actuary wants a two-sided 90% confidence standard with allowable relative error \(k=10\%\). Assume a Poisson claim count and use \(z_{0.90}=1.645\). Estimate the exposure target.

Method. Use the claim-count limited-fluctuation formula. The calculation assumes the Poisson model is an adequate approximation and does not include variable severity.

First calculate the expected claim count target:

M\approx\left(\frac{z_p}{k}\right)^2 =\left(\frac{1.645}{0.10}\right)^2 =(16.45)^2 =270.6025.

Convert expected claims to exposure units using \(M=Eq\):

E\approx\frac{270.6025}{0.06} =4510.0417.

Thus the approximate planning target is 4,510 exposure units, rounded to the nearest whole exposure unit. As a check, 4,510 units at the stated frequency correspond to \(4510(0.06)=270.6\) expected claims, essentially the calculated target of 270.6025. This is a planning approximation, not a claim that the observed count will always be within 10% of its mean.

Common mistake avoided: dividing by the confidence level, 0.90, instead of using its critical value, 1.645, would not implement the stated normal-approximation standard.

What Tightening a Standard Costs

The cleanest way to compare choices is to hold the model and exposure basis fixed. If only the allowable error changes from \(k_1\) to \(k_2\), the ratio of exposure targets is \((k_1/k_2)^2\). If only confidence changes, the ratio is \((z_{p_2}/z_{p_1})^2\). These comparisons separate the effect of the design choice from changes in frequency or severity.

Worked Example: Comparing Confidence Levels and Error Tolerances

A hypothetical repair program has a Poisson claim count with constant payment per claim and expected frequency \(q=0.20\) per exposure. Compare the approximate exposure target for a 95% confidence level and 5% allowable error with the target for a 90% confidence level and 10% allowable error. Use \(z_{0.95}=1.96\) and \(z_{0.90}=1.645\).

Method. Apply \(E\approx z_p^2/(k^2q)\) separately to each standard.

For 95% confidence and 5% error:

E_{95,5\%}\approx \frac{1.96^2}{0.05^2(0.20)} =\frac{3.8416}{0.0025(0.20)} =\frac{3.8416}{0.0005} =7683.2.

The approximate target is 7,683 exposure units, rounded to the nearest whole unit. For 90% confidence and 10% error:

E_{90,10\%}\approx \frac{1.645^2}{0.10^2(0.20)} =\frac{2.706025}{0.01(0.20)} =\frac{2.706025}{0.002} =1353.0125.

The approximate target is 1,353 exposure units. The ratio is \(7683.2/1353.0125\approx5.68\). This is also obtained from the design choices alone:

\left(\frac{1.96}{1.645}\right)^2 \left(\frac{0.10}{0.05}\right)^2 \approx 1.42(4) \approx 5.68.

So the combined move to higher confidence and half the allowable error requires about 5.68 times as much experience under the model. The common trap is to attribute the whole increase only to the stricter confidence level; the halved tolerance contributes a factor of four by itself.

Severity Variability Also Affects Stability

A count standard and an aggregate-loss standard need not require the same amount of experience. If claim payments vary, two groups with the same expected number of claims can have different aggregate-loss fluctuation. Under the compound Poisson assumptions stated above, the factor \(1+\operatorname{CV}_X^2\) captures this effect. A larger severity coefficient of variation raises the expected claim count and exposure targets for the same \(p\) and \(k\).

Worked Example: Allowing for Variable Claim Payments

A hypothetical equipment coverage has expected frequency \(q=0.06\) per exposure. Individual claim payments have \(\operatorname{CV}_X=1\). Estimate the exposure target for 90% confidence and 10% allowable error under the compound Poisson model. Use \(z_{0.90}=1.645\).

Method. Use the severity-adjusted standard. The independence, identical severity distribution, and Poisson frequency assumptions are material; the calculation is not a general correction for any possible loss distribution.

First, \(1+\operatorname{CV}_X^2=1+1^2=2\). The expected claim target is:

M\approx \frac{1.645^2(2)}{0.10^2} =\frac{2.706025(2)}{0.01} =\frac{5.41205}{0.01} =541.205.

Convert this to exposures:

E\approx\frac{541.205}{0.06} =9020.0833.

The approximate planning target is 9,020 exposure units, rounded to the nearest whole unit. As a check, this is about twice the target for the same confidence, tolerance, and frequency when claim payments are constant: the severity multiplier is \(1+\operatorname{CV}_X^2=2\). The common mistake is to use the count-only formula for aggregate loss even though the payments vary.

Interpretation and Exam Traps

A full-credibility target describes the scale at which random variation is expected to meet a chosen standard. It does not establish that the experience is representative of the future, that the underlying rate is adequate, or that data problems have been corrected. The model and the definition of the claim measure must match the intended use.

Do not confuse confidence with the critical value. Choose \(z_p\) for the stated central probability before substituting into the formula.
Keep the tolerance as a proportion. A 10% allowable error is \(k=0.10\), not \(k=10\).
Do not treat the target as linear in tolerance. The formula contains \(1/k^2\), so changing tolerance has a squared effect.
Match frequency units to exposure units. If \(q\) is claims per policy-year, the result is policy-years, not necessarily distinct policies.
Check the severity assumption. A count standard does not automatically describe the stability of aggregate dollars when claim payments vary.

Key takeaway: Under a limited-fluctuation standard, higher confidence and smaller allowable error require more stable experience. For a Poisson count, the approximate exposure target is \(z_p^2/(k^2q)\); under a compound Poisson aggregate-loss model, severity variability adds the factor \(1+\operatorname{CV}_X^2\). These are model-based planning standards, not guarantees about a particular observed result.

Check Your Understanding

Use the normal-approximation limited-fluctuation formulas above. Treat each result as an approximate planning target and round exposure counts to the nearest whole unit.

A hypothetical class has Poisson claim frequency \(q=0.10\) per exposure. For 90% confidence and \(k=0.08\), estimate the exposure target. Use \(z_{0.90}=1.645\). (Answer: 4,228 exposure units.)
A hypothetical policy group has Poisson frequency \(q=0.04\). Estimate the target for 95% confidence and \(k=0.12\), using \(z_{0.95}=1.96\). (Answer: 6,669 exposure units.)
A compound Poisson portfolio has frequency \(q=0.05\), severity coefficient of variation \(0.5\), and a 90% confidence standard with \(k=0.10\). Use \(z_{0.90}=1.645\). (Answer: 6,765 exposure units.)
For a fixed Poisson frequency and confidence level, an analyst tightens the allowable error from 10% to 5%. By what factor does the approximate exposure target change? (Answer: It increases by a factor of 4.)
A Poisson claim-count model has \(q=0.08\) claims per exposure and \(E=5{,}000\) exposures. What is the largest approximate relative-error tolerance for a 95% central confidence standard, using \(z_{0.95}=1.96\)? (Answer: \(k=0.098\), or 9.8%.)

What You'll Learn