Full Credibility as a Probability Standard

What Full Credibility Means

A claim measure observed from limited experience varies from one sample to another. Full credibility expresses a probability standard for that variation: the experience is sufficient when the observed measure has a specified probability of being close enough to its expected value. The standard describes performance over repeated samples under a model; it does not promise that every particular observation will be close to the expectation.

For example, an insurer might assess whether an observed average claim index is likely to remain within a stated number of dollars of its expected value. The tolerance and the required probability must be specified before deciding whether the experience is sufficient. Both are part of the standard, not facts determined by the data after it has been observed.

Definition: Let $T_n$ be an observed claim measure based on $n$ units of experience, and let its expected value under the stated model be $\theta=E[T_n]$. For a specified tolerance $c>0$ and required probability $p$, the experience meets the full-credibility standard if $$ P\bigl(|T_n-\theta|\le c\bigr)\ge p. $$ The probability is evaluated under the assumed model for $T_n$. The tolerance $c$ is expressed in the same units as $T_n$, and $p$ is a probability between 0 and 1.

The absolute value makes this a two-sided condition: the observed measure may be too high or too low, and both departures count. The expectation $\theta$ is the model benchmark, not necessarily the value seen in a particular sample. A high probability of meeting the tolerance is a statement about the distribution of possible observations, not a guarantee about a realized observation.

Evaluating the Standard Under a Normal Model

A normal model gives a direct way to evaluate the probability. Suppose $X_1,\ldots,X_n$ are independent and identically distributed observations from a normal distribution with mean $\theta$ and variance $\sigma^2$, and the claim measure is their average, $\bar X$. Then $\bar X$ has mean $\theta$ and standard deviation $\sigma/\sqrt{n}$. These assumptions matter: the calculation below is not automatically valid for dependent observations, changing means, or a nonnormal model.

\bar X\sim N\left(\theta,\frac{\sigma^2}{n}\right), \qquad P\bigl(|\bar X-\theta|\le c\bigr) =2\Phi\left(\frac{c\sqrt n}{\sigma}\right)-1

Here $\Phi(z)$ is the cumulative distribution function of a standard normal random variable. The formula compares the probability actually provided by the model with the required probability $p$. It does not say that the observed average is exactly equal to $\theta$, or that the probability can be assessed without specifying a distribution.

Justification: Under the stated assumptions, standardizing the sample mean gives $Z=(\bar X-\theta)/(\sigma/\sqrt n)\sim N(0,1)$. The event $|\bar X-\theta|\le c$ is equivalent to $-c\sqrt n/\sigma\le Z\le c\sqrt n/\sigma$. By symmetry of the standard normal distribution, its probability is $\Phi(c\sqrt n/\sigma)-\Phi(-c\sqrt n/\sigma)=2\Phi(c\sqrt n/\sigma)-1$.

Worked Example: A Probability Just Below the Required Level

A hypothetical portfolio has independent, normally distributed claim-index observations with mean $\theta$ and standard deviation $\$25$. Its observed measure is the average of $n$ observations. Determine whether $n=24$ meets a standard requiring the average to be within $\$10$ of $\theta$ with probability at least $0.95$. If not, check whether $n=25$ meets it.

Method. Use the normal distribution of the sample average and compare its two-sided probability with $0.95$. The standard deviation of each observation is $\sigma=25$, and the tolerance is $c=10$.

For $n=24$, the standard deviation of the average is $$ \frac{25}{\sqrt{24}} \approx 5.1031036. $$ The standardized tolerance is $$ z=\frac{10}{25/\sqrt{24}}=\frac{10\sqrt{24}}{25}\approx 1.9595918. $$ Therefore, $$ P\bigl(|\bar X-\theta|\le 10\bigr) =2\Phi(1.9595918)-1 \approx 0.9499565. $$

This probability is just below $0.95$, so $n=24$ does not meet the standard. For $n=25$, the standard deviation of the average is $25/\sqrt{25}=5$, and the standardized tolerance is $10/5=2$. Thus $$ P\bigl(|\bar X-\theta|\le 10\bigr) =2\Phi(2)-1\approx 0.9545, $$ so $n=25$ does meet the standard. For this normal model, the probability increases with $n$; the two checks show that the smallest qualifying integer among these candidate experience levels is 25.

Check. The tolerance and standard deviation are both in dollars, so their ratio is dimensionless. The decision uses the unrounded probability at $n=24$: although $0.9499565$ rounds to $0.9500$ at four decimal places, it remains less than $0.95$.

Common mistake avoided. Rounding the probability to four decimal places before comparing it with the requirement would make $n=24$ appear to qualify. Compare using sufficient precision first; round only when reporting the result.

Checking a Different Probability Requirement

A probability standard has two separate thresholds to respect: how far the measure may depart from its expectation, and how likely it must be to stay within that distance. A candidate experience level can satisfy one threshold but fail the other. The model determines the probability; the standard determines whether that probability is sufficient.

Worked Example: A Stricter Required Probability

A hypothetical estimate of an average repair-cost index is normally distributed with mean $\theta$ and standard deviation $\$4$. Decide whether it meets a standard requiring the estimate to be within $\$10$ of $\theta$ with probability at least $0.99$.

Method. Standardize the tolerance using the supplied standard deviation of the estimate. No additional sample-size calculation is needed because the distribution of the estimate is given directly.

The standardized tolerance is $z=10/4=2.5$. Therefore, the probability of being within the tolerance is $$ P\bigl(|T-\theta|\le 10\bigr) =2\Phi(2.5)-1 =2(0.9937903)-1 \approx 0.9876. $$ Since $0.9876<0.99$, the estimate does not meet the standard.

Check. The probability is high, but the required probability is higher. The result is therefore a failure to meet this particular standard, not a conclusion that the estimate is generally inaccurate.

Common mistake avoided. Treating a two-sided probability as if a standardized tolerance of $2.5$ automatically meant $99\%$ coverage would give the wrong decision. The probability must be calculated and compared with the stated threshold.

Exact Probability for a Discrete Measure

The definition does not require a normal distribution. When the claim measure has a discrete distribution, its probability can be found by listing the possible outcomes and adding the probabilities of those within the tolerance. This is especially useful when a normal approximation is not part of the assumptions or could obscure the exact probability.

Worked Example: A Discrete Claim Index

A hypothetical claim index is based on four independent observations. Each observation equals $\theta+\$12$ or $\theta-\$12$, each with probability $0.5$. Determine whether the average of the four observations meets a standard requiring it to be within $\$6$ of $\theta$ with probability at least $0.35$.

Method. Count the equally likely combinations that put the sample average inside the specified interval. For each observation, write its deviation from $\theta$ as either $+12$ or $-12$.

If $k$ of the four deviations are $+12$, then $4-k$ are $-12$. The sum of the deviations is $$ 12k-12(4-k)=24k-48, $$ so the average deviation is $(24k-48)/4=6k-12$. For $k=0,1,2,3,4$, the possible average deviations are respectively $-12,-6,0,6,12$ dollars.

Only the average deviation of $0$ is within $\$6$ of $\theta$ if the boundary is interpreted inclusively? In fact, the deviations $-6$, $0$, and $6$ all satisfy the condition $|\bar X-\theta|\le 6$. They occur when $k=1,2,3$, respectively. The number of qualifying sign patterns is $$ \binom{4}{1}+\binom{4}{2}+\binom{4}{3}=4+6+4=14. $$ There are $2^4=16$ equally likely patterns, so the probability is $14/16=0.875$. Since $0.875\ge 0.35$, the experience meets the standard.

Check. The boundary cases matter: both $-\$6$ and $+\$6$ are included by the stated “within $\$6$” condition. The exact probability is $0.875$, not merely the probability of an average equal to $\theta$.

Common mistake avoided. Counting only the four-observation outcomes whose average equals $\theta$ would omit the boundary outcomes. The definition uses an interval, not exact equality.

A Practical Decision Sequence

A reliable evaluation keeps the model, the event, and the comparison distinct. The following sequence applies whether the probability is calculated from a continuous distribution or exactly from a discrete one.

Identify the claim measure and its benchmark.
Name the random measure $T_n$ and its expectation $\theta$ under the assumed model.

Write the tolerance event.
Translate the stated requirement into $|T_n-\theta|\le c$, preserving the units and including the endpoints.

Evaluate the probability under the model.
Use a distributional formula only when its assumptions hold; otherwise use the specified distribution directly.

Compare before rounding.
The standard is met if and only if the calculated probability is at least the required probability $p$.

This distinction is important in actuarial work because a statement such as “the result is likely to be near its expectation” is incomplete. The tolerance, the probability level, the claim measure, and the probability model together determine what “likely” and “near” mean. A standard met under one model need not be met under a different model with a different variance or shape.

Common Interpretation Traps

Confusing a probability standard with a guarantee. A $0.95$ requirement allows a $0.05$ probability of falling outside the tolerance under the model.
Using the wrong center. The event is measured around the model expectation $\theta$, not automatically around a selected estimate or an observed value.
Dropping one side of the interval. The event $|T_n-\theta|\le c$ includes both upward and downward departures.
Applying a distribution without checking assumptions. The normal formula for an average requires the stated normal, independent, identically distributed observations. A different model requires its own probability calculation.
Rounding across the decision boundary. If the calculated probability is just below the requirement, a rounded display that looks equal does not change the comparison.

Key takeaway: Full credibility, in this probability-standard formulation, means that the probability of an observed claim measure lying within a specified tolerance of its expectation is at least the required level. State the model, calculate that probability, and compare it with the requirement without rounding first.

Check Your Understanding

For each problem, decide whether the stated experience meets the full-credibility standard. Use the supplied model and compare probabilities before rounding.

A hypothetical estimate $T$ is normally distributed with mean $\theta$ and standard deviation $3$. Is it fully credible under a tolerance of $6$ and a required probability of $0.95$? (Answer: Yes; $2\Phi(2)-1\approx0.9545$.)
A normally distributed estimate has mean $\theta$ and standard deviation $5$. The required tolerance is $10$, and the required probability is $0.96$. Does it qualify? (Answer: No; $2\Phi(2)-1\approx0.9545<0.96$.)
Independent normal observations have standard deviation $12$. Their average is based on $36$ observations. Test a tolerance of $4$ around the expected average with a required probability of $0.95$. (Answer: Yes; the average has standard deviation $12/\sqrt{36}=2$, so the probability is $2\Phi(2)-1\approx0.9545$.)
Two independent observations each equal $\theta+8$ or $\theta-8$, with equal probability. Their average must be within $4$ of $\theta$, with probability at least $0.60$. Does it qualify? (Answer: No; the four equally likely pairs produce average deviations $-8,0,0,8$, so the probability is $2/4=0.50<0.60$.)

What You'll Learn