The Limited-Fluctuation Criterion

What a Limited-Fluctuation Standard Requires

A credibility standard needs to say what counts as sufficiently stable experience. The limited-fluctuation criterion does this by setting a band around the expected value and requiring a specified probability that the observed experience falls inside that band. The band describes an acceptable relative deviation; the probability describes how often that requirement should be met under the assumed model.

For example, a hypothetical insurer might ask whether the observed claim count is likely to be within 20% of its expected count. That question is incomplete until the insurer also specifies a confidence level, such as 90%. A narrow band is harder to meet than a wide one, while a higher required probability is harder to meet than a lower one.

Definition: Let $X$ be the random quantity being assessed, such as a claim count or aggregate loss, and suppose its expected value $\mu=E[X]$ is finite and positive. For a chosen relative tolerance $r$, with $0<r<1$, and required probability $p$, the limited-fluctuation criterion is $$ P\left((1-r)\mu \leq X \leq (1+r)\mu\right)\geq p. $$ The quantity $r$ is the maximum permitted deviation as a fraction of the expected value, and $p$ is the required probability of being within that band.

The criterion concerns the random experience itself under an assumed probability model. It is not a guarantee that every future observation will be in the band, nor is it a confidence interval for an unknown parameter. The probability is calculated from the model for $X$. If the expected value is zero, a relative band around that mean is not useful; a different standard must be specified.

The band is centered on the expected value. If $r=0.15$, the acceptable values range from 85% to 115% of the mean. Because the criterion is relative, the same tolerance has a comparable interpretation for a small expected count and a large expected loss amount. The actual probability of meeting the criterion still depends on the distribution of $X$.

Writing the Criterion for Different Quantities

The limited-fluctuation criterion can be applied to different random quantities, but the quantity and its model must be identified clearly. A standard for claim counts is not automatically a standard for aggregate losses: losses vary both because the number of claims varies and because the amounts of the claims vary. The event in the criterion must match the quantity the standard is intended to assess.

For a nonnegative discrete quantity such as a claim count, the band endpoints need not be integers. The acceptable outcomes are the integers that lie inside the band. Thus, if the band is from 7.2 through 12.8, the acceptable integer outcomes are 8 through 12, inclusive. Rounding the endpoints to the nearest integer could incorrectly include outcomes outside the stated band.

For a continuous quantity, such as a modeled aggregate loss, probabilities at individual endpoints are usually zero. The probability within the band is then calculated over the interval. A distributional assumption is needed to evaluate that probability; the limited-fluctuation definition itself does not require a normal distribution.

Worked Example: A Band for a Discrete Claim Count

A hypothetical program models the number of claims in a period as $N\sim\operatorname{Binomial}(8,0.5)$. Determine the probability that the observed count is within 25% of its expected value, and assess whether the model meets a required probability of 70%.

Method and model. For a binomial count with 8 independent trials and claim probability 0.5 on each trial, $E[N]=8(0.5)=4$. The limited-fluctuation band with $r=0.25$ is $4(1-0.25)$ through $4(1+0.25)$, or 3 through 5.

Identify the acceptable integer outcomes. The endpoints happen to be integers, so the acceptable counts are $N=3,4,5$, inclusive. Using the binomial probability formula, $$ P(N=k)=\binom{8}{k}(0.5)^k(0.5)^{8-k}=\frac{\binom{8}{k}}{256}. $$

Calculate the probability. The three relevant probabilities are $P(N=3)=56/256$, $P(N=4)=70/256$, and $P(N=5)=56/256$. Therefore, $$ P(3\leq N\leq 5)=\frac{56+70+56}{256}=\frac{182}{256}=0.7109375. $$ Rounded to four decimal places, the probability is 0.7109.

Check the requirement. Since $0.7109\geq0.70$, the model meets the stated 70% criterion. The common mistake is to calculate the probability at just one endpoint or to treat the band as a strict inequality. The criterion includes both endpoints, and all three acceptable counts must be included.

Normal Models and Standardization

When the quantity $X$ is modeled as normal, the probability within the relative band can be calculated by standardizing. Assume $X\sim N(\mu,\sigma^2)$, where $\mu>0$ and $\sigma>0$. Here $\sigma$ is the standard deviation; the second parameter in the normal notation is the variance. Let $\Phi$ denote the standard normal cumulative distribution function.

Derivation: Under the normal model, $Z=(X-\mu)/\sigma$ has a standard normal distribution. Subtracting $\mu$ from both sides of the band event and dividing by the positive value $\sigma$ gives $$ P\left((1-r)\mu\leq X\leq(1+r)\mu\right) =P\left(-\frac{r\mu}{\sigma}\leq Z\leq\frac{r\mu}{\sigma}\right). $$ By symmetry of the standard normal distribution, this probability is $2\Phi(r\mu/\sigma)-1$. The formula holds under the stated normal assumption; it should not be used as an exact probability for a nonnormal quantity.

This derivation shows what determines the probability in the normal case: the tolerance measured in standard deviations, $r\mu/\sigma$. A larger tolerance or larger mean relative to the standard deviation makes the band more likely to contain the observation. To check a specified probability requirement, compare the calculated probability with $p$; equality meets the criterion.

Worked Example: Testing a Normal Aggregate-Loss Standard

A hypothetical portfolio models aggregate loss for a period as $X\sim N(800{,}000,(80{,}000)^2)$, with all amounts in dollars. The standard requires losses to be within 15% of the mean with probability at least 85%. Determine whether the model meets the requirement.

Method and assumptions. Use the normal limited-fluctuation probability, which applies here because the problem specifies a normal model with mean $\mu=800{,}000$ and standard deviation $\sigma=80{,}000$. The tolerance is $r=0.15$.

Calculate the band. The permitted deviation is $0.15(800{,}000)=120{,}000$. Thus, the lower endpoint is $800{,}000-120{,}000=680{,}000$, and the upper endpoint is $800{,}000+120{,}000=920{,}000$.

Standardize the deviation. The deviation $120{,}000$ is $120{,}000/80{,}000=1.5$ standard deviations. Therefore, $$ P(680{,}000\leq X\leq920{,}000)=P(-1.5\leq Z\leq1.5)=2\Phi(1.5)-1. $$ Using $\Phi(1.5)\approx0.9332$, the probability is $2(0.9332)-1=0.8664$.

Check the requirement. Since $0.8664\geq0.85$, the model meets the requirement. A common mistake is to compare the one-sided probability $\Phi(1.5)$ with 85%. The criterion is two-sided: it counts observations both below the lower endpoint and above the upper endpoint as outside the band.

Discrete Probabilities and Boundary Care

For a discrete model, the probability should be obtained by summing the probabilities of every integer outcome inside the band. The endpoints are determined by the original percentage band, not by rounding the mean or the tolerance first. This matters particularly when the expected count is small, because excluding or including a single count can materially change the result.

Worked Example: A Poisson Count Near the Required Probability

A hypothetical service unit models its period claim count as $N\sim\operatorname{Poisson}(4)$. The target is to be within 50% of the expected count with probability at least 80%. Calculate the probability and determine whether the target is met.

Method and assumptions. Under a Poisson model with parameter $\lambda=4$, $E[N]=4$ and $P(N=k)=e^{-4}4^k/k!$ for nonnegative integers $k$. The tolerance is $r=0.50$.

Find the acceptable counts. The band is $4(1-0.50)=2$ through $4(1+0.50)=6$. The acceptable outcomes are therefore $N=2,3,4,5,6$.

Sum their probabilities. The individual probabilities, rounded to six decimal places, are $P(N=2)=0.146525$, $P(N=3)=0.195367$, $P(N=4)=0.195367$, $P(N=5)=0.156293$, and $P(N=6)=0.104196$. Adding them gives $$ P(2\leq N\leq6)=0.146525+0.195367+0.195367+0.156293+0.104196=0.797748. $$ Rounded to four decimal places, the probability is 0.7977.

Check the requirement. Since $0.7977<0.80$, the target is not met. It is close, but rounding 0.797748 to 0.80 before comparing would incorrectly declare success. Compare using sufficient precision, then report the result at the requested precision.

Interpreting and Using the Criterion

A limited-fluctuation standard is a statement about modeled variation, not a statement that the observed experience has already been assigned a credibility weight. Earlier material on full and partial credibility distinguishes a full-credibility standard from the use of a partial-credibility factor. Here the key task is to state precisely what “full” means for the quantity and assumptions at hand: a chosen relative band and a chosen probability.

The standard is only as informative as its inputs. The analyst must specify the quantity being assessed, its expected value, the tolerance, the required probability, and the model used to calculate the probability. If the model does not fit the data or the assumptions are inappropriate, a precisely calculated probability may still be a poor description of actual fluctuation.

Keep tolerance and probability distinct. The tolerance is the permitted percentage deviation; the probability is the chance of staying within that deviation.
State whether the band is two-sided. The usual limited-fluctuation criterion counts deviations both below and above the mean.
Use the specified quantity. A criterion for claim count does not establish the same probability for aggregate loss.
Respect discreteness. For counts, include only integer values inside the unrounded band.
Match the calculation to the model. The normal standardization formula is not an exact result for a Poisson or other nonnormal quantity.

This framework is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus. In applications, a complete statement of a standard lets another actuary identify the event being tested and reproduce the probability calculation under the stated model.

Check Your Understanding

For each problem, identify the band before calculating or interpreting its probability. Report probabilities to four decimal places.

A hypothetical random outcome is $X\sim N(120,15^2)$. What is the probability that $X$ is within 12.5% of its mean? Use the unrounded standard normal cumulative distribution function in the calculation. (Answer: 0.6827)
A modeled aggregate loss is $X\sim N(2{,}000{,}000,(160{,}000)^2)$. Find the probability that it is within 12% of its mean, and state whether it meets a required probability of 85%. (Answer: Probability 0.8664; yes)
A hypothetical count has distribution $N\sim\operatorname{Binomial}(6,0.5)$. Find the probability that it is within one-third of its expected value. (Answer: 0.7813)
A hypothetical claim count has distribution $N\sim\operatorname{Poisson}(5)$. For a tolerance of 40%, calculate the probability of falling in the band and determine whether a 75% requirement is met. (Answer: 0.7420; no)
A hypothetical aggregate loss is $X\sim N(400{,}000,(50{,}000)^2)$. What is the largest symmetric relative tolerance that gives a 95% probability of being within the band? Use $z_{0.975}=1.96$. (Answer: 24.5%)

What You'll Learn