Limited-Fluctuation Standards for Binomial Counts

Why a Fixed-Exposure Count Has a Binomial Standard

A count of claims can arise in more than one way. In the binomial model, the number of exposure units is fixed, and each unit either produces a claim during the stated period or does not. This structure gives the count a variance smaller than its mean when the claim probability is below one. A limited-fluctuation standard can use that variance to estimate how many units are needed for the count to stay near its expected value with a specified probability.

The result is an approximate full-credibility standard: it identifies a minimum exposure size for a chosen relative tolerance and probability requirement under a normal approximation. The assumptions matter. This is not a universal standard for all claim counts, and the formula does not itself guarantee an exact binomial probability.

Definition: Suppose there are $n$ fixed exposure units. For each unit, a claim occurs with probability $q$, independently of the outcomes for all other units, with the same $q$ for every unit. The total number of claims $N$ then has a binomial distribution, $N\sim\operatorname{Bin}(n,q)$, with $$ E[N]=nq,\qquad \operatorname{Var}(N)=nq(1-q). $$

Here $n$ is the number of exposure units observed over the specified period, and $q$ is the probability that one unit has at least one claim in that period. For this tutorial, the model treats an exposure unit as having a binary outcome: claim or no claim. It does not model the number or dollar amount of multiple claims from one unit.

Let $k$ be the allowed relative deviation, expressed as a proportion, and let $p$ be the required probability. A two-sided limited-fluctuation requirement is

P\bigl(\lvert N-nq\rvert\leq k n q\bigr)\geq p.

The deviation is measured relative to the expected count $nq$, not relative to the number of exposure units $n$. The stated probability level determines the normal quantile used to approximate the count’s distribution.

Deriving the Two-Sided Standard

For sufficiently large $n$, a normal approximation to the binomial count may be used. Under that approximation, $N$ has mean $nq$ and standard deviation $\sqrt{nq(1-q)}$. For a central probability $p$, let $z_{(1+p)/2}$ denote the positive standard-normal quantile with probability $(1+p)/2$ to its left. The approximate requirement is that the allowed deviation be at least this many standard deviations.

k n q \geq z_{(1+p)/2}\sqrt{nq(1-q)}.

Both sides are nonnegative. Squaring the inequality and dividing by $k^2nq$, which is positive for $n>0$, $q>0$, and $k>0$, gives

k^2 n^2 q^2 \geq z_{(1+p)/2}^2nq(1-q) \quad\Longrightarrow\quad n\geq \frac{z_{(1+p)/2}^2(1-q)}{k^2q}.

This derivation shows why the standard depends on $q$ and $1-q$: they enter through the binomial variance. It also shows why a stricter relative tolerance (smaller $k$) requires more exposure, and why a higher probability requirement (larger quantile) requires more exposure. The formula is for the specified binomial model and normal approximation; it is not an exact binomial-tail calculation.

Theorem: Under the independent, equal-probability binomial model, the approximate two-sided full-credibility standard for relative tolerance $k$ and central probability $p$ is $$ n_{\mathrm{full}} = \left\lceil \frac{z_{(1+p)/2}^2(1-q)}{k^2q} \right\rceil. $$ The ceiling means the smallest whole number of exposure units at least as large as the calculated bound.

The ceiling is important: rounding to the nearest integer can leave the selected number below the requirement. When using a rounded value of $z$, keep enough precision in the calculation and still round the final exposure requirement upward.

Worked Applications

Worked Example: A Two-Sided Standard for a Small Claim Probability

A hypothetical warranty program observes a fixed number $n$ of independent devices for one year. Each device has probability $q=0.20$ of generating a claim during that year. Find the approximate full-credibility standard for the claim count to be within 15% of its expected value with 90% central probability. Use $z_{0.95}=1.645$.

Method and assumptions. Each device has one binary outcome for the year, and the outcomes are independent with common probability $0.20$. Thus $N\sim\operatorname{Bin}(n,0.20)$, $k=0.15$, and the two-sided central probability uses $z_{(1+0.90)/2}=z_{0.95}=1.645$.

Substitute into the two-sided standard:

n\geq \frac{1.645^2(1-0.20)}{0.15^2(0.20)} = \frac{2.706025(0.80)}{0.0225(0.20)} = \frac{2.16482}{0.0045} = 481.071111\ldots.

The required number of devices is therefore $n=482$, after rounding upward. At 482 devices, the expected count is $482(0.20)=96.4$, and the variance is $482(0.20)(0.80)=77.12$. The standard deviation is $\sqrt{77.12}\approx8.7818$, and the allowed deviation is $0.15(96.4)=14.46$. The normal threshold is $1.645(8.7818)\approx14.446$, which is below the allowed deviation. The fractional boundary is 481.071111…, so 481 units do not meet the approximate bound: $481(0.0045)=2.1645<2.16482$.

Common mistake avoided. Rounding 481.071111… to 481 would round to the nearest integer, not to a sufficient exposure size. The standard requires rounding up to 482.

Worked Example: A 95% Standard With a Different Claim Probability

A hypothetical group policy covers a fixed set of independent contracts for one year. Each contract has probability $q=0.25$ of at least one claim. Find the approximate number of contracts needed for the claim count to be within 20% of its expected value with 95% central probability. Use $z_{0.975}=1.96$.

Method and assumptions. The one-year outcomes are independent Bernoulli trials with common probability $0.25$, so the total count is binomial. Here $k=0.20$, and a 95% central requirement uses $z_{(1+0.95)/2}=z_{0.975}=1.96$.

Calculate the lower bound:

n\geq \frac{1.96^2(1-0.25)}{0.20^2(0.25)} = \frac{3.8416(0.75)}{0.04(0.25)} = \frac{2.8812}{0.01} = 288.12.

Rounding upward gives $n=289$ contracts. At 289 contracts, the expected count is $289(0.25)=72.25$. The variance is $289(0.25)(0.75)=54.1875$, so the standard deviation is $\sqrt{54.1875}\approx7.3619$. The allowed deviation is $0.20(72.25)=14.45$, while the normal threshold is $1.96(7.3619)\approx14.4293$. Thus the rounded-up exposure size meets the approximate inequality.

Common mistake avoided. The variance is $nq(1-q)$, not $nq$. Using a Poisson variance here would discard the fixed-unit binomial assumption and produce a different standard.

Worked Example: A One-Sided Count Requirement

A hypothetical monitoring program follows a fixed number of independent installations for a year. Each installation has probability $q=0.10$ of a claim. Find the approximate number needed so that the count does not exceed its expected value by more than 25%, with at least 95% probability. Use the one-sided quantile $z_{0.95}=1.645$.

Method and assumptions. The count $N$ is binomial because the number of installations is fixed, each has a common claim probability, and outcomes are independent. The condition is one-sided: only an excess above the expected count is restricted. Under the normal approximation, a one-sided 95% requirement uses $z_{0.95}=1.645$, rather than the two-sided 95% quantile $z_{0.975}=1.96$.

The one-sided inequality is $k nq\geq z_{0.95}\sqrt{nq(1-q)}$. Solving as in the two-sided derivation gives

n\geq \frac{1.645^2(1-0.10)}{0.25^2(0.10)} = \frac{2.706025(0.90)}{0.0625(0.10)} = \frac{2.4354225}{0.00625} = 389.6676.

The required whole-number exposure is $n=390$. Its expected count is $390(0.10)=39$, its variance is $390(0.10)(0.90)=35.1$, and its standard deviation is $\sqrt{35.1}\approx5.9245$. The allowed excess is $0.25(39)=9.75$; the one-sided normal threshold is $1.645(5.9245)\approx9.746$. The inequality is met. The calculation addresses an upper deviation only; it does not impose the same bound on unusually low counts.

Common mistake avoided. A one-sided 95% requirement does not use the two-sided 95% quantile. Using 1.96 would impose a different, more conservative probability criterion.

Interpreting and Applying the Standard

The standard answers a specific question: how many independent, equally exposed units are needed for the count to satisfy a relative-fluctuation condition under the stated probability approximation? It does not directly specify a credible claim frequency or an expected loss amount. The count $N$ is the random quantity in the condition, and its reference value is $nq$.

The assumptions should be checked before using the formula:

Fixed exposure count: $n$ is set rather than random.
Binary outcome per unit: each unit contributes either zero or one to this count in the period.
Common probability: every unit has the same claim probability $q$.
Independence: one unit’s claim outcome does not change another’s probability or outcome.
Normal approximation: the standard is based on approximating the binomial distribution by a normal distribution.

If units have different probabilities or are dependent, the count need not have the stated binomial distribution. If a unit can generate multiple claims and the count includes all of them, a binomial model is not appropriate for that total. Likewise, a count over several years can be treated this way only if the defined exposure units have the required common probabilities and independence. The period is part of the model: changing it can change $q$.

For probabilities near zero or one, or for a small expected count, the normal approximation may be poor. The formula remains a useful approximate standard when its assumptions and approximation are reasonable, but it should not be described as an exact binomial guarantee. If an exact guarantee is required, evaluate the relevant binomial probabilities directly for candidate integer values of $n$.

Key takeaway: For a fixed number of independent units with common claim probability $q$, start with $E[N]=nq$ and $\operatorname{Var}(N)=nq(1-q)$. Apply the appropriate normal quantile to the specified deviation condition, solve for $n$, and round upward. Use a one-sided or two-sided quantile to match the actual requirement.

This binomial standard is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus. Its central exam skill is matching the probability condition to the right distribution, variance, quantile, and rounding rule.

Check Your Understanding

For each problem, assume independent units with the stated common claim probability and use the normal-approximation standard. Round the required exposure upward to a whole unit.

A fixed portfolio has $q=0.10$. Find the two-sided standard for 20% relative precision with 90% central probability, using $z_{0.95}=1.645$. (Answer: 609 units)
A hypothetical service contract group has $q=0.40$. Find the two-sided standard for 10% relative precision with 95% central probability, using $z_{0.975}=1.96$. (Answer: 577 units)
A monitoring program has $q=0.05$. Find the one-sided standard for an upper deviation of at most 30% of the expected count with 99% probability, using $z_{0.99}=2.326$. (Answer: 1,143 units)
For fixed independent units with $q=0.20$, find the two-sided standard for 15% relative precision with 95% central probability, using $z_{0.975}=1.96$. (Answer: 683 units)
A hypothetical pool has $q=0.60$. Find the two-sided standard for 5% relative precision with 90% central probability, using $z_{0.95}=1.645$. (Answer: 722 units)

What You'll Learn