Deriving Claim Count Standards from Probability Bounds

Why Derive a Count Standard From a Probability Bound?

A claim count varies from one observation period to another, even when the underlying claim frequency is stable. A full-credibility standard asks how large the expected count must be so that the observed count is unlikely to differ too much from its expectation. A probability bound provides a direct, conservative answer without requiring an exact calculation of Poisson probabilities.

This tutorial uses Chebyshev’s inequality. The result is a sufficient standard: meeting it guarantees at least the selected confidence under the stated model, but failing it does not prove that the actual probability is too low. The bound can be conservative because it uses only the mean and variance, not the full shape of the Poisson distribution.

Definition: Let $N$ be the number of claims in a fixed observation period, with $N\sim\operatorname{Poisson}(\lambda)$, where $\lambda$ is the expected claim count. For confidence level $p$ and relative tolerance $r>0$, the count is within tolerance when $$ |N-\lambda|<r\lambda. $$ A bound-based full-credibility standard is a value of $\lambda$ large enough to guarantee $P(|N-\lambda|<r\lambda)\ge p$ using the selected probability bound.

Assumptions and the Poisson Model

The model assumes a fixed observation period and a stable claim-generating process for which the total claim count is Poisson. In particular, the count has mean $\lambda$ and variance $\lambda$. The derivation below needs these two moments; it does not require claim severities, premium, or a mortality model. No interest rate or discounting assumption enters this frequency calculation.

The relative tolerance $r$ is measured as a fraction of the expected count. For example, $r=0.10$ means that the deviation is judged against 10% of $\lambda$, not against 10 claims and not against the standard deviation. The confidence level $p$ is the required lower bound on the probability of being within that tolerance.

A Poisson model is not appropriate merely because the observations are counts. If the claim process is materially overdispersed, clustered, changing over time, or otherwise inconsistent with the assumed model, this standard does not automatically apply. The calculation also does not establish that the underlying expected frequency is correct; it assesses random fluctuation around the model’s $\lambda$.

Deriving the Chebyshev Claim-Count Standard

Chebyshev’s inequality states that a random variable with finite mean $\mu$ and variance $\sigma^2$ satisfies $P(|X-\mu|\ge a)\le \sigma^2/a^2$ for $a>0$. Apply it to the Poisson count $N$, taking $\mu=\lambda$, $\sigma^2=\lambda$, and $a=r\lambda$.

P(|N-\lambda|\ge r\lambda) \le \frac{\operatorname{Var}(N)}{(r\lambda)^2} = \frac{\lambda}{r^2\lambda^2} = \frac{1}{r^2\lambda}.

The event of being outside the tolerance is bounded above by $1/(r^2\lambda)$. Taking its complement gives a lower bound on the probability of being within tolerance:

P(|N-\lambda|<r\lambda) = 1-P(|N-\lambda|\ge r\lambda) \ge 1-\frac{1}{r^2\lambda}.

To guarantee confidence of at least $p$, it is sufficient that the right-hand side be at least $p$. Rearranging gives the standard:

Theorem: For a Poisson claim count with mean and variance $\lambda$, Chebyshev’s inequality guarantees $P(|N-\lambda|<r\lambda)\ge p$ whenever $$ \lambda\ge \frac{1}{(1-p)r^2}. $$ This is a sufficient bound-based standard, not a claim that the exact Poisson probability equals $p$, nor necessarily the smallest count that would meet the exact probability requirement.

The derivation can also be seen directly from the squared deviation. On the event $|N-\lambda|\ge a$, the squared deviation $(N-\lambda)^2$ is at least $a^2$. Therefore $$ E[(N-\lambda)^2]\ge a^2P(|N-\lambda|\ge a), $$ so $P(|N-\lambda|\ge a)\le \operatorname{Var}(N)/a^2$. Substituting $a=r\lambda$ and $\operatorname{Var}(N)=\lambda$ yields the result above. The bound becomes more demanding as the desired confidence rises or the allowed relative deviation shrinks.

Applying the Standard

Worked Example: Required Count at 90% Confidence

A hypothetical roadside-assistance program models its claim count in a fixed observation period as $N\sim\operatorname{Poisson}(\lambda)$. Determine the expected count sufficient to guarantee at least 90% confidence that the observed count is within 10% of its mean. Use the Chebyshev bound.

Method and assumptions. The count is Poisson with mean and variance $\lambda$; there is no interest or mortality assumption. Set $p=0.90$ and $r=0.10$. The event of interest is $|N-\lambda|<0.10\lambda$.

Substitute. The required expected count satisfies $$ \lambda\ge \frac{1}{(1-p)r^2} =\frac{1}{(1-0.90)(0.10)^2} =\frac{1}{(0.10)(0.01)} =\frac{1}{0.001} =1000. $$

Check. At $\lambda=1000$, the bound on the probability of being outside tolerance is $$ P(|N-\lambda|\ge 100)\le \frac{1}{(0.10)^2(1000)} =\frac{1}{10} =0.10. $$ Thus the probability of being within tolerance is at least $1-0.10=0.90$, as required. The result is an expected count threshold; it does not assert that every observed count will lie within 10% of the mean.

Common mistake avoided. The tolerance is 10% of $\lambda$, so the deviation allowance is $0.10\lambda$. Treating it as a fixed allowance of 10 claims would change the event being bounded.

Worked Example: A Tighter Standard Requires More Count

A hypothetical specialty insurer wants a Chebyshev-based standard with 95% confidence and a 5% relative tolerance for a Poisson claim count. Find the required expected count and verify the bound.

Method and assumptions. Assume $N\sim\operatorname{Poisson}(\lambda)$, so $E[N]=\operatorname{Var}(N)=\lambda$. No discounting or mortality assumptions are involved. Use $p=0.95$ and $r=0.05$.

Calculate. $$ \lambda\ge \frac{1}{(1-0.95)(0.05)^2} =\frac{1}{(0.05)(0.0025)} =\frac{1}{0.000125} =8000. $$

Check. At $\lambda=8000$, the allowed deviation is $0.05(8000)=400$ claims. The probability bound for being outside that range is $$ P(|N-8000|\ge 400) \le \frac{8000}{400^2} =\frac{8000}{160000} =0.05. $$ Therefore the probability of being within 400 claims of the mean is at least 0.95.

Common mistake avoided. The confidence shortfall is $1-p=0.05$, while the tolerance is squared as $r^2=(0.05)^2$. Substituting $p$ where $1-p$ belongs would make the standard far too small.

Worked Example: Test a Proposed Count Against Two Confidence Targets

For a hypothetical product, assume the claim count in the observation period follows $N\sim\operatorname{Poisson}(1200)$. A reviewer requires a 10% relative tolerance. Determine the confidence lower bound from Chebyshev’s inequality, then assess whether the count meets standards of 90% and 95%.

Method and assumptions. Here $\lambda=1200$, $r=0.10$, and the mean and variance are both 1200. There is no interest or mortality assumption.

Find the bound. $$ P(|N-1200|\ge 0.10(1200)) \le \frac{1}{(0.10)^2(1200)} =\frac{1}{12} \approx 0.0833. $$ Thus, $$ P(|N-1200|<120)\ge 1-\frac{1}{12} =\frac{11}{12} \approx 0.9167. $$

Compare with the requirements. Since $0.9167\ge0.90$, the count meets the 90% bound-based standard. Since $0.9167<0.95$, this bound does not certify the 95% requirement. To see the latter directly, the 95% standard is $$ \lambda\ge\frac{1}{(0.05)(0.10)^2} =\frac{1}{0.0005} =2000, $$ which exceeds 1200.

Check and interpretation. The 90% standard is $\lambda\ge1000$, so 1200 passes it. The 95% standard is $\lambda\ge2000$, so 1200 does not pass it. Failure to pass the Chebyshev standard does not prove that the actual Poisson probability is below 95%; it means only that this particular bound is insufficient to guarantee 95%.

Common mistake avoided. A lower bound of 91.67% is not the exact probability. It supports the 90% guarantee, but it cannot be reported as the exact chance that the count falls in the specified interval.

Worked Example: Find the Tolerance for a Fixed Expected Count

A hypothetical portfolio has a Poisson claim-count mean of $\lambda=2500$. What is the smallest relative tolerance for which the Chebyshev bound guarantees at least 90% confidence?

Method and assumptions. Assume $N\sim\operatorname{Poisson}(2500)$, with mean and variance 2500. Set $p=0.90$. Starting from $\lambda\ge1/((1-p)r^2)$, solve for $r$. No interest or mortality assumption is needed.

Rearrange and calculate. $$ r^2\ge \frac{1}{(1-p)\lambda} =\frac{1}{(0.10)(2500)} =\frac{1}{250} =0.004, $$ so $$ r\ge\sqrt{0.004}\approx0.06325. $$ Expressed as a percentage, the smallest tolerance certified by this bound is approximately $6.325\%$.

Check. Using $r=0.06325$ to the shown precision gives $(1-p)r^2\lambda\approx(0.10)(0.004)(2500)=1$, so the bound guarantees confidence of at least 90%. A larger tolerance is easier to meet; a smaller tolerance is not certified by this bound.

Common mistake avoided. Solving for tolerance requires taking the square root after isolating $r^2$. Reporting $0.004$, or 0.4%, as the tolerance confuses the squared tolerance with the tolerance itself.

Interpreting and Using the Result

The formula $$ \lambda\ge \frac{1}{(1-p)r^2} $$ makes the effect of each design choice explicit. Holding tolerance fixed, increasing $p$ decreases $1-p$ and raises the required count. Holding confidence fixed, reducing $r$ raises the required count in inverse proportion to $r^2$. For example, halving the tolerance multiplies the bound-based count by four.

This standard is deliberately conservative. Chebyshev’s inequality applies broadly to random variables with finite variance, so it does not use the full probability shape of a Poisson count. A more distribution-specific calculation may give a different, often lower, threshold, but that is a different method. The Chebyshev result should be described as a guarantee from a bound, not as an exact probability calculation or as the unique possible credibility standard.

A common exam trap is to confuse three separate quantities: the confidence $p$, the probability left outside the tolerance $1-p$, and the relative tolerance $r$. The bound controls the outside probability, which is why $1-p$ appears in the denominator. Another trap is to use a deviation of $r\sqrt{\lambda}$; the stated relative tolerance is $r\lambda$. Applying the bound to that deviation produces the $r^2\lambda$ denominator.

The result is aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus. Its proper use depends on stating the Poisson assumption, defining the tolerance relative to the expected count, and identifying the standard as sufficient rather than necessary.

Check Your Understanding

For each problem, use the Poisson mean-variance assumption and the Chebyshev bound unless stated otherwise. Treat each count as hypothetical; no interest or mortality assumptions apply.

A hypothetical warranty program requires 80% confidence within 10% of its Poisson expected claim count. Find the minimum expected count under the bound. (Answer: $\lambda\ge500$ claims.)
A hypothetical fleet account seeks 95% confidence within 10% of its expected count. Calculate the bound-based standard and state how it compares with a proposed mean of 1,800 claims. (Answer: $\lambda\ge2{,}000$; 1,800 does not meet the Chebyshev standard.)
A Poisson claim count has mean 400. For a 10% tolerance, calculate the lower confidence bound and determine whether it certifies a 70% requirement. (Answer: $P(|N-400|<40)\ge0.75$; yes, it certifies 70%.)
A hypothetical portfolio has Poisson mean 1,600. Find the smallest relative tolerance certified at 90% confidence, as a percentage to four decimal places. (Answer: $r\ge7.9057\%$.)
A proposed count standard requires 90% confidence within 20% of the Poisson mean. Find the minimum expected count, then calculate the confidence lower bound if the actual model mean is 900. (Answer: required $\lambda\ge250$; at $\lambda=900$, confidence is at least $1-1/(0.20^2\cdot900)=0.9722$, so the bound exceeds 90%.)

What You'll Learn