Why a Claim-Count Threshold Depends on Confidence and Tolerance
A claim-count credibility standard asks whether a count is sufficiently large that random fluctuation is likely to stay within a chosen relative distance of its expected value. A higher confidence requirement demands a larger count. A tighter tolerance does too. Under the Poisson model, these requirements combine in a simple threshold formula.
This tutorial focuses on using the Poisson mean-variance equality to express that threshold. The result is a benchmark for the expected number of claims under the model. It is not a prediction that the realized count will equal the threshold, nor a conversion of that count into a number of policies or exposure units.
Here, \(0<p<1\) is a two-sided central confidence level, and \(r>0\) is a relative tolerance expressed as a decimal. For example, a 6% tolerance means \(r=0.06\), so the allowed deviation from the expected count is 6% of that expected count. The target event is \( |N-\lambda|\leq r\lambda \).
The standard claim-count threshold uses a normal approximation to the Poisson count. This approximation is the basis for the familiar limited-fluctuation formula; it is not an exact calculation of a discrete Poisson probability. The model also presumes that the count in the period and population being analyzed is appropriately represented by a Poisson distribution. If counts are materially overdispersed relative to Poisson, the mean-variance equality used here does not hold and this threshold is not justified by the stated model.
Deriving the Poisson Count Threshold
For a Poisson count, the standard deviation is \(\sqrt{\lambda}\), because the variance is \(\lambda\). Standardizing the count therefore expresses the allowed deviation in units of its standard deviation. The derivation below shows why a relative tolerance leads to a threshold proportional to the square of a normal quantile divided by the tolerance.
The count meets the tolerance when \( |N-\lambda|\leq r\lambda \). Dividing by the Poisson standard deviation \(\sqrt{\lambda}\) gives \( \left|(N-\lambda)/\sqrt{\lambda}\right|\leq r\sqrt{\lambda} \).
For the standard full-credibility count criterion, approximate \( (N-\lambda)/\sqrt{\lambda} \) by a standard normal variable \(Z\). Thus the probability of meeting the tolerance is approximated by \(P(|Z|\leq r\sqrt{\lambda})\).
Define \(z_p=\Phi^{-1}((1+p)/2)\), where \(\Phi\) is the standard normal cumulative distribution function. The central interval from \(-z_p\) to \(z_p\) has probability \(p\). To make the approximate probability at least \(p\), require \(r\sqrt{\lambda}\geq z_p\).
Since \(r>0\) and \(\lambda>0\), square both sides to obtain the minimum Poisson mean under this criterion.
The formula follows directly from the Poisson mean-variance equality: the tolerance measured in standard deviations is \(r\lambda/\sqrt{\lambda}=r\sqrt{\lambda}\). Equating that quantity to the required normal quantile and squaring produces the threshold. The formula applies to a two-sided criterion around \(\lambda\); a one-sided probability uses a different quantile and describes a different criterion.
Applying the Threshold
A calculation has four parts: identify the central confidence level, obtain its two-sided normal quantile, express the relative tolerance as a decimal, and compare the proposed expected count with the resulting threshold. Quantiles commonly used in calculations include \(z_p\approx1.645\) for 90% central confidence, \(z_p\approx1.960\) for 95%, and \(z_p\approx2.576\) for 99%.
Worked Example: Checking a 6% Tolerance
A hypothetical book of small-equipment coverage is modeled as having a Poisson claim count. The expected count is 900 claims for the defined period. Determine whether it meets a 90% central confidence standard with a 6% relative tolerance. Use \(z_{0.90}=1.645\).
Method. Apply the Poisson count full-credibility threshold. The stipulated distribution is \(N\sim\operatorname{Poisson}(\lambda)\), with \(\lambda\) equal to the expected count being assessed. The chosen relative tolerance is \(r=0.06\).
First calculate the ratio of the quantile to the tolerance: \(1.645/0.06=27.416666\ldots\). Then square it: \(27.416666\ldots^2=751.673611\ldots\). Thus the threshold is approximately 751.67 expected claims.
The proposed expected count is 900, and \(900\geq751.673611\ldots\). It meets the standard under the stated normal approximation. As a check, the allowed deviation at \(\lambda=900\) is \(0.06(900)=54\), the standard deviation is \(\sqrt{900}=30\), and the allowed deviation is \(54/30=1.8\) standard deviations. Since \(1.8>1.645\), the count meets the quantile requirement.
Common mistake avoided. The threshold is approximately 751.67, not 751.17. The comparison is with the expected count, not with a claim count that has already been observed.
Worked Example: A 95% Standard With a 10% Tolerance
A hypothetical regional warranty program has an expected count of 400 claims for a specified period. Assess it against a 95% central confidence standard with a 10% relative tolerance. Use \(z_{0.95}=1.960\).
Method. Use \(N\sim\operatorname{Poisson}(\lambda)\), so the mean and variance are both \(\lambda\). Set \(r=0.10\) and calculate the count threshold: \(1.960/0.10=19.6\), and \(19.6^2=384.16\). The required expected count is 384.16.
The proposed mean count is 400, so \(400\geq384.16\) and it meets the standard. A direct check gives an allowed deviation of \(0.10(400)=40\), a standard deviation of \(\sqrt{400}=20\), and a standardized allowance of \(40/20=2.00\). Because \(2.00\geq1.960\), the criterion is satisfied.
For contrast, an expected count of 360 would not meet the threshold, since \(360<384.16\). At 360, the standardized allowance would be \(0.10\sqrt{360}\approx1.897\), below 1.960.
Common mistake avoided. The 10% tolerance is \(0.10\), not 10. Substituting 10 into the formula would make the threshold far too small and would not represent a 10% relative deviation.
Worked Example: A High-Confidence, Tight-Tolerance Standard
A hypothetical specialty coverage segment is modeled with a Poisson claim count and has an expected count of 4,200. Determine whether it meets a 99% central confidence standard with a 4% relative tolerance. Use \(z_{0.99}=2.576\).
Method. The model is \(N\sim\operatorname{Poisson}(\lambda)\); set \(p=0.99\) and \(r=0.04\). The quantile-to-tolerance ratio is \(2.576/0.04=64.4\). Squaring gives \(64.4^2=4,147.36\). The required expected count is therefore 4,147.36.
Since \(4,200\geq4,147.36\), the segment meets the standard, though with a relatively small margin. Check the comparison in standardized units: the tolerance allows \(0.04(4,200)=168\) claims of deviation, while the Poisson standard deviation is \(\sqrt{4,200}\approx64.8074\). Thus the allowed deviation is \(168/64.8074\approx2.5923\) standard deviations, which exceeds 2.576.
Common mistake avoided. Do not use the 95% quantile just because it is familiar. A 99% central confidence level requires the larger quantile 2.576, which increases the threshold.
Interpreting the Result and Avoiding Traps
The threshold combines confidence and tolerance in a transparent way. Holding \(r\) fixed, increasing the confidence level increases \(z_p\), and therefore increases the required count. Holding the confidence level fixed, decreasing the tolerance increases \(z_p/r\) and hence increases the threshold. For example, halving the tolerance multiplies the threshold by four, because the tolerance is squared in the denominator.
The result is a condition on \(\lambda\), the Poisson mean. It does not assert that every realization of \(N\) lies within the tolerance, and meeting the standard does not eliminate random variation. It says that, under the model and normal approximation, the probability of lying within the specified band is at least the selected central confidence level.
A common wrong calculation for a 95% central standard is to use the one-sided 95th-percentile quantile, 1.645, rather than the central 95% quantile, 1.960. With a 10% tolerance, that incorrect choice gives \((1.645/0.10)^2=270.6025\). The correction is to use \(1.960\), giving \((1.960/0.10)^2=384.16\). The one-sided quantile would be appropriate only for a one-sided criterion, not for the two-sided fluctuation band defined here.
Finally, avoid rounding intermediate values prematurely. Keep the quantile-to-tolerance ratio at full calculator precision, square it, and round the threshold only at the end. The resulting threshold may be non-integer because it is a required value of the Poisson mean; it need not itself be a possible realized claim count.
Check Your Understanding
For each problem, use the two-sided Poisson count standard and round the threshold to two decimal places. Use \(z_{0.90}=1.645\), \(z_{0.95}=1.960\), and \(z_{0.99}=2.576\), as applicable.
- A hypothetical inland marine segment has a Poisson expected count of 900. Does it meet a 90% central confidence standard with a 6% relative tolerance? Find the threshold before comparing. (Answer: threshold \(751.67\); yes, because \(900\geq751.67\).)
- Find the required expected count for a 95% central confidence standard with an 8% relative tolerance. (Answer: \(600.25\) expected claims.)
- Find the threshold for a 99% central confidence standard with a 5% relative tolerance. (Answer: \(2,654.31\) expected claims.)
- A hypothetical service contract portfolio has an expected Poisson count of 1,700. Does it meet a 90% central confidence standard with a 4% tolerance? (Answer: threshold \(1,691.27\); yes, narrowly.)
- Using a 95% central confidence level and a 7.5% tolerance, calculate the threshold. Show the decimal tolerance used in the formula. (Answer: use \(r=0.075\); threshold \(682.95\) expected claims.)