Introduction
A claim count is not described adequately by its average alone. Two portfolios can have the same expected number of claims but different exposure structures, possible count ranges, and degrees of variation. Choosing a frequency model therefore starts with the mechanism that generates claims, then uses features such as support and the mean-variance relationship as checks.
The binomial, Poisson, and negative binomial models are standard choices for claim counts, but they represent different structures. The binomial describes a fixed number of independent opportunities with a common claim probability. The Poisson describes events accumulating over exposure at a constant rate. The negative binomial can describe counts that vary more than a Poisson count, including variation arising from unobserved differences in risk levels. This selection framework is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.
Three Models and Their Structural Clues
The equal-probability condition in the binomial definition matters. Independence and a maximum of one claim per opportunity are not sufficient by themselves. If independent opportunities have different claim probabilities, their total is generally Poisson-binomial, not binomial. Similarly, a fixed maximum count can rule out an exact Poisson or negative binomial model, since both assign positive probability to every nonnegative integer.
For a Poisson claim count, \(\lambda\) is the expected count for the stated exposure. A constant rate per exposure unit gives \(\lambda\) proportional to exposure. The model is most plausible when events occur independently and the rate is reasonably homogeneous over the exposure being modeled. If risks have materially different rates or the rate changes over time, a single homogeneous Poisson model may not describe the aggregate count well.
The negative binomial’s dispersion parameter \(r\) controls its excess variation: as \(r\) increases, \(\mu^2/r\) decreases, and the variance approaches the Poisson variance \(\mu\). Smaller \(r\) means greater overdispersion. The variance-to-mean ratio is \(1+\mu/r\), which exceeds one for finite \(r\). This pattern can be consistent with heterogeneity, but it does not prove heterogeneity or uniquely identify a negative binomial distribution.
A Justification for the Negative Binomial Variance
One useful way to understand the negative binomial is as a Poisson count whose rate varies across risks. Suppose, conditionally on a risk-specific rate \(\Lambda\), the count satisfies \(N\mid\Lambda\sim\operatorname{Poisson}(\Lambda)\). Let \(\Lambda\) have a gamma distribution with mean \(\mu\) and variance \(\mu^2/r\), where \(r>0\). The conditional Poisson mean and variance are both \(\Lambda\). Applying the laws of total expectation and total variance gives:
The first variance term, \(\mu\), is the average Poisson variation within a risk level. The second, \(\mu^2/r\), is the additional variation from differences in rates among risks. Integrating over the gamma-distributed rates gives the negative binomial count model. This derivation explains why the negative binomial is a natural candidate when a Poisson model understates variation and latent rate heterogeneity is plausible; it does not establish that every overdispersed data set is negative binomial.
A Practical Model-Selection Sequence
Ask whether there is a fixed number of opportunities, a continuing event process over exposure, or a population of risks whose underlying rates may differ.
A known maximum count and identical independent Bernoulli opportunities support a binomial model. Unequal opportunity probabilities generally produce a Poisson-binomial count instead. A count without a fixed upper bound is compatible with Poisson or negative binomial support.
For a Poisson model the variance equals the mean; for the negative binomial it exceeds the mean. For a binomial model the variance is below the mean when \(0<p<1\), because \(np(1-p)<np\).
Overdispersion may point toward heterogeneous rates, clustering, or another departure from the assumed process. It is a reason to investigate a negative binomial model, not a rule that selects it automatically.
These checks complement one another. Support can rule out a model on structural grounds, while sample mean and variance provide evidence rather than certainty. In particular, finite samples have variable empirical moments; a sample variance slightly above or below its sample mean does not by itself settle the model choice.
Worked Example: Fixed, Identical Claim Opportunities
A hypothetical device warranty covers eight identical units for one year. Assume each unit independently generates either zero or one claim during that year, and each unit has the same claim probability \(p=0.08\). Find the appropriate model, its mean and variance, and the probability of no claims.
Method and model. There are exactly \(n=8\) opportunities; they are independent, have the same probability \(p=0.08\), and each allows at most one claim. Therefore \(N\sim\operatorname{Bin}(8,0.08)\).
Mean and variance. \(E[N]=np=8(0.08)=0.64\). Also, \(\operatorname{Var}(N)=np(1-p)=8(0.08)(0.92)=0.5888\). The variance is below the mean, as expected for a non-degenerate binomial count.
Probability of no claims. With \(N=0\), all eight units must have no claim: \(P(N=0)=\binom{8}{0}(0.08)^0(0.92)^8=(0.92)^8=0.513218\ldots\). Thus the probability is \(0.5132\) to four decimal places.
Check and trap avoided. The largest possible count is eight, consistent with the binomial support. A Poisson model would assign positive probability to counts above eight. The common mistake is to call any count of independent, one-claim opportunities binomial while ignoring whether the probabilities are equal; here the common probability has been explicitly specified.
Worked Example: A Homogeneous Poisson Exposure Model
For a hypothetical portfolio, assume claims occur independently at a constant rate of \(1.5\) claims per policy-year. Find the model and parameter for three policy-years of exposure, then calculate the probability of exactly two claims.
Method and model. Under the assumed constant rate and independent event process, the count over exposure \(e\) is Poisson with parameter equal to rate times exposure. Here \(e=3\) policy-years, so \(\lambda=(1.5)(3)=4.5\), and \(N\sim\operatorname{Poisson}(4.5)\).
Mean and variance. Both equal \(\lambda\), so \(E[N]=4.5\) claims and \(\operatorname{Var}(N)=4.5\) claims squared.
Probability of two claims. \(P(N=2)=e^{-4.5}4.5^2/2!\). Since \(4.5^2=20.25\) and \(2!=2\), this is \(10.125e^{-4.5}=0.1125\) to four decimal places.
Check and trap avoided. The expected count scales with exposure: three policy-years produce a parameter of \(4.5\), not \(1.5\). The calculation uses exposure in policy-years, matching the units of the stated rate. A fixed number of policies alone does not make this binomial: the model here allows multiple events over exposure and is specified as a Poisson process.
Worked Example: Overdispersion and a Negative Binomial Candidate
A hypothetical group of insured risks has an estimated mean annual claim count of \(\mu=2\) and an estimated variance of \(5\). Assume a gamma-Poisson mixture is a suitable candidate model. Find \(r\) in the mean-dispersion parameterization and compare its probability of no claims with the Poisson model having the same mean.
Method and dispersion parameter. The negative binomial variance is \(\mu+\mu^2/r\). Substituting the stated moments gives \(5=2+2^2/r\), so \(3=4/r\), and \(r=4/3\).
Probability of no claims. In the gamma-Poisson mixture parameterization with mean \(\mu\) and shape \(r\), the probability of zero is \(P(N=0)=(r/(r+\mu))^r\). Therefore, \(P(N=0)=((4/3)/(4/3+2))^{4/3}=((4/3)/(10/3))^{4/3}=0.4^{4/3}=0.2947\) to four decimal places.
For a Poisson model with the same mean, the corresponding probability is \(e^{-2}=0.1353\) to four decimal places. The negative binomial candidate has greater variance, \(2+4/(4/3)=5\), matching the stated estimate.
Check and trap avoided. The fitted \(r\) is positive, and substitution recovers the given variance. The common error is to set the negative binomial variance equal to \(\mu^2/r\) and omit the Poisson component \(\mu\). The observed mean and variance support examining an overdispersed model, but the assumed mixture structure is an additional modeling choice, not a conclusion forced by two sample statistics.
Common Selection Errors
- Using “independent” as a complete binomial justification. Each opportunity must also have the same claim probability, in addition to allowing at most one claim. If independent opportunities have probabilities \(p_1,\ldots,p_n\) that are not all equal, the count is generally Poisson-binomial.
- Choosing a model from the mean alone. A mean does not reveal whether the count is capped, event-driven over exposure, or affected by varying risk levels.
- Treating overdispersion as proof. A variance greater than the mean is inconsistent with the exact Poisson mean-variance relationship at the population level, but it can have several causes. It motivates further model assessment rather than automatically proving a negative binomial mechanism.
- Ignoring units and exposure. A Poisson rate has units such as claims per policy-year. Multiply by exposure in the matching unit before using it as \(\lambda\).
- Using support without context. Although Poisson and negative binomial counts are unbounded, the relevant question is whether the proposed data-generating process can genuinely produce counts beyond any stated maximum.
In practice, begin with what one risk can generate and how exposure is accumulated. Then check whether the implied support and mean-variance pattern are consistent with the data. This keeps a convenient formula from substituting for a defensible frequency model.
Check Your Understanding
For each problem, select or characterize the model using the stated assumptions, and check the requested numerical result.
- A hypothetical contract has ten independent, identical opportunities, each with claim probability \(0.03\), and each can produce at most one claim. Give the binomial model, mean, and variance. (Answer: \(N\sim\operatorname{Bin}(10,0.03)\), mean \(0.30\), variance \(0.291\).)
- A hypothetical event process has a constant rate of \(0.4\) claims per month over five months. Under a Poisson model, find the parameter and probability of exactly one claim. (Answer: \(\lambda=2\), \(P(N=1)=2e^{-2}=0.2707\).)
- A count model has mean \(3\) and variance \(6\), and a negative binomial mean-dispersion model is assumed. Find \(r\). (Answer: \(r=3\), since \(6=3+9/r\).)
- Three independent opportunities have respective claim probabilities \(0.10\), \(0.20\), and \(0.30\), with at most one claim per opportunity. Identify the count model type and calculate its mean and variance. (Answer: generally Poisson-binomial, not binomial; mean \(0.60\), variance \(0.10(0.90)+0.20(0.80)+0.30(0.70)=0.46\).)
- A count is known to be at most four because exactly four independent, identical trials each allow at most one claim. The estimated mean is \(1.2\) and variance is \(0.72\). Which standard model best matches the stated structure, and what is \(p\)? (Answer: binomial with \(n=4\); \(p=1.2/4=0.30\), giving variance \(4(0.30)(0.70)=0.84\). The stated variance \(0.72\) does not match these binomial parameters, so the moments require investigation rather than changing the model without reconciling the trial assumptions.)