Why the Claim-Count Model Matters
Before estimating the size of losses, an actuary often needs to describe how many claims may occur in a specified period or across a specified set of exposures. The appropriate count model depends on how claims can arise. A single yes-or-no opportunity suggests a Bernoulli model; a fixed number of independent opportunities suggests a binomial model; and events occurring over a period at a stable average rate are often represented by a Poisson model.
The distinction is more than terminology. Each model has different assumptions, probability calculations, and implications for the mean and variance of the count. This tutorial treats the number of claims only; it does not model the dollar amount of any claim. No interest or mortality assumption is needed for these count calculations.
Bernoulli Counts: One Opportunity
A Bernoulli count records whether a single specified opportunity produces a claim. Let \(I=1\) if a claim occurs and \(I=0\) otherwise, with \(P(I=1)=p\). The possible counts are only zero and one. The model is appropriate only when there is one relevant binary opportunity; it does not allow two claims from that opportunity.
The mean is \(p\) because \(E[I]=1p+0(1-p)\). Also \(I^2=I\), so \(E[I^2]=p\). Therefore \(\operatorname{Var}(I)=E[I^2]-(E[I])^2=p-p^2=p(1-p)\).
Worked Example: One Annual Claim Opportunity
A hypothetical specialist policy can produce at most one covered claim during a particular annual review period. Based on the model assumption, the probability of a claim is \(p=0.08\). Find the probabilities of zero and one claim, and calculate the mean and variance of the annual count.
Method and assumptions. There is one binary opportunity, and the claim probability for that opportunity is \(0.08\). Thus the count \(I\) is Bernoulli with parameter \(p=0.08\).
The probability of one claim is \(P(I=1)=p=0.08\). The probability of no claim is \(P(I=0)=1-p=1-0.08=0.92\). These sum to \(0.08+0.92=1.00\).
The mean count is \(E[I]=p=0.08\) claims per annual period. The variance is \(\operatorname{Var}(I)=p(1-p)=0.08(0.92)=0.0736\) claims squared. As a check, \(E[I^2]=0.08\), so \(E[I^2]-(E[I])^2=0.08-(0.08)^2=0.0736\).
Common mistake avoided: Treating this as a model that permits two or more claims would contradict the stated one-opportunity assumption. A Bernoulli count has only the outcomes zero and one.
Binomial Counts: A Fixed Number of Opportunities
Suppose there are exactly \(n\) opportunities, each of which either produces one claim or does not. If the opportunities are independent and have the same claim probability \(p\), the total count \(N\) has a binomial distribution with parameters \(n\) and \(p\), written \(N\sim\operatorname{Binomial}(n,p)\). The fixed number of opportunities and the common probability matter: a binomial model does not automatically fit merely because the count is a whole number.
The factor \(\binom{n}{k}\) counts the ways to select which \(k\) opportunities produce claims. Independence gives probability \(p^k(1-p)^{n-k}\) for any one such arrangement. Since the arrangements are disjoint, multiplying by their number gives the probability of exactly \(k\) claims.
For justification, represent the count as \(N=I_1+\cdots+I_n\), where each \(I_j\) is a Bernoulli indicator with probability \(p\). Linearity of expectation gives \(E[N]=\sum_{j=1}^{n}E[I_j]=np\). Independence makes the covariances zero, so \(\operatorname{Var}(N)=\sum_{j=1}^{n}\operatorname{Var}(I_j)=np(1-p)\). If opportunities are dependent or have different probabilities, these binomial formulas do not generally apply.
Worked Example: Claims Across Six Independent Exposures
A hypothetical portfolio contains six comparable one-period exposures. For this model, assume each exposure produces either zero or one claim, the claim probability for each is \(p=0.10\), and outcomes are independent. Find the probability of exactly two claims, the probability of at most one claim, and the mean and variance of the total count.
Method and assumptions. There are \(n=6\) fixed, independent opportunities with common claim probability \(p=0.10\). Therefore \(N\sim\operatorname{Binomial}(6,0.10)\).
For exactly two claims, use \(k=2\):
For at most one claim, add the probabilities of zero and one claim:
The mean count is \(E[N]=np=6(0.10)=0.60\). The variance is \(\operatorname{Var}(N)=np(1-p)=6(0.10)(0.90)=0.54\). As a check on the at-most-one result, \(P(N\geq2)=1-0.885735=0.114265\); the probabilities of at least two and at most one sum to one.
Common mistake avoided: The probability of at most one claim includes both zero and one claim. Using only \(P(N=1)\) would omit the no-claim outcome.
Poisson Counts: Events Over Exposure
A Poisson model is commonly used for a count of events over a defined time period, area, or other exposure when the process can be represented by a stable average rate. If \(N\) is Poisson with parameter \(\lambda>0\), then \(\lambda\) is the expected number of events in the specified exposure. The model permits any nonnegative integer count; unlike the binomial model, it has no fixed upper bound.
The parameter must be stated for the exposure period being modeled. If a hypothetical process has a constant rate of \(r\) events per unit of exposure and the period contains \(t\) units, then its Poisson parameter is \(\lambda=rt\). This rate adjustment is valid when the same rate applies across the specified exposure; using a rate measured per year directly for a shorter period would give the wrong parameter.
A Poisson model is most defensible when events arise independently over the exposure, the rate is stable, and the process does not impose a fixed number of opportunities. The equality of mean and variance is a useful implication and diagnostic, not a guarantee that every observed count process is Poisson. Strong clustering, changing rates, or a fixed opportunity structure can make the model unsuitable.
Worked Example: Quarterly Event Counts
A hypothetical service program models the number \(N\) of reported incidents in a quarter as Poisson with parameter \(\lambda=2\). Calculate the probability of exactly one incident and the probability of at least three incidents. Give probabilities to four decimal places.
Method and assumptions. The count for this quarter is assumed to follow a Poisson distribution with mean \(\lambda=2\) incidents per quarter. Use the probability mass function for exactly one incident and the complement of zero, one, and two incidents for the tail.
For exactly one incident:
which is \(0.2707\) to four decimal places.
For at least three incidents, retain the common exponential factor while summing the probabilities of zero, one, and two:
Thus \(P(N\geq3)=0.3233\) to four decimal places. The mean and variance are both \(2\) incidents and \(2\) incidents squared, respectively. As a check, the complement \(P(N\leq2)=5e^{-2}=0.6766764162\ldots\), and the two complementary probabilities sum to one.
Common mistake avoided: Do not subtract separately rounded probabilities for zero, one, and two incidents if reporting more precision. Their rounded values can introduce a small discrepancy. The exact tail expression is \(1-5e^{-2}\); rounding it only at the end gives \(0.3233\).
Selecting and Comparing the Models
Model selection begins with the process that generates the count, not with a preference for a convenient formula. A Bernoulli model is the one-trial special case of the binomial model. A binomial model has a fixed number \(n\) of independent opportunities and cannot produce more than \(n\) claims. A Poisson model represents events over exposure and has no fixed upper bound.
| Model | Generating structure | Mean | Variance | Key restriction |
|---|---|---|---|---|
| Bernoulli | One binary opportunity | \(p\) | \(p(1-p)\) | Count is only zero or one |
| Binomial | \(n\) independent binary opportunities, common probability \(p\) | \(np\) | \(np(1-p)\) | Count cannot exceed \(n\) |
| Poisson | Events over a defined exposure at a modeled rate | \(\lambda\) | \(\lambda\) | Assumes a suitable stable-rate event process |
The Poisson distribution can approximate a binomial distribution when \(n\) is large, \(p\) is small, and \(\lambda=np\) is held as the approximate mean. This is an approximation, not an identity. For an important tail probability or a small number of opportunities, the exact binomial calculation may be preferable if its assumptions are appropriate.
Worked Example: Checking a Poisson Approximation
A hypothetical system has 100 independent components, each with probability \(p=0.02\) of producing one incident in a period. Find the exact binomial probability of no incidents and compare it with the Poisson approximation having the same mean.
Method and assumptions. The exact model is binomial because there are \(n=100\) fixed independent components, each with common incident probability \(0.02\). Its mean is \(np=100(0.02)=2\). For the approximation, use a Poisson model with \(\lambda=2\).
The exact probability of no incidents is:
or \(0.1326\) to four decimal places.
Under the Poisson approximation:
or \(0.1353\) to four decimal places. The absolute difference using the unrounded values is \(0.1353352832\ldots-0.1326195559\ldots=0.0027157273\ldots\), which is \(0.0027\) to four decimal places. The approximation is close here, but it is not exact.
Common mistake avoided: Matching the means does not make the distributions identical. The exact binomial model retains the fixed limit of 100 opportunities, while the Poisson approximation does not.
Common Exam Traps
- Ignoring the number of opportunities. If at most one claim can arise from each of \(n\) policies, a binomial model may fit under independence and equal probabilities. A Poisson model has no such upper bound.
- Using the wrong exposure unit. A rate per month and a rate per year imply different parameters for the same calendar period. Convert the rate to the modeled exposure before calculating probabilities.
- Confusing “exactly” and “at least.” \(P(N=k)\) is one outcome; \(P(N\geq k)\) includes every count from \(k\) upward. For an infinite tail, a complement is often efficient.
- Assuming independence without a process reason. The binomial variance formula relies on independent opportunities. Dependence changes the variance and can invalidate the binomial model.
- Rounding too early. Retain exact expressions or adequate precision through intermediate steps, then round the requested final probability. This is especially important when a tail is calculated by subtracting several probabilities from one.
Check Your Understanding
For each problem, identify the model from the stated assumptions before calculating. Round probabilities to four decimal places.
- A hypothetical coverage has one claim opportunity in the period, with claim probability \(0.07\). Find the probability of no claim and the mean and variance of the count. (Answer: \(P(N=0)=0.9300\); mean \(=0.0700\); variance \(=0.0651\).)
- Eight independent exposures each have claim probability \(0.10\), with at most one claim per exposure. Find the probability of exactly two claims and the probability of at least one claim. (Answer: \(P(N=2)=0.1488\); \(P(N\geq1)=0.5695\).)
- A hypothetical count is Poisson with mean \(3\) per defined period. Find the probability of zero events and the probability of at least two events. (Answer: \(P(N=0)=0.0498\); \(P(N\geq2)=1-4e^{-3}=0.8009\).)
- A hypothetical system has 12 independent units, each with probability \(0.05\) of one incident during the period. Find the probability of at least one incident and the mean and variance of the count. (Answer: \(P(N\geq1)=1-(0.95)^{12}=0.4596\); mean \(=0.6000\); variance \(=0.5700\).)
- A process is modeled as Poisson at a constant rate of \(0.4\) events per month. For a three-month period, find the parameter and the probability of exactly one event. (Answer: \(\lambda=1.2\); \(P(N=1)=1.2e^{-1.2}=0.3614\).)