Poisson Frequency Models And Rate Parameters

Introduction

A claim count is a nonnegative whole number, but the number of claims that will occur during a future period is uncertain. A frequency model describes that uncertainty. The Poisson model is a standard choice when the count is represented by one parameter, $\lambda$, whose meaning depends on the exposure period or unit specified in the model.

For sound interpretation, always identify three things before using the model: what is counted, over what exposure, and what value of $\lambda$ applies to that exposure. The count is not a claim amount, and $\lambda$ is not a maximum possible count.

Definition: A random variable $N$ has a Poisson distribution with parameter $\lambda$, written $N\sim\operatorname{Poisson}(\lambda)$, if $\lambda>0$ and $$ P(N=k)=e^{-\lambda}\frac{\lambda^k}{k!},\qquad k=0,1,2,\ldots. $$ The support is the set of all nonnegative integers. The parameter $\lambda$ is the expected number of events over the exposure period or unit to which the model refers.

What the Rate Parameter Means

The word rate can refer to a rate per unit of exposure, such as claims per vehicle-year. In the Poisson probability function, however, $\lambda$ is the parameter for the particular count being modeled. If $N$ counts claims over a specified policy term, $\lambda$ is the expected number of claims over that term. State the term or exposure whenever reporting the parameter.

For example, “$\lambda=1.8$” is incomplete if the exposure is not known. “The number of claims during one six-month policy term is Poisson with parameter $\lambda=1.8$” identifies both the random variable and the period. The value 1.8 is an expected count, not a prediction that 1.8 claims will actually occur. An observed count must be a whole number.

A model with a common rate per exposure unit can use different parameter values for different exposure amounts. The rate and the parameter are related, but they are not interchangeable: the parameter refers to a particular amount of exposure. When converting between units, make sure the rate and exposure use matching units. Detailed probability calculations are separate from identifying and interpreting the model.

Worked Example: Interpreting a Policy-Term Parameter

A hypothetical insurer models the number of windshield claims on one policy during a six-month term as $N\sim\operatorname{Poisson}(1.8)$. State what the parameter means, calculate the mean and variance of the count, and give its support.

Method. Treat $N$ as the count over the stated six-month policy term. For a Poisson variable, both the expected count and the variance equal its parameter.

Parameter interpretation. Here $\lambda=1.8$, so the expected number of windshield claims per six-month term is 1.8. This does not mean every policy has exactly 1.8 claims; each realized count is an integer.

Mean and variance. $$ E[N]=\lambda=1.8,\qquad \operatorname{Var}(N)=\lambda=1.8. $$ The standard deviation is $\sqrt{1.8}\approx1.3416$ claims.

Support. The possible values are $0,1,2,3,\ldots$. Zero claims is possible, and there is no finite maximum count in the model.

Check. The expected count and variance use the same parameter, and the support includes only nonnegative whole-number counts. The common mistake avoided here is to interpret 1.8 as a claim amount or as a maximum number of claims.

Support and Basic Consequences

The support specifies the values to which the probability function assigns probability. For a Poisson count, zero is included: a period may have no claims. Negative counts and fractional counts are not possible. The support has no finite upper endpoint, even though very large counts may be unlikely for a particular parameter.

A Poisson model with $\lambda=0$ is sometimes used as a degenerate boundary case: the count is zero with probability 1. The usual Poisson model for claim frequency assumes $\lambda>0$, as in the definition above. In either case, the count cannot be negative or fractional.

Theorem: If $N\sim\operatorname{Poisson}(\lambda)$, where $\lambda>0$, then $E[N]=\lambda$ and $\operatorname{Var}(N)=\lambda$. Thus the parameter is both the expected count and the variance of the count, for the exposure to which the model applies.

The mean and variance result follows directly from the probability function. This also explains why $\lambda$ has a specific role beyond being a label for the distribution.

Find the mean.
Using the probability function and the fact that the $k=0$ term contributes zero to the mean, $$ E[N]=\sum_{k=1}^{\infty}k e^{-\lambda}\frac{\lambda^k}{k!} =\lambda e^{-\lambda}\sum_{k=1}^{\infty}\frac{\lambda^{k-1}}{(k-1)!} =\lambda. $$ In the last step, the sum is the exponential series $e^\lambda$, which cancels $e^{-\lambda}$.

Find the second factorial moment.
Terms for $k=0$ and $k=1$ contribute zero, so $$ E[N(N-1)]=\sum_{k=2}^{\infty}k(k-1)e^{-\lambda}\frac{\lambda^k}{k!} =\lambda^2 e^{-\lambda}\sum_{k=2}^{\infty}\frac{\lambda^{k-2}}{(k-2)!} =\lambda^2. $$ Again, the remaining sum is $e^\lambda$.

Convert to variance.
Since $N^2=N(N-1)+N$, $$ E[N^2]=E[N(N-1)]+E[N]=\lambda^2+\lambda. $$ Therefore $$ \operatorname{Var}(N)=E[N^2]-(E[N])^2=(\lambda^2+\lambda)-\lambda^2=\lambda. $$

This result concerns the count for the specific exposure used to define $\lambda$. If a model is described using a different exposure unit, the parameter must be interpreted consistently with that unit. The mean-variance equality is a property of the Poisson model; it is not a general fact about every claim-count distribution.

Worked Example: Separating a Rate per Exposure From a Parameter

A hypothetical equipment warranty portfolio is modeled at a rate of $0.12$ claims per machine-year. For a group with $15$ machine-years of exposure, suppose the claim count is Poisson under this common-rate model. Find the Poisson parameter for the group, interpret it, and state the mean, variance, and support.

Method. Multiply the rate per machine-year by the exposure in machine-years. The units cancel: $$ \lambda=(0.12\text{ claims per machine-year})(15\text{ machine-years})=1.8\text{ claims}. $$ The group count is therefore $N\sim\operatorname{Poisson}(1.8)$.

Interpretation. The expected count for the full group exposure of 15 machine-years is 1.8 claims. The rate is $0.12$ per machine-year; the Poisson parameter for this group is $1.8$. They are different quantities because they refer to different amounts of exposure.

Mean, variance, and support. $$ E[N]=1.8,\qquad \operatorname{Var}(N)=1.8,\qquad N\in\{0,1,2,\ldots\}. $$ Check. The product has units of claims, as required for an expected count. The common mistake avoided is using $0.12$ as the group parameter without accounting for the 15 machine-years of exposure.

Worked Example: Keeping Time Units Consistent

A hypothetical mobile-device protection program has a modeled frequency of $0.04$ claims per covered device-month. A block of $30$ covered device-months is observed. Under the stated rate model, identify the count distribution and its parameter, then find its mean, variance, and support.

Method. Use matching units: device-months in the exposure and claims per device-month in the rate. $$ \lambda=(0.04\text{ claims per device-month})(30\text{ device-months})=1.2. $$ Thus $N\sim\operatorname{Poisson}(1.2)$ for the block.

Interpretation and consequences. The block has an expected 1.2 claims. The variance is also 1.2, and its possible counts are $0,1,2,\ldots$. A realized count is always an integer, despite the noninteger expected count.

Check. The device-month unit cancels in the rate-exposure product. The common mistake avoided is using an annual unit or multiplying by 30 devices when the supplied exposure is already expressed in device-months.

Modeling Choices and Exam Traps

The Poisson distribution specifies a count model, not the severity of the claims. If claim amounts are also modeled, they are separate random quantities. Likewise, the meaning of a count depends on the event definition: “claims reported,” “claims incurred,” and “claims paid” need not describe the same random variable. Define the event and the observation period before assigning a distribution.

The Poisson probability function assigns probabilities to every nonnegative integer and totals to 1. Its shape is controlled by $\lambda$, but the presence of a Poisson formula alone does not decide whether the model is appropriate for a particular portfolio. The modeling assumption must be supplied or justified in the surrounding analysis. In particular, do not infer that a rate is per year if the problem states a different unit or period.

A frequent error is to confuse a rate with a count parameter. If a rate is given per exposure unit, it must be paired with the exposure being modeled before it becomes the parameter for that total count. Another error is to treat $\lambda$ as a maximum number of claims; the model has no finite upper support. Finally, never include negative or fractional values as possible observed claim counts.

Key takeaway: For $N\sim\operatorname{Poisson}(\lambda)$, $\lambda$ is the expected claim count for a clearly stated exposure, the variance is also $\lambda$, and the count can take any value in $0,1,2,\ldots$. Always make the exposure and event definition explicit.

Check Your Understanding

For each item, identify the relevant parameter or support carefully and keep the stated exposure units intact.

A hypothetical travel policy has a Poisson count of baggage claims during a one-year term, with $\lambda=0.7$. State the expected count, variance, and support. (Answer: Mean $0.7$, variance $0.7$, support $0,1,2,\ldots$.)
A hypothetical service plan has a rate of $0.08$ claims per device-year and $10$ device-years of exposure. Under a common-rate Poisson model, what is the parameter for the total count? (Answer: $\lambda=0.08\times10=0.8$ claims.)
A hypothetical agricultural program models claims at $0.03$ per insured-acre-month. For $40$ insured-acre-months, state the Poisson parameter and its interpretation. (Answer: $\lambda=0.03\times40=1.2$, the expected count over the stated exposure.)
A Poisson claim count for a specified contract period has parameter $2.5$. A candidate says the possible counts are $1,2,3,\ldots$ and the maximum is 2.5. Identify both errors and give the correct description. (Answer: Zero is also possible, and there is no finite maximum; the support is $0,1,2,\ldots$, while $2.5$ is the expected count.)
A hypothetical repair portfolio has a frequency of $0.15$ claims per repair-year and $24$ repair-years of exposure. Under the stated common-rate model, give the parameter, mean, variance, and support for the total count. (Answer: $\lambda=0.15\times24=3.6$; mean $3.6$, variance $3.6$, support $0,1,2,\ldots$.)

What You'll Learn