Tweedie Distributions As Compound Losses

Why a Tweedie Model Can Describe Aggregate Losses

For an annual insurance outcome, the total loss can be zero because no claims occurred, or positive because one or more claims occurred. A model that represents both possibilities naturally is useful when the response is an aggregate such as annual loss per policy. For Tweedie power between one and two, the distribution has a compound Poisson-Gamma interpretation: the number of claims is random, and each claim amount is positive and Gamma distributed.

The interpretation is more than a convenient description. It lets an actuary translate a fitted Tweedie model into a claim frequency and a claim severity distribution, subject to the assumptions below. Conversely, a compound Poisson-Gamma model can be summarized using Tweedie parameters. The central task is to keep the parameterizations consistent.

Definition: Let $N\sim\operatorname{Poisson}(\lambda)$, and let $Z_1,Z_2,\ldots$ be independent and identically distributed claim amounts, independent of $N$. If each $Z_j$ has a Gamma distribution with shape $\alpha>0$ and scale $\theta>0$, then the aggregate loss is $$ X=\sum_{j=1}^{N} Z_j, $$ where the sum is defined to be zero when $N=0$. This is a compound Poisson-Gamma model. Here $\lambda$ is the expected claim count, and the Gamma severity has mean $\alpha\theta$ and variance $\alpha\theta^2$.

The Gamma convention matters: throughout this tutorial, $\theta$ is the scale, not the rate. Thus a Gamma variable with shape $\alpha$ and scale $\theta$ has density proportional to $z^{\alpha-1}e^{-z/\theta}$ for $z>0$. An equivalent rate parameter would be $1/\theta$.

For the Tweedie description, write the mean as $\mu>0$, the dispersion as $\phi>0$, and the power as $p$. In the range $1<p<2$, the Tweedie model is exactly representable as the compound Poisson-Gamma model above. Its variance function is $\operatorname{Var}(X)=\phi\mu^p$; this identity is useful for translating and checking parameters.

Theorem: For a Tweedie model with mean $\mu$, dispersion $\phi$, and $1<p<2$, the equivalent compound Poisson-Gamma parameters are $$ \alpha=\frac{2-p}{p-1},\qquad \theta=\phi(p-1)\mu^{p-1},\qquad \lambda=\frac{\mu^{2-p}}{\phi(2-p)}. $$ Thus $N\sim\operatorname{Poisson}(\lambda)$, and conditional claim amounts are independent $\operatorname{Gamma}(\alpha,\theta)$ variables in the shape-scale convention.

The restrictions $1<p<2$ ensure that both $\alpha$ and $\theta$ are positive and that the count parameter is positive. As $p$ approaches either endpoint, the parameter behavior changes; this compound Poisson-Gamma correspondence should not be applied at $p=1$, $p=2$, or outside that interval.

Why the Parameter Conversion Works

A useful justification begins with the moments of the compound model. Conditional on $N$, the expected aggregate is $N\alpha\theta$, and its conditional variance is $N\alpha\theta^2$. Applying the law of total expectation and the law of total variance, with $E[N]=\operatorname{Var}(N)=\lambda$, gives:

E[X]=\lambda\alpha\theta,\qquad \operatorname{Var}(X) =\lambda\alpha\theta^2+\lambda(\alpha\theta)^2 =\lambda\alpha(\alpha+1)\theta^2.

The first variance term reflects variation in severity conditional on the count; the second reflects variation in the count itself. In particular, a Poisson claim count contributes variability even if the average severity is held fixed.

To check the Tweedie-to-compound conversion, first note that the shape relation implies $\alpha+1=1/(p-1)$. Substituting the stated parameters into the compound mean gives:

\lambda\alpha\theta = \frac{\mu^{2-p}}{\phi(2-p)} \cdot\frac{2-p}{p-1} \cdot\phi(p-1)\mu^{p-1} =\mu.

For the variance, use $\alpha(\alpha+1)=(2-p)/(p-1)^2$. Then:

\lambda\alpha(\alpha+1)\theta^2 = \frac{\mu^{2-p}}{\phi(2-p)} \cdot\frac{2-p}{(p-1)^2} \cdot\phi^2(p-1)^2\mu^{2p-2} =\phi\mu^p.

So the conversion gives the Tweedie mean and variance. The compound Poisson-Gamma representation is an exact distributional correspondence in this power range, not a claim that any two distributions with matching first two moments must be identical.

Start with the Tweedie parameters.
Record $\mu$, $\phi$, and $p$, and confirm that $1<p<2$.

Calculate the severity shape and scale.
Use $\alpha=(2-p)/(p-1)$ and $\theta=\phi(p-1)\mu^{p-1}$. Keep the scale convention explicit.

Calculate the claim-count parameter.
Use $\lambda=\mu^{2-p}/[\phi(2-p)]$. This is the expected count over the exposure period represented by $X$.

Check the implied aggregate.
Verify $E[X]=\lambda\alpha\theta=\mu$ and $\operatorname{Var}(X)=\lambda\alpha(\alpha+1)\theta^2=\phi\mu^p$.

Worked Conversions

Worked Example: Translate a Tweedie Model Into Claim Frequency and Severity

A hypothetical insurer models annual aggregate claim dollars for one exposure unit using a Tweedie distribution with $\mu=4$ thousand dollars, $\phi=0.5$, and $p=1.5$. Find the equivalent claim-count and claim-amount distributions, then check the aggregate mean and variance. Assume no discounting or mortality adjustment: this is an annual loss model. Use a Gamma shape-scale parameterization.

Method and assumptions. Apply the Tweedie-to-compound formulas. Claims are assumed independent and identically distributed conditional on the Poisson count, and the count is independent of the claim amounts.

First calculate the Gamma shape: $\alpha=(2-1.5)/(1.5-1)=0.5/0.5=1$. Next calculate the scale: $\theta=0.5(0.5)(4^{0.5})=0.25(2)=0.5$ thousand dollars. Finally calculate the Poisson mean: $\lambda=4^{0.5}/[0.5(2-1.5)]=2/(0.25)=8$.

Therefore $N\sim\operatorname{Poisson}(8)$, and each claim amount is $\operatorname{Gamma}(1,0.5)$ thousand dollars. Its mean is $\alpha\theta=1(0.5)=0.5$ thousand dollars, or $500.

Check the aggregate mean: $8(1)(0.5)=4$ thousand dollars, matching $\mu$. The Gamma second moment is $\alpha(\alpha+1)\theta^2=1(2)(0.5^2)=0.5$ in squared thousands of dollars. The compound aggregate variance is $8(0.5)=4$ squared thousands of dollars. The Tweedie variance gives $0.5(4^{1.5})=0.5(8)=4$, also matching.

Check. Both the implied expected aggregate and variance agree with the Tweedie parameters. The conversion avoids treating $\mu$ as a claim severity: here $\mu$ is the aggregate mean, while the average claim is $500.

Worked Example: Recover Tweedie Parameters From a Compound Model

A hypothetical equipment policy has annual claim count $N\sim\operatorname{Poisson}(8)$. Claim amounts are independent $\operatorname{Gamma}(2,0.5)$ thousand dollars, independent of $N$. Find the corresponding Tweedie power, mean, and dispersion, and verify the variance. Assume no discounting or mortality adjustment.

Method and assumptions. Use the compound-model moments and reverse the shape relationship. The Gamma values use shape 2 and scale 0.5 thousand dollars.

From $\alpha=(2-p)/(p-1)$, solve for $p$: $\alpha(p-1)=2-p$, so $p(\alpha+1)=\alpha+2$, and therefore $p=(\alpha+2)/(\alpha+1)=(2+2)/(2+1)=4/3$.

The Gamma mean is $2(0.5)=1$ thousand dollars. Thus the aggregate mean is $\mu=\lambda\alpha\theta=8(2)(0.5)=8$ thousand dollars.

Rearrange the scale formula to find dispersion: $\phi=\theta/[(p-1)\mu^{p-1}]$. Here $p-1=1/3$, and $8^{1/3}=2$, so $\phi=0.5/[(1/3)(2)]=0.5/(2/3)=0.75$.

The Gamma second moment is $\alpha(\alpha+1)\theta^2=2(3)(0.5^2)=1.5$ squared thousands of dollars. The compound variance is $8(1.5)=12$. The Tweedie variance is $0.75(8^{4/3})=0.75(16)=12$ squared thousands of dollars.

Check. The reverse conversion gives $p=4/3$, $\mu=8$, and $\phi=0.75$, and both variance calculations give 12. A common error is to use the Gamma variance alone as the aggregate variance; that omits the contribution from the random Poisson count.

Worked Example: Interpret a Tweedie Model With a Different Power

A hypothetical annual aggregate model for a small commercial property portfolio has Tweedie parameters $\mu=8$ thousand dollars, $\phi=0.75$, and $p=4/3$. Translate it to the compound model and calculate the probability of no claims. Assume the stated annual period and no discounting or mortality adjustment.

Method and assumptions. Convert to the Poisson-Gamma representation. Under this model, no claims means exactly $N=0$; no separate claim-severity calculation is needed for that event.

The shape is $\alpha=(2-4/3)/(4/3-1)=(2/3)/(1/3)=2$. The scale is $\theta=0.75(1/3)(8^{1/3})=0.75(1/3)(2)=0.5$ thousand dollars. The Poisson mean is $\lambda=8^{2/3}/[0.75(2/3)]=4/0.5=8$. Thus the claim count is $\operatorname{Poisson}(8)$, and each claim is $\operatorname{Gamma}(2,0.5)$ thousand dollars.

For a Poisson random variable, $P(N=0)=e^{-\lambda}$. Consequently, $P(X=0)=e^{-8}=0.00033546\ldots$, or 0.0003 to four decimal places.

As a parameter check, claim severity has mean $2(0.5)=1$ thousand dollars, giving aggregate mean $8(1)=8$ thousand dollars. Its second moment is $2(3)(0.5^2)=1.5$, so aggregate variance is $8(1.5)=12$; the Tweedie calculation is $0.75(8^{4/3})=12$. The check confirms the translated parameters.

Check. The zero probability depends on $\lambda$, not on the Gamma scale. Multiplying the probability by a severity probability would incorrectly mix the count event with claim amounts.

Assumptions and Common Exam Traps

The compound interpretation is useful only when its assumptions fit the outcome being modeled. It treats the count as Poisson, so counts have equal mean and variance conditional on the modeled parameters. It treats claim amounts as positive Gamma variables, independent across claims and independent of the count. It also assumes the same severity distribution for every claim in the aggregate. Changes in exposure or risk characteristics may require different model parameters; a single aggregate Tweedie distribution does not by itself explain those differences.

A frequent calculation error is to use $\lambda\operatorname{Var}(Z)$ for the aggregate variance. That is not the variance of a random sum with a Poisson count. For a compound Poisson total, the correct variance is $\lambda E[Z^2]$, which includes the contribution from uncertainty in the number of claims.

\lambda\operatorname{Var}(Z) \quad\text{is not generally}\quad \operatorname{Var}(X); \qquad \operatorname{Var}(X)=\lambda E[Z^2] =\lambda\alpha(\alpha+1)\theta^2.

For example, in the first worked example, the Gamma severity variance is $1(0.5^2)=0.25$ squared thousands of dollars. The incorrect calculation would give $8(0.25)=2$. The correct calculation uses the severity second moment $0.5$, giving $8(0.5)=4$. The omitted 2 is the count-variation contribution.

Other traps to guard against include using a Gamma rate where a scale is required, confusing aggregate mean $\mu$ with average claim severity $\alpha\theta$, and applying the conversion when $p$ is not strictly between one and two. State the parameterization and units before calculating. If loss is in thousands of dollars, then severity and aggregate means must use the same units, and variances use squared units.

Key takeaway: For $1<p<2$, a Tweedie aggregate has an exact compound Poisson-Gamma representation. Convert using $\alpha=(2-p)/(p-1)$, $\theta=\phi(p-1)\mu^{p-1}$, and $\lambda=\mu^{2-p}/[\phi(2-p)]$; then verify the aggregate mean and variance before interpreting the frequency and severity.

Check Your Understanding

For each problem, identify the relevant model and check that the requested aggregate quantities use both count and severity information.

A hypothetical annual loss has a compound Poisson-Gamma model with $\lambda=5$, shape $\alpha=4$, and scale $\theta=2$ dollars. Find $p$, the aggregate mean, and the aggregate variance. (Answer: $p=6/5$, mean $40$ dollars, variance $200$ dollars squared.)
A Tweedie aggregate has $\mu=9$ thousand dollars, $\phi=2$, and $p=1.5$. Find its equivalent Poisson mean and Gamma shape and scale. (Answer: $\lambda=3$, $\alpha=1$, $\theta=3$ thousand dollars.)
A hypothetical compound model has $\lambda=6$ and Gamma severity with shape 2 and scale 1 thousand dollars. Find its Tweedie power, mean, and dispersion, and verify the variance. (Answer: $p=4/3$, $\mu=12$ thousand dollars, $\phi=3/\sqrt[3]{12}$; variance $36$ squared thousands of dollars.)
A Tweedie aggregate has $\mu=16$ thousand dollars, $\phi=0.5$, and $p=1.25$. Find $\alpha$, $\theta$, and $\lambda$, and check the aggregate mean. (Answer: $\alpha=3$, $\theta=0.25$ thousand dollars, $\lambda=64/3$; aggregate mean $16$ thousand dollars.)
A compound Poisson-Gamma model has $\lambda=2$ and Gamma severity with shape 1 and scale 3 dollars. Calculate the correct aggregate variance and compare it with the incorrect calculation $2\operatorname{Var}(Z)$. (Answer: correct variance $36$ dollars squared; incorrect value $18$ dollars squared.)

What You'll Learn