Tweedie Variance Power and Dispersion

Aggregate Pure Premium Variability

A pure premium is an expected aggregate loss per chosen exposure unit. Policies with the same expected pure premium can still have different degrees of variability: one may have relatively frequent, moderate claims, while another may have rarer, larger claims. A Tweedie model summarizes the relationship between the mean and variance of an aggregate response with a power parameter and a dispersion parameter.

Here, $Y$ denotes the aggregate loss for one fixed, standardized exposure unit, and $\mu=E[Y]$ is its expected pure premium. The response is aggregate loss, not the severity of an individual claim. The distinction matters: the Tweedie variance function describes variability across aggregate responses at the policy or exposure-unit level.

Definition: A Tweedie variance function specifies that an aggregate response $Y$, with mean $\mu>0$, has variance $$ \operatorname{Var}(Y)=\phi\mu^p, $$ where $\phi>0$ is the dispersion parameter and $p$ is the variance power. For the compound Poisson-Gamma aggregate-loss interpretation, $1<p<2$. The mean and variance are conditional on the modeled risk characteristics and the fixed exposure basis.

The equation specifies the variance given the mean; it does not say that the variance equals the mean, nor does it make the variance constant across risks. For a fixed $p$ and $\mu$, increasing $\phi$ increases the variance in direct proportion. The power $p$ controls how rapidly variance changes as the mean changes. Taking logarithms gives $\log\operatorname{Var}(Y)=\log\phi+p\log\mu$, so $p$ is the elasticity of variance with respect to the mean when $\phi$ is held fixed.

What Dispersion and Power Tell Us

At fixed $\mu$ and $p$, multiplying $\phi$ by a factor $c>0$ multiplies the variance by $c$, the standard deviation by $\sqrt{c}$, and the coefficient of variation by $\sqrt{c}$. The coefficient of variation is the standard deviation divided by the mean. For the Tweedie variance function,

\operatorname{SD}(Y)=\sqrt{\phi}\,\mu^{p/2}, \qquad \operatorname{CV}(Y)=\frac{\sqrt{\operatorname{Var}(Y)}}{\mu} =\sqrt{\phi}\,\mu^{(p-2)/2}.

For $1<p<2$, the exponent $(p-2)/2$ is negative. Thus, if $p$ and $\phi$ are fixed, increasing the mean increases the variance, but decreases the coefficient of variation. The absolute spread grows while the spread relative to the mean shrinks. This conclusion is conditional on holding the dispersion and power fixed; it is not a claim that every portfolio will have lower relative variability at higher expected loss.

The numerical value of $\phi$ depends on the units used for $Y$, and its units depend on $p$. If $Y$ is measured in dollars, then variance is measured in dollars squared, so $\phi$ has units of dollars raised to $2-p$. Rescaling losses from dollars to thousands of dollars changes both the numerical mean and the numerical dispersion. In addition, when comparing models with different powers, their dispersion parameters are generally estimated separately. Holding the same numerical $\phi$ across different powers is useful for illustrating the formula, but is not by itself a fair model-fit comparison.

Worked Example: Calculating Variance and Relative Variability

A hypothetical insurer models annual aggregate loss in dollars for a fixed group of risks with a Tweedie variance function. The fitted mean is $\mu=400$, the power is $p=1.5$, and the dispersion is $\phi=0.05$. Calculate the variance, standard deviation, and coefficient of variation. Round final dollar amounts to the cent and the coefficient of variation to four decimals.

Method. Apply the variance function to the aggregate response, then take its square root for the standard deviation and divide by the mean for the coefficient of variation.

First, $400^{1.5}=400\sqrt{400}=400(20)=8{,}000$. Therefore, $$ \operatorname{Var}(Y)=0.05(8{,}000)=400\text{ dollars}^2. $$

The standard deviation is $\sqrt{400}=20$, or $\$20.00$. The coefficient of variation is $20/400=0.05$, or $0.0500$. Thus this model has variance $400\text{ dollars}^2$, standard deviation $\$20.00$, and coefficient of variation $0.0500$.

Check and trap avoided. Squaring the standard deviation returns $20^2=400$, which checks the variance. A common mistake is to report $400$ as the standard deviation; it is the variance and has squared-dollar units.

Worked Example: Comparing Variance Powers

For a hypothetical aggregate-loss response measured in a fixed unit, consider three candidate variance functions with the same mean $\mu=100$ and the same illustrative numerical dispersion $\phi=0.02$, but powers $p=1.2$, $1.5$, and $1.8$. Calculate the variance under each candidate and identify the largest. Round variances to four decimals.

Method. Evaluate $0.02(100)^p$ separately for each power. Since $100^p=10^{2p}$, the powers can be calculated directly.

For $p=1.2$, $100^{1.2}=251.1886$, so the variance is $0.02(251.1886)=5.0238$. For $p=1.5$, $100^{1.5}=1{,}000$, so the variance is $0.02(1{,}000)=20.0000$. For $p=1.8$, $100^{1.8}=3{,}981.0717$, so the variance is $0.02(3{,}981.0717)=79.6214$.

The largest variance among these calculations is $79.6214$, for $p=1.8$. This is an illustration of the formula with $\phi$ held fixed numerically. In actual model comparisons, dispersion is estimated for each candidate, and its units vary with the power. The calculation alone does not establish which model fits data better.

Check and trap avoided. At a fixed positive mean greater than one and a fixed numerical $\phi$, a larger power produces a larger value of $\phi\mu^p$. Do not interpret the exercise as permission to compare fitted models while ignoring their separately estimated dispersions.

Why the Compound Poisson-Gamma Variance Has This Form

For $1<p<2$, the Tweedie aggregate-loss model has a compound Poisson-Gamma representation. This is useful for understanding why its variance increases faster than its mean. Let $N$ be Poisson with mean $\lambda$, and let the independent claim amounts $Z_1,Z_2,\ldots$ be Gamma with shape $\alpha$ and scale $\theta$. Under the shape-scale convention, $E[Z_j]=\alpha\theta$ and $\operatorname{Var}(Z_j)=\alpha\theta^2$. The aggregate is $Y=\sum_{j=1}^{N}Z_j$, with an empty sum equal to zero.

Conditional on $N$, the sum has mean $N\alpha\theta$ and variance $N\alpha\theta^2$. Applying the law of total variance gives a short derivation:

\begin{aligned} \operatorname{Var}(Y) &=E[\operatorname{Var}(Y\mid N)] +\operatorname{Var}(E[Y\mid N])\\ &=E[N]\alpha\theta^2+\operatorname{Var}(N\alpha\theta)\\ &=\lambda\alpha\theta^2+\lambda\alpha^2\theta^2\\ &=\lambda\alpha(\alpha+1)\theta^2 =\lambda E[Z_1^2]. \end{aligned}

The last equality follows because $E[Z_1^2]=\operatorname{Var}(Z_1)+(E[Z_1])^2=\alpha\theta^2+\alpha^2\theta^2$. The first term in the total-variance calculation represents variation in claim amounts given the claim count; the second represents variation in the count itself.

The Tweedie parameters correspond to the following compound Poisson-Gamma parameters when $\mu>0$, $\phi>0$, and $1<p<2$:

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

These relationships give the stated mean and variance. For the mean,

E[Y]=\lambda\alpha\theta =\frac{\mu^{2-p}}{\phi(2-p)} \frac{2-p}{p-1}\phi(p-1)\mu^{p-1} =\mu.

For the variance, $\alpha+1=1/(p-1)$, so substitution into the compound variance yields

\begin{aligned} \operatorname{Var}(Y) &=\frac{\mu^{2-p}}{\phi(2-p)} \frac{2-p}{p-1} \frac{1}{p-1} \phi^2(p-1)^2\mu^{2p-2}\\ &=\phi\mu^p. \end{aligned}

This derivation also clarifies the scope of the interpretation: this particular positive-aggregate-loss construction uses $1<p<2$, with positive Poisson frequency and positive Gamma claim amounts. A variance function with another power may be useful in another statistical setting, but this compound Poisson-Gamma construction does not automatically apply to it.

Worked Example: Verifying the Tweedie Variance From Claim Components

A hypothetical portfolio has aggregate losses measured in thousands of dollars per fixed exposure unit. Suppose the claim count $N$ is Poisson with mean $\lambda=200$, and each claim amount is Gamma with shape $\alpha=1$ and scale $\theta=0.5$ thousand dollars. Find the aggregate mean and variance, then identify the corresponding Tweedie power and dispersion.

Method. Calculate the compound Poisson moments and match them to $\operatorname{Var}(Y)=\phi\mu^p$. The Poisson count has mean and variance 200. Each Gamma claim has mean $\alpha\theta=1(0.5)=0.5$ thousand dollars and variance $\alpha\theta^2=1(0.5)^2=0.25$ thousand-dollars squared.

The aggregate mean is $$ E[Y]=\lambda E[Z_1]=200(0.5)=100\text{ thousand dollars}. $$ The aggregate variance is $$ \operatorname{Var}(Y) =\lambda\left(\operatorname{Var}(Z_1)+(E[Z_1])^2\right) =200(0.25+0.5^2) =200(0.50) =100\text{ thousand-dollars}^2. $$

Because the Gamma shape is $\alpha=1$, the parameter relationship gives $1=(2-p)/(p-1)$. Thus $p-1=2-p$, so $p=1.5$. Matching $\theta=\phi(p-1)\mu^{p-1}$ gives $$ 0.5=\phi(0.5)(100^{0.5})=\phi(0.5)(10)=5\phi, $$ so $\phi=0.1$. The Tweedie variance checks: $0.1(100^{1.5})=0.1(1{,}000)=100$ thousand-dollars squared.

Check and trap avoided. The aggregate variance includes both the individual claim variance and the squared claim mean. Using only $200(0.25)=50$ would omit the claim-count contribution and fail the Tweedie check.

Using the Variance Function Carefully

In a fitted ratemaking model, $\mu$ varies by risk characteristics, while a chosen power and dispersion describe how the conditional variance relates to that mean. The function can therefore distinguish a model in which variance grows approximately in proportion to the mean from one in which it grows more quickly. It is a model for conditional variability, not a guarantee that every observed aggregate will lie near its fitted mean.

A reliable calculation follows a fixed sequence:

Identify the response and exposure basis.
Confirm that $Y$ is the aggregate response being modeled and that $\mu$, $\phi$, and the response units refer to the same fixed basis.

Check the parameters.
Use $\mu>0$, $\phi>0$, and, for the compound Poisson-Gamma aggregate-loss interpretation, $1<p<2$.

Calculate variance first.
Evaluate $\phi\mu^p$; take a square root only if standard deviation is requested.

Interpret the result in context.
Report variance in squared response units, standard deviation in response units, and any coefficient of variation as a unitless ratio.

The most frequent exam trap is to confuse the power in the variance function with the exponent needed for a standard deviation, or to confuse variance with standard deviation. For the first worked example, the incorrect calculation $0.05(400^{0.5})=0.05(20)=1$ uses $p-1$ instead of the stated power $p=1.5$. The corrected variance is $0.05(400^{1.5})=400\text{ dollars}^2$, and the standard deviation is $\sqrt{400}=\$20.00$. Always apply the stated formula to variance before deriving other measures of spread.

Key takeaway: In $\operatorname{Var}(Y)=\phi\mu^p$, dispersion scales variance at a fixed mean and power, while the power governs how variance responds to changes in the mean. For $1<p<2$, the Tweedie aggregate-loss model has a compound Poisson-Gamma interpretation, and its variance includes both claim-count and claim-amount variability.

Check Your Understanding

For each problem, use the stated units and round only as directed.

A hypothetical aggregate response has $\mu=100$, $\phi=0.03$, and $p=1.5$. Find its variance and standard deviation. (Answer: variance $=30$ response-units squared; standard deviation $=5.4772$ response units.)
With $\phi$ fixed, the mean doubles and $p=1.6$. By what factor does the variance change? (Answer: $2^{1.6}=3.0314$, so variance is multiplied by $3.0314$.)
A model has variance $90$, mean $\mu=300$, and power $p=1.5$. Find $\phi$, rounded to five decimal places. (Answer: $\phi=90/300^{1.5}=90/5{,}196.1524=0.01732$.)
For a compound Poisson-Gamma aggregate, $p=1.5$, $\mu=25$, and $\phi=0.2$. Use the parameter relationships to find $\lambda$, $\alpha$, and $\theta$, then verify the variance. (Answer: $\lambda=50$, $\alpha=1$, $\theta=0.5$; variance $=50(1)(2)(0.5)^2=25=\phi\mu^p$.)
At fixed $\phi$ and $p=1.25$, the mean increases from $\mu$ to $4\mu$. By what factor does the coefficient of variation change? (Answer: $4^{(1.25-2)/2}=4^{-0.375}=0.5946$; the coefficient of variation is multiplied by $0.5946$.)

What You'll Learn