Tweedie Zero Mass And Positive Losses

Why a Tweedie Aggregate Can Be Exactly Zero

A claim count and an aggregate loss answer different questions. The count records how many claim events occurred; the aggregate records the sum of their positive severities. In a compound Poisson-Gamma model, an annual aggregate loss can therefore be exactly zero: it is zero when no claim events occur. If one or more events occur, the aggregate is positive and varies continuously.

This distinction matters when interpreting a Tweedie model with power between 1 and 2. Such a model represents aggregate outcomes using a Poisson event count and positive Gamma severities. Its probability of a zero outcome is not an extra parameter added independently of the count process: under the model assumptions, it follows directly from the probability of zero events.

Definition: Let $N\sim\operatorname{Poisson}(\lambda)$, where $\lambda>0$, and let $Y_1,Y_2,\ldots$ be independent and identically distributed Gamma random variables with shape $\alpha>0$ and scale $\theta>0$, independent of $N$. Define the aggregate loss by $S=\sum_{j=1}^{N}Y_j$, with the empty sum equal to zero. Then $S$ has a point mass at zero and a continuous distribution over positive losses.

The Gamma convention here is shape and scale, so a Gamma$(\alpha,\theta)$ severity has density $y^{\alpha-1}e^{-y/\theta}/[\Gamma(\alpha)\theta^\alpha]$ for $y>0$. In particular, each severity is strictly positive with probability one. The event count is the only source of an exact zero in this model.

Deriving the Zero Probability

Condition on the number of events. If $N=0$, there are no terms in the sum defining $S$, so $S=0$. If $N\geq 1$, every $Y_j>0$ with probability one, and hence their sum is strictly positive. Thus the events $\{S=0\}$ and $\{N=0\}$ agree, apart from probability-zero exceptions.

P(S=0)=P(N=0)=e^{-\lambda}.

Consequently, the probability of a positive aggregate is $P(S>0)=1-e^{-\lambda}$. The Gamma shape and scale determine the sizes and distribution of positive outcomes, but do not change the probability of an exact zero when $\lambda$ is fixed. Increasing the Poisson intensity makes no-event outcomes less likely.

For completeness, conditional on $N=n\geq1$, the sum of $n$ independent Gamma severities with the same scale $\theta$ has a Gamma$(n\alpha,\theta)$ distribution. This follows, for example, from multiplying their moment-generating functions: each contributes $(1-\theta t)^{-\alpha}$, so the product is $(1-\theta t)^{-n\alpha}$, the moment-generating function of Gamma$(n\alpha,\theta)$. Mixing over the possible positive counts gives a continuous density on $s>0$:

f_S(s)=\sum_{n=1}^{\infty}e^{-\lambda}\frac{\lambda^n}{n!} \frac{s^{n\alpha-1}e^{-s/\theta}}{\Gamma(n\alpha)\theta^{n\alpha}}, \qquad s>0.

Each term is the probability of count $n$ multiplied by the conditional severity density for that count. The density over positive values has total probability $1-e^{-\lambda}$, not 1; the remaining probability $e^{-\lambda}$ is concentrated at the single point zero. This is a mixed distribution: it has both a discrete atom and a continuous component.

Connection to Tweedie Parameters

For a Tweedie model with mean $\mu>0$, dispersion $\phi>0$, and power $1<p<2$, the compound Poisson-Gamma representation uses the following parameters under the shape-scale convention above:

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

In this parameterization the zero probability is therefore

P(S=0)=\exp\!\left(-\frac{\mu^{2-p}}{\phi(2-p)}\right).

The event intensity $\lambda$ governs the probability of observing no events. Holding $p$ and $\phi$ fixed, a larger $\mu$ gives a larger $\lambda$ in this range and thus a smaller zero probability. A larger $\phi$, with the other parameters fixed, gives a smaller $\lambda$ and a larger zero probability. These are implications of the parameter mapping; they do not mean the severity parameters have no role in the positive outcomes.

If exposure is incorporated through a homogeneous Poisson process, its expected event count scales with exposure. For example, tripling the exposure triples $\lambda$, provided the event intensity per unit exposure is constant. The zero probability must then be calculated using the total expected count over that exposure, not the rate for one unit.

Worked Example: Zero and Two-Event Probabilities

A hypothetical annual aggregate-loss model has $N\sim\operatorname{Poisson}(0.8)$. Conditional on an event, the undiscounted claim severity is Gamma with shape 2 and scale $300. Find the probability of no aggregate loss, the probability of a positive aggregate loss, and the probability of exactly two events given a positive aggregate loss. No mortality or interest assumption is involved; the model is for one year and losses are not discounted.

Method and count distribution. The count has parameter $\lambda=0.8$, so $P(N=n)=e^{-0.8}(0.8)^n/n!$. The Gamma severity is positive with probability one. Therefore, $S=0$ exactly when $N=0$; the severity parameters do not enter the zero-probability calculation.

No aggregate loss. $P(S=0)=P(N=0)=e^{-0.8}=0.449328\ldots$, or 0.4493 to four decimal places.

Positive aggregate loss. $P(S>0)=1-P(N=0)=1-0.449328\ldots=0.550672\ldots$, or 0.5507.

Exactly two events, conditional on a positive aggregate. First calculate $P(N=2)=e^{-0.8}(0.8)^2/2=0.449328\ldots\times0.32=0.143785\ldots$. Since $S>0$ is equivalent to $N\geq1$, divide by $P(S>0)$: $P(N=2\mid S>0)=0.143785\ldots/0.550672\ldots=0.2611$, to four decimal places.

Check and common mistake. The unconditional probabilities of $N=0$ and $N\geq1$ add to $0.449328\ldots+0.550672\ldots=1$. A common error is to report $P(N=2)=0.1438$ as the requested conditional probability. The correction is to divide by $P(S>0)$, because the condition restricts the sample space to positive outcomes.

Worked Example: Calculate the Zero Mass from Tweedie Parameters

A hypothetical portfolio segment is modeled with Tweedie mean $\mu=\$400$, dispersion $\phi=40$, and power $p=1.5$. Find the corresponding Poisson intensity and Gamma severity parameters, then calculate the probability of zero aggregate loss and the probability of exactly one event given a positive aggregate. Treat losses as undiscounted annual amounts; no mortality assumption applies.

Method. Use the compound Poisson-Gamma parameter mapping. Here $2-p=0.5$ and $p-1=0.5$. Thus $\lambda=400^{0.5}/[40(0.5)]=20/20=1$. The Gamma shape is $\alpha=0.5/0.5=1$, and its scale is $\theta=40(0.5)400^{0.5}=20(20)=\$400$.

The event count is therefore Poisson with parameter 1, and each event has a positive Gamma$(1,\$400)$ severity. The zero probability is $P(S=0)=e^{-1}=0.367879\ldots$, or 0.3679 to four decimal places. The positive probability is $1-e^{-1}=0.632121\ldots$, or 0.6321.

The unconditional probability of exactly one event is $P(N=1)=e^{-1}(1)^1/1!=e^{-1}=0.367879\ldots$. Conditional on positive aggregate loss, $P(N=1\mid S>0)=e^{-1}/(1-e^{-1})=0.367879\ldots/0.632121\ldots=0.5820$, to four decimal places.

Check and common mistake. The unconditional probabilities for zero and one event are each $e^{-1}$, but the conditional one-event probability is larger than the unconditional one because it excludes the zero-event outcomes. Using $e^{-1}$ directly for the conditional answer omits the required division by $P(S>0)$. The positive aggregate distribution is a mixture over one or more event counts, not simply the single-event Gamma distribution.

Worked Example: Adjust for Total Exposure

A hypothetical insurer models event frequency at 0.6 events per policy-year. Three independent policies each contribute one full year of exposure, and claim severities are independent Gamma with shape 2 and scale $500. Find the probability of no aggregate loss across the three policy-years and the probability of exactly one event conditional on a positive aggregate. Amounts are undiscounted, and no mortality assumption is used.

Method and total count. The total exposure is three policy-years. For independent Poisson counts, the total count is Poisson with parameter equal to the sum of the individual parameters: $\lambda_{\mathrm{total}}=3(0.6)=1.8$. The severity parameters do not affect whether the total is zero.

No aggregate loss. $P(S=0)=e^{-1.8}=0.165298\ldots$, or 0.1653 to four decimal places. Thus $P(S>0)=1-e^{-1.8}=0.834701\ldots$, or 0.8347.

Exactly one event given a positive aggregate. The unconditional one-event probability is $P(N=1)=1.8e^{-1.8}=1.8(0.165298\ldots)=0.297538\ldots$. Hence $P(N=1\mid S>0)=0.297538\ldots/0.834701\ldots=0.3564$, to four decimal places.

Check and common mistake. The zero probability uses the full-exposure intensity, and the zero and positive probabilities add to 1. Using $e^{-0.6}$ would calculate the chance of zero events for only one policy-year, not for the three policy-years combined.

Interpretation and Practical Traps

The compound model gives a precise interpretation of a zero: no events occurred during the modeled exposure. This interpretation depends on the assumptions. In particular, the frequency count must be Poisson with the stated intensity, and event severities must be strictly positive. If event severities can themselves be zero, or if the recorded outcome is modified, a zero aggregate need not mean that the original event count was zero.

For example, a deductible can make the insurer's payment zero even when a loss event occurred. Rounding small payments to zero, incomplete reporting, or a coverage rule that excludes certain losses can also create recorded zeros. Those mechanisms are not part of the basic compound Poisson-Gamma model. Before interpreting a fitted zero probability as a no-event probability, identify whether the modeled response is ground-up loss, insurer payment, or another transformed amount.

Quantity or assumption	Role in the zero probability
Poisson intensity $\lambda$	Determines $P(S=0)=e^{-\lambda}$
Gamma shape $\alpha$ and scale $\theta$	Determine the distribution of positive severity sums; do not alter the zero mass at fixed $\lambda$
Exposure represented in $\lambda$	Must match the period or risk group for which the aggregate is calculated
Strictly positive event severities	Ensure that any event makes the aggregate positive

A frequent exam trap is to treat the aggregate as an ordinary Gamma variable. A Gamma distribution with positive shape and scale has no point mass at zero, so it cannot by itself represent the full compound outcome. Another is to confuse the unconditional probability of one or more events with a conditional probability among positive outcomes. For any $n\geq1$, the correct conditioning calculation is

P(N=n\mid S>0)=\frac{e^{-\lambda}\lambda^n/n!}{1-e^{-\lambda}}.

The denominator is necessary because positive aggregate loss means that the zero-count case has been removed. These distinctions are aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus, where selecting and interpreting a response distribution requires attention to its support and assumptions.

Key takeaway: Under a compound Poisson-Gamma model with strictly positive severities, $S=0$ if and only if $N=0$ with probability one. Therefore the zero mass is $e^{-\lambda}$, while the positive outcomes form a continuous mixture over the possible event counts $N\geq1$.

Check Your Understanding

For each problem, use the count model and exposure stated. Give probabilities to four decimal places.

A one-year model has $N\sim\operatorname{Poisson}(1.2)$ and strictly positive Gamma severities with shape 3 and scale $250. Find the probabilities of zero and positive aggregate loss. (Answer: $P(S=0)=0.3010$; $P(S>0)=0.6990$.)
A Tweedie model has $\mu=900$, $\phi=30$, and $p=1.5$. Find its Poisson intensity, zero probability, and probability of exactly one event given positive aggregate loss. (Answer: $\lambda=2$; $P(S=0)=0.1353$; $P(N=1\mid S>0)=0.3130$.)
A hypothetical event process has intensity 0.25 per month for 12 months, with strictly positive severities. Find the probability of no events and the probability of at least one event over the year. (Answer: $\lambda=3$; $P(N=0)=0.0498$; $P(N\geq1)=0.9502$.)
For $N\sim\operatorname{Poisson}(0.7)$ and strictly positive severities, calculate the probability of exactly two events conditional on a positive aggregate. (Answer: $\frac{e^{-0.7}(0.7)^2/2}{1-e^{-0.7}}=0.2416$.)
A Tweedie model has $\mu=256$, $\phi=16$, and $p=1.25$. Find the Poisson intensity, Gamma shape and scale, and zero probability. (Answer: $\lambda=16/3$, $\alpha=3$, $\theta=\$16$, and $P(S=0)=e^{-16/3}=0.0048$.)

Quantity or assumption	Role in the zero probability
Poisson intensity \(\lambda\)	Determines \(P(S=0)=e^{-\lambda}\)
Gamma shape \(\alpha\) and scale \(\theta\)	Determine the distribution of positive severity sums; do not alter the zero mass at fixed \(\lambda\)
Exposure represented in \(\lambda\)	Must match the period or risk group for which the aggregate is calculated
Strictly positive event severities	Ensure that any event makes the aggregate positive

What You'll Learn