Tweedie Probability of Zero Loss

Why a Tweedie Model Can Predict Zero Loss

A policy can have a positive expected loss and still have no loss during a particular period. In a Tweedie pure premium model with variance power between 1 and 2, aggregate loss has a compound Poisson-Gamma representation: a random number of positive claim amounts are added together. This structure makes the probability of zero aggregate loss especially interpretable. It is the probability that the claim count is zero.

That interpretation depends on the model’s assumptions. In particular, claim amounts must be positive, and the aggregate must be zero only when there are no claims. A Tweedie mean alone does not determine the zero-loss probability; the dispersion and variance power are also needed.

Definition: Let \(S\) be aggregate loss for one observation period, with Tweedie mean \(\mu=E[S]>0\), dispersion \(\phi>0\), and variance power \(p\), where \(1<p<2\). Under the compound Poisson-Gamma representation, \(S=\sum_{j=1}^{N}Y_j\), where \(N\) is Poisson and the \(Y_j\) are independent, positive Gamma claim amounts, independent of \(N\). The zero-loss probability is \(P(S=0)\). Since each \(Y_j>0\), \(S=0\) if and only if \(N=0\).

From the Tweedie Parameters to the Poisson Mean

Use the Tweedie variance convention \(\operatorname{Var}(S)=\phi\mu^p\). For the compound Poisson-Gamma representation, write \(N\sim\operatorname{Poisson}(\lambda)\) and \(Y_j\sim\operatorname{Gamma}(\alpha,\theta)\), using shape \(\alpha\) and scale \(\theta\). Then \(E[Y_j]=\alpha\theta\), and \(\operatorname{Var}(Y_j)=\alpha\theta^2\). Matching this compound model to the Tweedie mean and variance gives the following parameter relationships.

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

The parameter \(\lambda\) is the expected number of claims in the observation period. Because a Poisson random variable with mean \(\lambda\) has \(P(N=0)=e^{-\lambda}\), the Tweedie zero-loss probability is

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

Important condition: This formula applies to the compound Poisson-Gamma Tweedie model with \(1<p<2\) and positive claim amounts. If an individual claim payment can itself be zero, or if zeros can arise from a separate mechanism, then \(S=0\) need not mean \(N=0\); the formula does not describe that additional source of zeros.

Why the Formula Holds

The key step is matching the compound model’s moments to the Tweedie mean and variance. Conditional on \(N\), the mean of \(S\) is \(N\alpha\theta\), and its conditional variance is \(N\alpha\theta^2\). Applying the mean and variance formulas for a random sum gives

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

The Tweedie specification requires \(E[S]=\mu\) and \(\operatorname{Var}(S)=\phi\mu^p\). The compound Poisson-Gamma relationships above satisfy these requirements: substituting \(\alpha\) and \(\theta\) into the mean gives \(\lambda\alpha\theta=\mu\), and substituting all three parameters into the variance gives \(\lambda\alpha(\alpha+1)\theta^2=\phi\mu^p\). In particular, the Poisson mean can be read from the compound representation as the value \(\mu^{2-p}/[\phi(2-p)]\).

Once \(\lambda\) is identified, the zero probability follows directly from the Poisson probability mass function. For \(N\sim\operatorname{Poisson}(\lambda)\), \(P(N=0)=e^{-\lambda}\); and, with positive claim amounts, \(N=0\) is equivalent to \(S=0\). No probability of a zero-valued Gamma claim needs to be added.

Identify the period and model inputs.
Record the Tweedie mean \(\mu\), dispersion \(\phi\), and variance power \(p\), ensuring \(1<p<2\).

Calculate the implied claim-count mean.
Use \(\lambda=\mu^{2-p}/[\phi(2-p)]\). This is an expected count for the specified observation period, not a probability.

Convert the count mean to a zero probability.
Use \(P(S=0)=e^{-\lambda}\), provided claims have positive amounts and there is no separate zero-payment mechanism.

Worked Applications

Worked Example: Finding the Zero Probability From Tweedie Parameters

A hypothetical policy’s aggregate loss for one year is modeled as Tweedie with \(\mu=400\), \(\phi=10\), and \(p=1.5\). Find the expected claim count and probability of zero aggregate loss. Assume the compound Poisson-Gamma representation, positive claim amounts, and no separate zero-payment mechanism. Interest and mortality assumptions are not relevant to this one-year loss calculation.

Method. Calculate the Poisson mean implied by the Tweedie parameters, then use the Poisson probability of no claims. Here \(2-p=0.5\) and \(\mu^{2-p}=\sqrt{400}=20\), so

\lambda=\frac{20}{10(0.5)}=\frac{20}{5}=4.

Thus the expected claim count is 4. The probability of no claims is

P(S=0)=e^{-4}=0.0183156\ldots

Rounded to four decimal places, the zero-loss probability is 0.0183. As a parameter check, the implied Gamma shape is \(\alpha=0.5/0.5=1\), and the scale is \(\theta=10(0.5)\sqrt{400}=100\). The Gamma claim amount therefore has mean \(100\), and \(\lambda E[Y_j]=4(100)=400=\mu\). The calculation avoids treating the expected count of 4 as the zero probability.

Worked Example: Adjusting the Probability for Exposure

A hypothetical portfolio uses a Tweedie model per unit of exposure with \(\mu=100\), \(\phi=50\), and \(p=1.5\). A policy has 2.5 units of exposure. Assume claim frequency is proportional to exposure and the claim severity distribution is unchanged across those units. Find the expected count and the probability of zero claims over the policy’s exposure. Interest and mortality are not part of this loss model.

Method. First calculate the implied expected count per exposure unit. Since \(\sqrt{100}=10\) and \(2-p=0.5\),

\lambda_{\text{unit}}=\frac{10}{50(0.5)}=\frac{10}{25}=0.4.

The policy has 2.5 exposure units, so its expected count is \(2.5(0.4)=1\). Equivalently, independent Poisson counts across equal exposure increments add to a Poisson count with the sum of their means. The probability of no claims across the policy’s exposure is therefore

P(N=0)=e^{-1}=0.3678794\ldots

Rounded to four decimal places, the answer is 0.3679. The implied severity parameters per unit are \(\alpha=1\) and \(\theta=50(0.5)(10)=250\), giving mean claim severity \(250\). The expected aggregate loss over 2.5 units is \(1(250)=250\), also equal to \(2.5(100)\). The exposure adjustment applies to the expected count; applying the one-unit zero probability just once would ignore the policy’s additional exposure.

Worked Example: Inferring Dispersion From a Zero-Loss Probability

A hypothetical one-period Tweedie model has mean aggregate loss \(\mu=16\), variance power \(p=1.5\), and observed model zero-loss probability \(P(S=0)=0.2500\). Infer the dispersion \(\phi\). Assume the compound Poisson-Gamma model with positive claim amounts. Interest and mortality assumptions are not relevant.

Method. Take natural logarithms of \(P(S=0)=e^{-\lambda}\) to obtain the implied Poisson mean:

\lambda=-\ln(0.2500)=\ln(4)=1.386294361\ldots

For \(p=1.5\), \(\mu^{2-p}=\sqrt{16}=4\) and \(2-p=0.5\). Rearranging \(\lambda=\mu^{2-p}/[\phi(2-p)]\) gives

\phi=\frac{4}{1.386294361\ldots(0.5)} =\frac{4}{0.693147181\ldots} =5.770780\ldots

To four significant figures, \(\phi=\mathbf{5.771}\). As a check, substituting the unrounded dispersion back into the count formula gives \(\lambda=1.386294361\ldots\), and \(e^{-\lambda}=0.2500\). The implied Gamma shape is \(1\), and the scale is \(\theta=5.770780\ldots(0.5)(4)=11.54156\ldots\); hence \(\lambda\alpha\theta=16\), as required. The common mistake is to round \(\lambda\) too early or to use \(\mu\) rather than \(\mu^{2-p}\) in the dispersion calculation.

Interpreting Frequency, Exposure, and Dispersion

For fixed \(\mu\) and \(p\), increasing \(\phi\) decreases \(\lambda\), and therefore increases \(e^{-\lambda}\). For fixed \(\phi\) and \(p\), increasing \(\mu\) increases \(\lambda\), so the probability of zero loss falls. These comparisons hold with the other stated parameters fixed; they are not claims that changing a rating variable in a fitted model affects only one parameter.

Exposure gives the count probability a direct frequency interpretation. If the expected claim frequency per exposure unit is \(\nu\), and the count follows a Poisson process with constant rate across exposure, then over \(e\) exposure units the expected count is \(e\nu\), and the probability of no claims is \(e^{-e\nu}\). In a Tweedie model, \(\nu\) can be obtained from the per-unit Tweedie parameters using \(\nu=\mu^{2-p}/[\phi(2-p)]\). This relates the Tweedie zero probability to frequency without requiring a separate severity calculation.

The model’s variance power also matters. It affects the implied count mean through both the exponent \(2-p\) and the factor \(2-p\) in the denominator. Therefore, it is not valid to infer the zero probability from the mean and dispersion while ignoring \(p\). Nor should the Tweedie dispersion be confused with the Poisson mean: \(\phi\) is a variance parameter, while \(\lambda\) is an expected count.

Common Exam Trap

A frequent error is to report \(\lambda\) as the probability of zero loss, or to use \(1-\lambda\). Neither is correct. A Poisson mean is not a probability, and the probability of zero claims is an exponential function of that mean.

Correction: First calculate \(\lambda=\mu^{2-p}/[\phi(2-p)]\); then calculate \(P(S=0)=e^{-\lambda}\). For example, if \(\lambda=1\), the zero probability is \(e^{-1}=0.3679\), not \(1\) and not \(0\).

A second check is conceptual: the probability must lie between 0 and 1, and it decreases as the implied Poisson mean increases. If a calculation produces a negative probability, a probability greater than 1, or a zero probability that rises when \(\lambda\) rises, revisit the calculation.

Check Your Understanding

Assume the compound Poisson-Gamma representation with positive claim amounts wherever a Tweedie model is specified. Round probabilities to four decimal places and rates to four significant figures.

A one-period model has \(\mu=625\), \(\phi=25\), and \(p=1.5\). Find its expected claim count and zero-loss probability. (Answer: \(\lambda=2\), \(P(S=0)=0.1353\))
A model has \(\mu=81\), \(\phi=9\), and \(p=1.5\). Find the implied Poisson mean and probability of zero loss. (Answer: \(\lambda=2\), \(P(S=0)=0.1353\))
A Tweedie model has \(\mu=100\), \(p=1.25\), and implied expected claim count \(\lambda=2\). Find \(\phi\) and the zero-loss probability. (Answer: \(\phi=21.08\), \(P(S=0)=0.1353\))
A hypothetical policy has 1.8 exposure units and a Poisson expected claim frequency of 0.3 per exposure unit. Under a constant-rate Poisson count model, find its expected claim count and probability of no claims. (Answer: \(\lambda=0.54\), \(P(N=0)=0.5827\))
A model has \(\mu=64\), \(p=1.5\), and \(P(S=0)=0.5000\). Find the implied dispersion, using the natural logarithm. (Answer: \(\phi=23.08\), to four significant figures)

What You'll Learn