Tweedie Fit and Aggregate Loss Shape

What Aggregate Loss Shape Does the Model Imply?

A pure premium response is an aggregate loss over a specified policy and period. It may be zero when no covered claim occurs, but it cannot be negative when it represents gross claim payments. Those features suggest a model with a point mass at zero and a continuous distribution for positive losses. A compound Poisson-Gamma model gives a precise version of that structure.

Compatibility is more than checking that an average or variance looks reasonable. The model makes linked predictions about the probability of zero, the distribution of positive amounts, and the relationship between mean and variance. If these predictions conflict substantially with the observed response, the model may be inappropriate even if its mean prediction seems plausible.

Definition: In a compound Poisson-Gamma model, the claim count $N$ has a Poisson distribution with mean $\lambda$. Conditional on $N$, claim severities $X_1,\ldots,X_N$ are independent and identically distributed Gamma random variables, independent of $N$, with shape $\alpha>0$ and scale $\theta>0$. Aggregate loss is $S=\sum_{j=1}^{N}X_j$, with $S=0$ when $N=0$. The Gamma scale parameterization has $E[X_j]=\alpha\theta$ and $\operatorname{Var}(X_j)=\alpha\theta^2$.

This setup assumes a homogeneous claim process over the modeled period, positive claim severities, and no separate source of zero payments such as coverage exclusions or reporting failures. The claims are assumed independent conditional on the model parameters. A response that includes recoveries or accounting adjustments can be negative, so it does not have the support of this model without further transformation or a different modeling framework.

Structural Implications for Zeros and Positive Losses

The aggregate is exactly zero if and only if the claim count is zero, because every modeled severity is strictly positive. Thus the zero probability is determined by the Poisson frequency parameter. For positive aggregate losses, the number of claims is at least one. Conditional on exactly $n$ claims, the sum of the $n$ independent Gamma severities has a Gamma distribution with shape $n\alpha$ and the same scale $\theta$. Consequently, conditional on $S>0$, the aggregate distribution is a mixture over possible positive claim counts—not generally a single Gamma distribution.

P(S=0)=P(N=0)=e^{-\lambda}, \qquad P(N=n\mid S>0)=\frac{e^{-\lambda}\lambda^n/n!}{1-e^{-\lambda}}, \quad n=1,2,\ldots

This mixture gives the positive part a potentially long right tail: larger claim counts produce sums with larger means and greater spread. When $\lambda$ is small, most positive aggregate losses may come from one claim, and their distribution can resemble the severity distribution. As $\lambda$ increases, multi-claim sums matter more. It is therefore a mistake to require the positive aggregate observations to follow exactly the same Gamma distribution as individual claim severity.

Result: If the severity has Gamma shape $\alpha$ and scale $\theta$, then $E[S]=\lambda\alpha\theta$ and $\operatorname{Var}(S)=\lambda\alpha(\alpha+1)\theta^2$. Also, $E[S\mid S>0]=\lambda\alpha\theta/(1-e^{-\lambda})$. The last identity follows because $S=0$ contributes nothing to the unconditional mean: $E[S]=P(S>0)E[S\mid S>0]$.

For the variance, condition first on $N$. The conditional mean and variance are $N\alpha\theta$ and $N\alpha\theta^2$, respectively. The law of total variance gives

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

The two variance terms have different sources: variation in severities and variation in claim count. This derivation also shows why using only the severity variance times the expected claim count understates aggregate variance: it omits the variability contributed by the random count.

Connection to the Tweedie Model

For a compound Poisson-Gamma distribution, the corresponding Tweedie variance power $p$ lies strictly between 1 and 2. Write the aggregate mean as $\mu>0$, and define the dispersion parameter $\phi>0$ by $\operatorname{Var}(S)=\phi\mu^p$. With the Gamma scale convention used above, the parameters are related as follows:

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

These equations are not three unrelated fitting choices. They are the parameter conversion that makes the compound Poisson-Gamma mean, variance, and variance power agree. For example, substituting the frequency and severity parameters into the aggregate mean gives $\lambda\alpha\theta=\mu$. Substitution into the aggregate variance gives $\phi\mu^p$. The implied zero probability is then $e^{-\lambda}$. A fitted Tweedie model therefore entails a specific zero probability as well as a specific positive-loss shape.

Worked Example: Check a Tweedie Parameter Conversion

A hypothetical aggregate-loss model has mean $\mu=1$ dollar, dispersion $\phi=2$, and variance power $p=1.5$. Under the compound Poisson-Gamma representation, calculate the severity shape and scale, claim-count mean, zero probability, aggregate variance, and mean conditional on positive aggregate loss.

Method and assumptions. Use the compound Poisson-Gamma conversion, with independent Poisson claim count and Gamma severities, and assume the stated mean and dispersion apply to this aggregate period.

Severity shape. $\alpha=(2-1.5)/(1.5-1)=0.5/0.5=1$.

Severity scale. $\theta=\phi(p-1)\mu^{p-1}=2(0.5)(1^{0.5})=1$ dollar. The individual-severity mean is $\alpha\theta=1$ dollar.

Claim-count mean. $\lambda=\mu^{2-p}/[\phi(2-p)]=1^{0.5}/[2(0.5)]=1$.

Zero probability. $P(S=0)=e^{-1}=0.3679$, to four decimals.

Aggregate variance. $\operatorname{Var}(S)=\lambda\alpha(\alpha+1)\theta^2=1(1)(2)(1^2)=2$ dollars squared. The Tweedie formula independently gives $\phi\mu^p=2(1^{1.5})=2$, so the variance agrees.

Conditional positive mean. $E[S\mid S>0]=1/(1-e^{-1})=1/0.6321206=1.58198$, or $\$1.58$ to the nearest cent.

Check and common mistake. The unconditional mean is $\lambda\alpha\theta=1$ dollar, and multiplying the conditional mean by $P(S>0)$ gives $1.58198(0.6321206)=1.0000$ dollar. Dividing by $P(S>0)$ is essential: the conditional positive mean is not the unconditional mean.

Assessing Observed Data

A practical compatibility assessment starts with the response definition and exposure period. Then compare the observed data with the model’s structural consequences. The goal at this stage is a reasoned diagnostic, not a declaration that a model is true. Sampling variation, heterogeneity between policies, and data-processing choices can all affect the observed pattern.

Check the support.
Confirm that aggregate responses are nonnegative and that zero represents no modeled payment. Investigate negative values, caps, and separately recorded adjustments before fitting this distribution.

Check the zero mechanism.
Compare the observed zero proportion with $e^{-\lambda}$ implied by the fitted frequency or Tweedie parameters. Interpret the comparison in light of the number of observations and the policy exposure structure.

Inspect positive amounts.
Look for a nonnegative, right-skewed pattern consistent with a mixture of Gamma sums. Consider whether there is a plausible severity tail and whether the observed positive values show a shape the model cannot represent.

Check linked moments and context.
Compare the model’s mean and variance implications with the data, allowing for sampling variability. Revisit assumptions if claim frequency, severity, or exposure differs materially across observations.

Zero-frequency comparisons require care. If each observation has a different exposure or fitted mean, its implied zero probability may differ as well; comparing one pooled empirical proportion with a single $e^{-\lambda}$ can then be misleading. Likewise, an unusual number of zeros may indicate a model mismatch, but can also reflect small samples or unmodeled groups with different risk levels.

Worked Example: Compare an Observed Zero Rate With the Model

In a hypothetical portfolio, 400 equal-exposure policy-periods are modeled with a Poisson claim-count mean $\lambda=0.4$ per period. There are 280 observed zero-loss periods. Calculate the model-implied zero rate and expected number of zero periods, then compare the observed rate using a simple standardized diagnostic.

Method and assumptions. Under the fitted model, each period has zero probability $q=e^{-0.4}$, and the 400 zero indicators are treated as independent Bernoulli observations with that common probability.

Implied zero rate. $q=e^{-0.4}=0.670320$, or $0.6703$ to four decimals.

Expected zero count. $400q=400(0.670320)=268.128$, approximately 268.1 periods.

Observed rate and standard error. The observed rate is $280/400=0.7000$. Under the model, the standard error of this rate is $\sqrt{q(1-q)/400}=\sqrt{(0.670320)(0.329680)/400}=0.0235$.

Standardized difference. $(0.7000-0.670320)/0.0235=1.26$, approximately. The observed rate is about 1.26 model-based standard errors above the implied rate. This difference alone is not compelling evidence of incompatibility; it is a descriptive check, not a complete formal test.

Check and common mistake. The fitted zero probability is $e^{-\lambda}$, not $1-e^{-\lambda}$; the latter is the probability of a positive loss. The calculation uses the common exposure assumption stated in the problem.

Why the Positive Part Is a Mixture

Suppose severity is Gamma with shape $\alpha$ and scale $\theta$. For exactly one claim, the aggregate is Gamma with shape $\alpha$ and scale $\theta$. For exactly two claims, it is Gamma with shape $2\alpha$ and scale $\theta$. More generally, the shape is $n\alpha$ given $N=n$. Conditioning only on $S>0$ combines these distributions with the conditional count probabilities. Therefore, fitting or judging the model by demanding that all positive aggregate losses follow one Gamma distribution imposes the wrong condition.

Worked Example: Interpret the Positive-Loss Shape

A hypothetical model has $\lambda=0.3$, Gamma severity shape $\alpha=2$, and scale $\theta=\$100$. Find the probability that a positive aggregate loss comes from exactly one claim, the conditional mean aggregate loss given a positive loss, and the component distributions for one and two claims.

Method and assumptions. Use the compound Poisson-Gamma model with independent claims and the stated severity parameterization.

Positive probability. $P(S>0)=1-e^{-0.3}=1-0.7408182=0.2591818$.

One-claim share among positive losses. $P(N=1\mid S>0)=P(N=1)/P(N>0)=[0.3e^{-0.3}]/[1-e^{-0.3}]$. The numerator is $0.3(0.7408182)=0.2222455$, so the ratio is $0.2222455/0.2591818=0.8575$, to four decimals.

Conditional positive mean. The unconditional mean is $\lambda\alpha\theta=0.3(2)(\$100)=\$60$. Thus $E[S\mid S>0]=\$60/0.2591818=\$231.50$, to the nearest cent.

Component distributions. Given one claim, $S$ is Gamma with shape $2$ and scale $\$100$, with mean $2(\$100)=\$200$. Given two claims, it is Gamma with shape $4$ and scale $\$100$, with mean $4(\$100)=\$400$. Positive aggregate losses combine these and all higher-count components.

Check and common mistake. The one-claim conditional probability is about 0.8575, so the one-claim component dominates but does not account for every positive observation. Treating the entire positive sample as exactly Gamma with shape 2 would ignore the remaining multi-claim probability.

Common Compatibility Traps

Using the wrong zero probability. $e^{-\lambda}$ is the probability of no claims and hence zero aggregate loss under the model. $1-e^{-\lambda}$ is the probability of at least one claim.
Confusing individual severity with positive aggregate loss. Positive aggregates can contain multiple claims; their conditional mean need not equal $\alpha\theta$.
Dropping claim-count variation from the variance. The correct aggregate variance is $\lambda\alpha(\alpha+1)\theta^2$, not merely $\lambda\alpha\theta^2$.
Reading model compatibility as proof of the assumptions. A plausible zero rate and positive shape do not establish that the count is Poisson, severities are Gamma, or observations are homogeneous.
Ignoring the response definition. A net response that can be negative, or a response with many administrative zeros unrelated to claim occurrence, may not have the stated compound Poisson-Gamma structure.

Key takeaway: A compound Poisson-Gamma fit has linked structural implications: nonnegative aggregate losses, a zero mass of $e^{-\lambda}$, and a positive-loss distribution that is a count-weighted mixture of Gamma sums. Assess all three features together rather than judging the model from its mean alone.

Check Your Understanding

For each hypothetical situation, use the compound Poisson-Gamma assumptions stated in the question. Give probabilities to four decimals and money to the nearest cent when applicable.

A model has $\lambda=0.2$, severity shape $\alpha=2$, and scale $\theta=\$50$. Calculate the zero probability, aggregate mean, aggregate variance, and mean conditional on a positive aggregate loss. (Answer: $P(S=0)=0.8187$, $E[S]=\$20.00$, $\operatorname{Var}(S)=3{,}000$ dollars squared, and $E[S\mid S>0]=\$110.33$.)
A Tweedie aggregate model has $\mu=9$, $\phi=1.5$, and $p=1.6$. Find the implied Gamma shape, Poisson mean, and zero probability. (Answer: $\alpha=0.6667$, $\lambda\approx4.0137$, and $P(S=0)\approx0.0180$.)
A fitted model assigns $\lambda=0.7$ to each of 500 independent equal-exposure periods, but 220 periods have zero loss. Calculate the model’s expected zero count and the standardized difference between observed and expected zero rates, using the model-based binomial standard error. (Answer: implied zero rate $0.4966$, expected count $248.3$, standard error about $0.0224$, and standardized difference about $-2.53$.)
A model has $\lambda=1$, $\alpha=1$, and $\theta=\$4$. Calculate the zero probability, aggregate mean, aggregate variance, conditional positive mean, and the probability that a positive loss comes from exactly one claim. (Answer: $P(S=0)=0.3679$, $E[S]=\$4.00$, $\operatorname{Var}(S)=32$ dollars squared, $E[S\mid S>0]=\$6.33$, and $P(N=1\mid S>0)=0.5820$.)
A model has $\lambda=0.3$, $\alpha=2$, and $\theta=\$100$. Is the positive aggregate distribution exactly Gamma with shape 2 and scale $\$100$? Calculate the one-claim share among positive losses and the conditional positive mean to support your answer. (Answer: No; the one-claim share is $0.8575$, while multi-claim outcomes also contribute, and $E[S\mid S>0]=\$231.50$.)

What You'll Learn