Zero Claims Under a Gamma-Poisson Prior

Why Zero Claims Do Not Mean Zero Risk

A policyholder may have no claims during a short observation period and still have a positive expected claim frequency. The key distinction is between the observed count and the unobserved rate that generated it. In a Bayesian frequency model, the prior distribution represents uncertainty about that rate; observing zero claims updates the distribution but does not ordinarily concentrate it at zero.

This tutorial uses a Gamma prior for a Poisson claim rate. It focuses on the consequences of zero claims: the posterior rate, the role of observed exposure, and probabilities for future claims. All scenarios below are hypothetical. Since these are claim-frequency calculations, no interest rate or mortality assumption is involved.

Definition: Let $\lambda$ be a risk's claim frequency per unit of exposure. Conditional on $\lambda$, the number of claims $N_e$ in exposure $e$ has a Poisson distribution with mean $\lambda e$: $$ N_e\mid\lambda\sim\operatorname{Poisson}(\lambda e). $$ A Gamma prior in shape-rate form is denoted $\lambda\sim\operatorname{Gamma}(\alpha,\beta)$, with $\alpha>0$, $\beta>0$, density $$ \pi(\lambda)=\frac{\beta^\alpha}{\Gamma(\alpha)}\lambda^{\alpha-1}e^{-\beta\lambda},\qquad \lambda>0. $$ Its mean is $\alpha/\beta$. Here $\beta$ is the rate parameter of the Gamma distribution, not its scale.

Exposure must use the same unit in the likelihood and in future projections. For example, if $\lambda$ is claims per policy-year, $e$ and any future exposure $h$ are measured in policy-years. The Poisson model assumes a constant rate over the exposure considered, conditional on $\lambda$, and independent increments across nonoverlapping periods. The Gamma prior describes variation or uncertainty in the rate across risks. These assumptions are needed for the formulas below; they may not suit data with changing risk conditions or other dependence not represented by the model.

Updating After Zero Claims

If the observed exposure is $e$ and the observed count is zero, its likelihood as a function of $\lambda$ is $\Pr(N_e=0\mid\lambda)=e^{-\lambda e}$. Multiplying this likelihood by the Gamma prior leaves a Gamma-shaped density. The resulting posterior is $\operatorname{Gamma}(\alpha,\beta+e)$, so the posterior mean frequency is $\alpha/(\beta+e)$. The shape parameter remains $\alpha$; the zero count adds exposure but no claim events.

\lambda\mid(N_e=0)\sim\operatorname{Gamma}(\alpha,\beta+e), \qquad E[\lambda\mid N_e=0]=\frac{\alpha}{\beta+e}.

The posterior mean is positive for every $\alpha>0$ and finite $e$. As zero-claim exposure grows, this mean decreases, but it does not become zero. This is the central protection against assigning a zero indicated frequency merely because no claim happened to be observed.

Interpretation: The prior mean is $\mu=\alpha/\beta$. If $\beta$ is viewed as prior pseudo-exposure and $\alpha$ as prior pseudo-claims, then zero observed claims over exposure $e$ give posterior mean $\alpha/(\beta+e)$. Greater zero-claim exposure lowers the rate, while the prior pseudo-claims keep the estimate positive.

Justification of the Posterior Update

Bayes' rule says that the posterior density is proportional to the likelihood multiplied by the prior density. With zero claims, the product is

e^{-\lambda e}\frac{\beta^\alpha}{\Gamma(\alpha)} \lambda^{\alpha-1}e^{-\beta\lambda} \ \propto\ \lambda^{\alpha-1}e^{-(\beta+e)\lambda}.

The Gamma density with shape $\alpha$ and rate $\beta+e$ is proportional to exactly this expression. Normalizing it therefore gives the stated posterior. This also explains why the rate parameter increases by the observed exposure: the likelihood contributes another exponential factor $e^{-\lambda e}$. Because the observed count is zero, there is no factor of $\lambda$ to increase the shape.

Predicting Future Claims

Let $M_h$ be the claim count over future exposure $h$. Conditional on $\lambda$, $M_h$ is Poisson with mean $\lambda h$. After observing zero claims over exposure $e$, the posterior distribution of $\lambda$ is Gamma with shape $\alpha$ and rate $r=\beta+e$. Averaging over that posterior gives a Gamma-Poisson predictive distribution:

\Pr(M_h=k\mid N_e=0) = \frac{\Gamma(\alpha+k)}{\Gamma(\alpha)\,k!} \left(\frac{r}{r+h}\right)^\alpha \left(\frac{h}{r+h}\right)^k, \qquad k=0,1,2,\ldots, \quad r=\beta+e.

In particular, the future probability of no claims is $(r/(r+h))^\alpha$, and the probability of at least one claim is its complement. The predictive mean is $h\alpha/r$, equal to future exposure multiplied by the posterior mean frequency. These are predictive quantities: they account for uncertainty about $\lambda$, rather than treating the posterior mean as a known, fixed rate.

\Pr(M_h=0\mid N_e=0)=\left(\frac{r}{r+h}\right)^\alpha, \qquad \Pr(M_h\geq 1\mid N_e=0)=1-\left(\frac{r}{r+h}\right)^\alpha, \qquad E[M_h\mid N_e=0]=\frac{h\alpha}{r}.

The predictive distribution is generally more variable than a Poisson distribution with the same mean because it allows the rate to vary according to its posterior distribution. The exact-count formula is useful when a question asks for two or more future claims; the complement formula is usually the shortest route to the probability of at least one.

Worked Example: A Short Zero-Claim Period

A hypothetical small commercial risk has a prior claim frequency $\lambda\sim\operatorname{Gamma}(3,6)$, in shape-rate form, where exposure is measured in policy-years. It has zero claims over $e=4$ policy-years. Find its posterior mean frequency and the probability of at least one claim over a future $h=2$ policy-years. The conditional count model is Poisson with mean $\lambda$ times exposure. No interest or mortality assumption is used.

Method. Apply the Gamma-Poisson conjugate update, then use the posterior predictive probability of zero future claims and take its complement. This avoids treating the observed frequency $0/4$ as the true claim rate.

Posterior. The prior parameters are $\alpha=3$ and $\beta=6$. After exposure $e=4$ with zero claims, the posterior is $\operatorname{Gamma}(3,6+4)=\operatorname{Gamma}(3,10)$. Its mean is $$ E[\lambda\mid N_4=0]=\frac{3}{10}=0.3 $$ claims per policy-year.

Future probability. Here $r=10$ and $h=2$, so $$ \Pr(M_2=0\mid N_4=0) =\left(\frac{10}{10+2}\right)^3 =\left(\frac{5}{6}\right)^3 =\frac{125}{216} =0.5787. $$ Therefore, $$ \Pr(M_2\geq1\mid N_4=0)=1-\frac{125}{216} =\frac{91}{216}=0.4213. $$ Probabilities are rounded to four decimal places. As a check, the predictive mean count is $2(3/10)=0.6$, consistent with a positive future risk despite the zero observed count.

Worked Example: How Additional Zero-Claim Exposure Changes the Estimate

A hypothetical class of equipment risks has $\lambda\sim\operatorname{Gamma}(4,8)$, with $\lambda$ measured per policy-year and conditional counts Poisson. Compare a risk with zero claims over $e=2$ policy-years to one with zero claims over $e=8$ policy-years. For each, calculate the posterior mean frequency and the probability of at least one claim in the next policy-year. There is no interest or mortality assumption.

Method. Update each risk separately using its own exposure. This illustrates why the count of zero alone is insufficient: the amount of exposure with no claims also matters.

Two policy-years with zero claims. The posterior rate parameter is $8+2=10$, giving posterior $\operatorname{Gamma}(4,10)$ and mean frequency $4/10=0.4$. For $h=1$, $$ \Pr(M_1\geq1\mid N_2=0) =1-\left(\frac{10}{11}\right)^4 =1-\frac{10000}{14641} =\frac{4641}{14641} =0.3170. $$

Eight policy-years with zero claims. The posterior rate parameter is $8+8=16$, giving posterior $\operatorname{Gamma}(4,16)$ and mean frequency $4/16=0.25$. For the next policy-year, $$ \Pr(M_1\geq1\mid N_8=0) =1-\left(\frac{16}{17}\right)^4 =1-\frac{65536}{83521} =\frac{17985}{83521} =0.2153. $$

The longer claim-free experience produces a lower posterior mean and a lower predictive probability. Both remain positive. Probabilities are rounded to four decimal places. The calculation avoids the mistake of treating both risks identically just because each has reported zero claims.

Worked Example: Predicting an Exact Future Count

A hypothetical delivery fleet has a prior frequency $\lambda\sim\operatorname{Gamma}(2,4)$, with exposure measured in vehicle-years. Conditional on $\lambda$, claims follow a Poisson model. The fleet has zero claims over $e=6$ vehicle-years. Find the probability of exactly two claims in the next $h=3$ vehicle-years and the probability of at least one claim. No interest or mortality assumption applies.

Method. First update the posterior; then use the Gamma-Poisson predictive count formula. The common mistake avoided here is using a Poisson distribution at the posterior mean as though the frequency were known exactly.

Posterior. The updated distribution is $\operatorname{Gamma}(2,4+6)=\operatorname{Gamma}(2,10)$, so $r=10$ and the posterior mean frequency is $2/10=0.2$ claims per vehicle-year.

Exactly two claims. Substitute $\alpha=2$, $k=2$, $r=10$, and $h=3$: $$ \begin{aligned} \Pr(M_3=2\mid N_6=0) &=\frac{\Gamma(4)}{\Gamma(2)\,2!} \left(\frac{10}{13}\right)^2 \left(\frac{3}{13}\right)^2\\ &=3\left(\frac{100}{169}\right)\left(\frac{9}{169}\right) =\frac{2700}{28561} =0.0945. \end{aligned} $$

At least one claim. The zero-claim probability and its complement are $$ \Pr(M_3=0\mid N_6=0)=\left(\frac{10}{13}\right)^2 =\frac{100}{169}, \qquad \Pr(M_3\geq1\mid N_6=0)=\frac{69}{169}=0.4083. $$

The predictive mean is $3(2/10)=0.6$ claims. The calculated probabilities are rounded to four decimal places. The exact-count result is positive, and the at-least-one probability exceeds the exactly-two probability, as expected because it includes every positive count.

Credibility Interpretation and an Exam Trap

Write the prior mean as $\mu=\alpha/\beta$, so $\alpha=\beta\mu$. After zero claims, the posterior mean can be expressed as $$ \frac{\alpha}{\beta+e} =\frac{\beta}{\beta+e}\mu+\frac{e}{\beta+e}\cdot 0. $$ This is a credibility-weighted average of the prior mean and the observed claim rate $0/e=0$, for $e>0$. The weight $e/(\beta+e)$ assigned to the observed rate increases with exposure. The prior strength $\beta$, interpreted as pseudo-exposure in this representation, prevents limited zero-claim experience from overwhelming the prior. This expression describes the posterior mean rate; predictive probabilities still require integrating over the posterior, as above.

Common exam trap: Calculating the observed rate $0/e=0$ and reporting it as the posterior rate or future premium ignores the prior and the uncertainty in $\lambda$. For example, with $\operatorname{Gamma}(3,6)$ and zero claims over four policy-years, the incorrect estimate is $0/4=0$. The correct posterior is $\operatorname{Gamma}(3,10)$, with mean $3/10=0.3$ claims per policy-year. Keep the exposure units consistent, update the Gamma rate by $e$, and use the posterior predictive formula when asked for a future claim probability.

These calculations matter whenever a rate estimate is based on sparse or claim-free experience. They give a principled distinction between “no claim was observed” and “the risk has no chance of a claim.” In pricing or risk assessment, the posterior mean may be used as an indicated frequency, while predictive probabilities answer questions about future counts and explicitly reflect rate uncertainty.

Check Your Understanding

For each hypothetical risk, use the stated shape-rate Gamma prior and conditional Poisson model. Exposure is measured in the same units for past and future periods; no interest or mortality assumption is involved.

A risk has prior $\operatorname{Gamma}(2,5)$, zero claims over exposure $3$, and future exposure $1$. Find the posterior mean frequency and the probability of at least one future claim. (Answer: posterior mean $2/8=0.25$; probability $17/81=0.2099$.)
A risk has prior $\operatorname{Gamma}(3,6)$, zero claims over exposure $6$, and future exposure $2$. Find the posterior mean frequency and the probability of at least one future claim. (Answer: posterior mean $3/12=0.25$; probability $127/343=0.3703$.)
A prior has mean frequency $0.4$ and pseudo-exposure strength $\beta=10$. There are zero claims over exposure $5$. Find the posterior mean and the probability of at least one claim in future exposure $1$. (Answer: $\alpha=4$, posterior mean $4/15=0.2667$; probability $1-(15/16)^4=14911/65536=0.2275$.)
A risk has prior $\operatorname{Gamma}(2,4)$, zero claims over exposure $4$, and future exposure $2$. Find the probability of exactly one future claim and the probability of at least one. (Answer: exactly one $=2(8/10)^2(2/10)=0.2560$; at least one $=1-(8/10)^2=0.3600$.)
A risk has prior $\operatorname{Gamma}(5,10)$. Compare zero claims over exposure $10$ with zero claims over exposure $20$. For each case, give the posterior mean and the probability of at least one claim over future exposure $1$. (Answer: exposure $10$: mean $5/20=0.25$, probability $1-(20/21)^5=0.2165$; exposure $20$: mean $5/30=0.1667$, probability $1-(30/31)^5=0.1512$.)

What You'll Learn