Forecasting Claim-Free Probabilities and Count Thresholds

Why Forecast Claim-Count Probabilities?

A forecast of the expected number of claims does not answer every pricing or portfolio question. An actuary may also need the probability that a policy has no claims, the probability of at least one claim, or the chance that a segment’s count exceeds a specified threshold. When the claim frequency is uncertain, these probabilities must account for both future claim randomness and uncertainty about the underlying frequency.

This tutorial uses the posterior predictive distribution for a Poisson claim count when the claim rate has a Gamma posterior. It focuses on calculating zero-count probabilities and count-threshold probabilities. The calculations are conditional on the experience and information already used to form the posterior.

Model and notation: Let $\lambda$ be a risk’s claim frequency per unit of exposure. After observing experience, suppose its posterior distribution is $\lambda\mid\text{data}\sim\operatorname{Gamma}(\alpha,\beta)$, using the shape-rate convention, with density $f(\lambda)=\beta^\alpha\lambda^{\alpha-1}e^{-\beta\lambda}/\Gamma(\alpha)$ for $\lambda>0$. Here $\alpha>0$ is the shape and $\beta>0$ is the rate. For future exposure $t$, assume $N\mid\lambda\sim\operatorname{Poisson}(t\lambda)$. The predictive probabilities below require these assumptions; they do not apply unchanged if the future count has a different conditional model or the frequency parameter does not have this Gamma posterior.

From a Posterior Rate to a Predictive Count

Conditional on a particular value of $\lambda$, the probability of $n$ future claims is $e^{-t\lambda}(t\lambda)^n/n!$, for $n=0,1,2,\ldots$. Since $\lambda$ is uncertain, the unconditional forecast given the observed data averages this conditional probability over the posterior distribution of $\lambda$.

\begin{aligned} P(N=n\mid\text{data}) &=\int_0^\infty P(N=n\mid\lambda)f(\lambda\mid\text{data})\,d\lambda\\ &=\frac{\beta^\alpha t^n}{n!\,\Gamma(\alpha)} \int_0^\infty \lambda^{\alpha+n-1}e^{-(\beta+t)\lambda}\,d\lambda\\ &=\frac{\Gamma(\alpha+n)}{\Gamma(\alpha)n!} \left(\frac{\beta}{\beta+t}\right)^\alpha \left(\frac{t}{\beta+t}\right)^n. \end{aligned}

The last step uses the Gamma integral: for $c>0$, $\int_0^\infty \lambda^{r-1}e^{-c\lambda}\,d\lambda=\Gamma(r)/c^r$. The resulting predictive count distribution is negative binomial. This derivation explains why the forecast is not generally Poisson: it averages Poisson probabilities over a distribution of possible rates.

Key formula: Define $p=\beta/(\beta+t)$ and $q=t/(\beta+t)$, so $p+q=1$. Then $$ P(N=n\mid\text{data})= \frac{\Gamma(\alpha+n)}{\Gamma(\alpha)n!}p^\alpha q^n. $$ In particular, $P(N=0\mid\text{data})=p^\alpha$ and $P(N\geq1\mid\text{data})=1-p^\alpha$. For an integer threshold $k\geq0$, $P(N>k\mid\text{data})=1-\sum_{n=0}^{k}P(N=n\mid\text{data})$, while $P(N\geq k)=1-\sum_{n=0}^{k-1}P(N=n\mid\text{data})$ for $k\geq1$.

The distinction between “greater than $k$” and “at least $k$” matters: $N>k$ starts at $k+1$, whereas $N\geq k$ includes $k$. For a claim-free probability, set $n=0$ in the predictive formula. The probability of at least one claim is its complement, not the probability of exactly one claim.

A Reliable Calculation Sequence

A useful way to organize threshold calculations is to first put the exposure and posterior rate on compatible units, then compute a short sequence of predictive probabilities, and finally take the appropriate complement. The flow below applies when the future exposure is fixed and known.

Record the posterior and exposure.
Identify $\alpha$, $\beta$, and future exposure $t$. Confirm that $\beta$ and $t$ use matching exposure units.

Calculate the predictive factors.
Set $p=\beta/(\beta+t)$ and $q=t/(\beta+t)$. Check that $p+q=1$.

Calculate the required count probabilities.
Use $p^\alpha$ for zero claims and the predictive mass formula for each count needed in a threshold sum.

Apply the event definition and check.
Take a complement for “at least one” or for an upper tail. Verify that the probabilities being added correspond exactly to the excluded counts.

Worked Probability Calculations

Worked Example: Claim-Free Probability and an Upper Threshold

A hypothetical small commercial segment has posterior claim frequency $\lambda\mid\text{data}\sim\operatorname{Gamma}(3,3)$, in shape-rate form per policy-year. For one future policy-year, calculate the probabilities of zero claims, at least one claim, and more than two claims.

Method and parameters. Use the Poisson-Gamma posterior predictive distribution with $\alpha=3$, $\beta=3$, and $t=1$. Thus $p=3/(3+1)=3/4$ and $q=1/(3+1)=1/4$.

Zero claims. The predictive probability is $p^\alpha$: $P(N=0\mid\text{data})=(3/4)^3=27/64=0.421875$, or $0.4219$ to four decimal places.

At least one claim. Use the complement: $P(N\geq1\mid\text{data})=1-27/64=37/64=0.578125$, or $0.5781$ to four decimal places.

More than two claims. First calculate the probabilities of counts zero, one, and two. For one claim, $P(N=1\mid\text{data})=3(3/4)^3(1/4)=81/256=0.31640625$. For two claims, $P(N=2\mid\text{data})=[\Gamma(5)/(\Gamma(3)2!)](3/4)^3(1/4)^2=6(27/64)(1/16)=162/1024=0.158203125$. Therefore, $P(N>2\mid\text{data})=1-(27/64+81/256+162/1024)=1-919/1024=105/1024=0.1025390625$, or $0.1025$ to four decimal places.

Check and common mistake. The four-decimal probabilities for zero claims and at least one claim sum to $1.0000$. A common error is to use $q^\alpha=(1/4)^3$ for the zero-count probability. In this parameterization, $q$ is the factor raised to the count $n$; when $n=0$, the zero-count probability is $p^\alpha=(3/4)^3$, not $(1/4)^3$.

Worked Example: Forecasting a Threshold Across Multiple Policy-Years

A hypothetical fleet account has posterior frequency $\lambda\mid\text{data}\sim\operatorname{Gamma}(2,4)$, in shape-rate form per policy-year. Forecast the count over two future policy-years. Find the probability of no claims and the probability of more than two claims.

Method and parameters. Here $\alpha=2$, $\beta=4$, and $t=2$ policy-years. The future mean conditional on $\lambda$ is $2\lambda$, so $p=4/(4+2)=2/3$ and $q=2/(4+2)=1/3$.

Zero claims. $P(N=0\mid\text{data})=(2/3)^2=4/9=0.4444$ to four decimal places.

More than two claims. The excluded counts are zero, one, and two. Their probabilities are $P(N=0)=4/9$, $P(N=1)=2(4/9)(1/3)=8/27$, and $P(N=2)=[\Gamma(4)/(\Gamma(2)2!)](4/9)(1/3)^2=3(4/9)(1/9)=4/27$. The sum through two is $4/9+8/27+4/27=12/27+8/27+4/27=24/27=8/9$. Hence $P(N>2\mid\text{data})=1-8/9=1/9=0.1111$ to four decimal places.

Check and common mistake. The probability of more than two claims is not the probability of at least two: the latter also includes $N=2$. The calculation above excludes counts 0, 1, and 2, as required by $N>2$.

Worked Example: A Zero-Claim Forecast With a Larger Exposure

For a hypothetical service-contract portfolio, the posterior frequency is $\lambda\mid\text{data}\sim\operatorname{Gamma}(4,2)$, in shape-rate form per contract-year. Forecast four contract-years of exposure. Calculate the probabilities of zero claims, at least one claim, and more than three claims.

Method and parameters. Take $\alpha=4$, $\beta=2$, and $t=4$. Then $p=2/(2+4)=1/3$ and $q=4/(2+4)=2/3$.

Zero and at least one claim. $P(N=0\mid\text{data})=(1/3)^4=1/81=0.012345679$, or $0.0123$ to four decimal places. Its complement is $P(N\geq1\mid\text{data})=80/81=0.987654321$, or $0.9877$.

More than three claims. Calculate the probabilities of counts zero through three: $P(N=0)=1/81=16/1296$; $P(N=1)=4(1/81)(2/3)=8/243=42.6667/1296$; $P(N=2)=10(1/81)(4/9)=40/729=71.1111/1296$; and $P(N=3)=20(1/81)(8/27)=160/2187=94.8148/1296$. Thus their sum is $224.5926/1296=0.1732968$, and $P(N>3\mid\text{data})=1-0.1732968=0.8267032$, or $0.8267$ to four decimal places. Equivalently, the exact excluded-count sum is $3799/2187$? No: converting carefully to denominator $2187$, the terms are $27/2187+72/2187+120/2187+160/2187=379/2187$. Therefore the exact tail is $1808/2187=0.8267032$.

Check and common mistake. The exact probabilities for zero through three sum to $379/2187$, approximately $0.1733$; adding the tail $1808/2187$ gives $1$. Do not keep the zero-count probability fixed when exposure changes: the exponent base is $\beta/(\beta+t)$, which depends on future exposure.

Interpretation and Exam Traps

The posterior predictive probability of zero claims averages the conditional zero-claim probability $e^{-t\lambda}$ over plausible values of $\lambda$. It is not obtained by simply substituting the posterior mean frequency into a Poisson model. For example, substituting a single rate would produce a Poisson forecast that ignores posterior uncertainty. The Gamma-Poisson mixture instead uses the full posterior distribution, as shown by the integral derivation.

The threshold method is straightforward but depends on careful event wording. For an integer $k$, “exceed $k$” means $N>k$, so subtract the probabilities for $0,1,\ldots,k$ from one. “At least $k$” means $N\geq k$, so subtract only the probabilities for $0,1,\ldots,k-1$. Write the excluded counts before calculating; this prevents off-by-one errors.

Using the wrong Gamma convention: Here $\beta$ is a rate, not a scale. The predictive factor is $\beta/(\beta+t)$, not a formula that treats $\beta$ as a scale.
Mixing exposure units: If $\beta$ is expressed per policy-year, $t$ must be measured in policy-years.
Confusing exact and cumulative probabilities: $P(N=1)$ is not $P(N\geq1)$, and $P(N>k)$ excludes the probability at $k$.
Rounding too early: Keep exact fractions or full precision through the finite sum and round the final probability to four decimal places.

Key takeaway: Under a Gamma posterior with shape $\alpha$ and rate $\beta$, combined with a conditional Poisson count over exposure $t$, set $p=\beta/(\beta+t)$. Then the claim-free probability is $p^\alpha$, the at-least-one probability is $1-p^\alpha$, and any integer count tail follows by summing the excluded predictive probabilities and taking a complement.

Check Your Understanding

For each problem, use the Gamma shape-rate convention and the posterior predictive distribution. Give probabilities to four decimal places.

A hypothetical risk has posterior frequency $\operatorname{Gamma}(2,2)$ per policy-year. For one future policy-year, find the probabilities of zero claims and at least one claim. (Answer: 0.4444 and 0.5556.)
A hypothetical group has posterior frequency $\operatorname{Gamma}(3,6)$ per policy-year and will be observed for three future policy-years. Find the probability of more than one claim. (Answer: 0.4074.)
A hypothetical account has posterior frequency $\operatorname{Gamma}(4,2)$ per policy-year. For one future policy-year, find the probability of at least two claims. (Answer: 0.5391.)
A hypothetical portfolio has posterior frequency $\operatorname{Gamma}(3,2)$ per contract-year. For two future contract-years, find the probabilities of zero claims and more than two claims. (Answer: 0.1250 and 0.3750.)
A hypothetical risk has posterior frequency $\operatorname{Gamma}(2,3)$ per unit of exposure and will have $1.5$ future exposure units. Find the probability of at least two claims. (Answer: 0.1563.)

What You'll Learn