Why Zero Observations Still Leave Future Risk
A risk with no reported claims in its observed period may still have a substantial forecast for future claims. The observed count is only one realization of a random process; it does not establish that the underlying claim frequency is zero. In a Bayesian model, the forecast combines that experience with the prior distribution for the risk’s unknown frequency.
This distinction matters particularly for short exposure periods and sparse data. A zero over a small exposure updates the frequency estimate less than a zero over a long exposure. The prior also matters: two prior distributions with the same mean can assign different strength to that mean and therefore produce different forecasts after the same zero-claim observation. The calculations below are aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.
Poisson-Gamma Model After Zero Claims
Suppose a risk has an unknown claim rate \(\lambda\) per unit of exposure. Conditional on \(\lambda\), assume claims follow a Poisson process: the count \(N\) over exposure \(e\) is Poisson with mean \(\lambda e\), and counts in nonoverlapping periods are conditionally independent. Exposure units must be consistent; for example, if \(\lambda\) is claims per vehicle-year, then \(e\) is measured in vehicle-years.
Represent prior uncertainty about the rate with a Gamma distribution in the shape-rate parameterization. Thus, \(\lambda\sim\operatorname{Gamma}(\alpha,\beta)\) has density proportional to \(\lambda^{\alpha-1}e^{-\beta\lambda}\), with mean \(\alpha/\beta\) and variance \(\alpha/\beta^2\). Here \(\alpha>0\) and \(\beta>0\).
For zero observed claims, the posterior mean rate is \(\alpha/(\beta+e_0)\), not zero. The future predictive mean is that rate multiplied by future exposure. The predictive variance is larger than the mean because it includes both Poisson variation and uncertainty about the rate.
Why the Predictive Distribution Is Negative Binomial
Let \(\alpha_*=\alpha+n\) and \(\beta_*=\beta+e_0\), the posterior shape and rate after observing \(n\) claims over exposure \(e_0\). Conditional on \(\lambda\), the future count \(M\) over exposure \(e_1\) is Poisson. Its unconditional probability is the integral of that Poisson probability against the posterior Gamma density.
To see the key step, combine the exponentials in the integrand to obtain \(e^{-(\beta_*+e_1)\lambda}\). The remaining integral is a Gamma integral: \(\int_0^\infty \lambda^{\alpha_*+m-1}e^{-c\lambda}\,d\lambda =\Gamma(\alpha_*+m)/c^{\alpha_*+m}\), for \(c>0\). Substituting \(c=\beta_*+e_1\) and collecting the powers gives the displayed probability formula. The mixing over \(\lambda\) creates extra variation, which is why this predictive distribution is not simply a Poisson distribution at the posterior mean rate.
Worked Predictive Calculations
Worked Example: Calculating the Distribution After Zero Claims
A hypothetical account has a prior claim rate \(\lambda\sim\operatorname{Gamma}(2,3)\), where the parameters are shape and rate and \(\lambda\) is measured per exposure unit. It has zero claims over 2 units of exposure. Calculate the posterior predictive probabilities of 0, 1, and 2 claims over the next 4 exposure units, the probability of at least 3 claims, and the predictive mean and variance.
Method. Apply the Gamma update, then use the negative-binomial predictive distribution. This avoids treating the observed zero count as if it were a known zero rate.
The posterior is \(\operatorname{Gamma}(2+0,3+2)=\operatorname{Gamma}(2,5)\). Thus \(\alpha_*=2\), \(\beta_*=5\), and \(e_1=4\). The probability formula becomes \(P(M=m)=\frac{\Gamma(2+m)}{\Gamma(2)m!}(5/9)^2(4/9)^m\).
For zero claims, \(P(M=0)=(5/9)^2=25/81=0.3086\). For one claim, the coefficient is \(\Gamma(3)/(\Gamma(2)1!)=2\), so \(P(M=1)=2(5/9)^2(4/9)=200/729=0.2743\).
For two claims, the coefficient is \(\Gamma(4)/(\Gamma(2)2!)=3\), giving \(P(M=2)=3(5/9)^2(4/9)^2=1200/6561=0.1829\).
Add the exact fractions before rounding: \(25/81+200/729+1200/6561=5025/6561\), which is approximately \(0.7659\). Therefore, \(P(M\ge 3)=1-5025/6561=1536/6561\), approximately \(0.2341\). The predictive mean is \(2(4)/5=1.6\). The predictive variance is \(1.6+2(4^2/5^2)=1.6+1.28=2.88\).
Check. The three stated probabilities sum to approximately \(0.7659\), and their complement is approximately \(0.2341\). The mean and variance are positive, and the variance exceeds the mean as expected when the rate is uncertain.
Common mistake avoided: Adding individually rounded probabilities can change the rounded complement. Calculate the complement from exact fractions, then round the final result to four decimal places.
How Prior Strength Changes the Forecast
The Gamma prior’s mean is \(\alpha/\beta\). When two priors have the same mean but different shapes, the prior with larger shape and rate has less relative uncertainty and exerts more influence after a limited amount of experience. A zero observation lowers each posterior mean, but need not lower them by the same amount.
Worked Example: Comparing Weak and Strong Priors
Consider two hypothetical risks with zero claims over 2 exposure units and a future period of 3 exposure units. For Risk A, the prior is \(\operatorname{Gamma}(1,2)\); for Risk B, it is \(\operatorname{Gamma}(4,8)\). Compare posterior means, predictive means, probabilities of no future claims, and predictive variances.
Method. Update each prior with the same zero-claim exposure. Both prior means are \(0.5\), but Risk B’s prior is more concentrated: its shape and rate are each four times those for Risk A.
For Risk A, the posterior is \(\operatorname{Gamma}(1,2+2)=\operatorname{Gamma}(1,4)\). Its posterior mean rate is \(1/4=0.25\), its predictive mean is \(0.25(3)=0.75\), and its probability of no future claims is \((4/(4+3))^1=4/7=0.5714\). Its predictive variance is \(0.75+1(3^2/4^2)=0.75+0.5625=1.3125\).
For Risk B, the posterior is \(\operatorname{Gamma}(4,8+2)=\operatorname{Gamma}(4,10)\). Its posterior mean rate is \(4/10=0.4\), its predictive mean is \(0.4(3)=1.2\), and its probability of no future claims is \((10/13)^4=10000/28561=0.3501\). Its predictive variance is \(1.2+4(3^2/10^2)=1.2+0.36=1.56\).
The data reduce both prior mean rates from \(0.5\), but the more concentrated prior retains a larger posterior mean after the same zero experience. Hence it forecasts more claims and a lower chance of zero claims in the future. This comparison illustrates prior sensitivity; it does not establish which prior is appropriate without considering the risk context and supporting information.
Check. Both posterior rates are below their prior mean of \(0.5\). The forecasts differ because the priors assign different weight to the common prior mean, not because the observed data differ.
Common mistake avoided: Equal prior means do not imply equal posterior forecasts. The prior’s concentration, as well as its mean, affects how strongly zero experience changes the estimate.
A Bernoulli Alternative for Claim Incidence
For a fixed set of individual risks, a claim-incidence model can be more natural than a claim-count process. Suppose each exposure independently has either one claim or no claim, conditional on a common unknown probability \(p\). A Beta prior \(p\sim\operatorname{Beta}(a,b)\) updates to \(\operatorname{Beta}(a+x,b+n-x)\) after \(x\) claims among \(n\) exposures. Thus, zero claims among \(n\) risks give posterior \(\operatorname{Beta}(a,b+n)\). For \(m\) future risks, the predictive count has a Beta-Binomial distribution and mean \(m a/(a+b+n)\).
Worked Example: Future Incidence After Five Claim-Free Risks
A hypothetical portfolio uses a common claim probability \(p\) with prior \(p\sim\operatorname{Beta}(2,8)\). Five independent risks each have zero or one claim in the observation period, and none has a claim. Find the posterior predictive mean number of claims among four future risks, and the probabilities of zero claims, exactly one claim, and at least two claims.
Method. Update the Beta prior using five failures (risks with no claim), then integrate the future binomial distribution over the posterior.
The posterior is \(\operatorname{Beta}(2,8+5)=\operatorname{Beta}(2,13)\), with posterior mean probability \(2/(2+13)=2/15\). The predictive mean for four risks is \(4(2/15)=8/15=0.5333\) claims.
For zero future claims, multiply the successive conditional probabilities of no claim: \(P(K=0)=\frac{13}{15}\frac{14}{16}\frac{15}{17}\frac{16}{18} =\frac{182}{306}=\frac{91}{153}=0.5948\).
For exactly one claim, the Beta-Binomial probability is \(\binom{4}{1}\frac{(2)(13)(14)(15)}{(15)(16)(17)(18)} =\frac{91}{306}=0.2974\). Consequently, \(P(K\ge2)=1-91/153-91/306=33/306=11/102=0.1078\).
Check. The probabilities for zero, one, and at least two claims sum to \(182/306+91/306+33/306=306/306=1\). The positive predictive mean is consistent with a posterior mean claim probability of \(2/15\), despite observing no claims.
Common mistake avoided: The five observed zeroes update the prior parameter \(b\); they do not set the unknown probability \(p\) to zero. This model applies to fixed individual trials with at most one claim per trial, not unrestricted counts over continuous exposure.
Interpretation and Common Exam Traps
Zero observations carry information, but their effect depends on how much exposure produced them and on the model for the unknown risk. In the Poisson-Gamma model, additional zero-claim exposure increases the posterior rate parameter from \(\beta\) to \(\beta+e_0\), lowering the posterior mean while leaving it positive when \(\alpha>0\). For a fixed prior, more zero-claim exposure therefore lowers the expected future count.
A tempting but incorrect calculation is: “The observed count was zero, so the forecast is zero.” That replaces an uncertain frequency parameter with an unsupported estimate of exactly zero. Under the Gamma model, the correct posterior mean future count is \(\alpha e_1/(\beta+e_0)>0\) for positive \(\alpha\) and future exposure. The forecast can approach zero as observed claim-free exposure grows, but does not become zero merely because one observation period had no claims.
Check these points before calculating:
- Keep exposure units consistent. A rate per vehicle-year must be multiplied by vehicle-years, not by calendar years alone unless there is exactly one vehicle at risk.
- Use the stated Gamma parameterization. These formulas use a rate parameter \(\beta\), so the prior mean is \(\alpha/\beta\). A scale parameterization requires different substitutions.
- Separate posterior and predictive uncertainty. A Poisson forecast at the posterior mean gives the predictive mean correctly, but generally understates predictive variance and does not give the full posterior predictive probabilities.
- Respect the model’s limits. The stable-rate assumption may be unsuitable if exposure, coverage, claim reporting, or underlying conditions change materially between periods.
Check Your Understanding
Use the stated model and parameterization in each problem. Round probabilities to four decimal places.
- A hypothetical risk has \(\lambda\sim\operatorname{Gamma}(3,6)\) and zero claims over 3 exposure units. What is its predictive mean over the next 2 exposure units, and what is the probability of zero future claims? (Answer: mean \(=0.6667\); \(P(M=0)=729/1331=0.5477\).)
- A risk has prior \(\operatorname{Gamma}(2,4)\), zero claims over 1 exposure unit, and future exposure 3. Find the predictive mean and variance. (Answer: mean \(=1.2000\); variance \(=1.9200\).)
- For a Bernoulli incidence model, let \(p\sim\operatorname{Beta}(1,9)\). There are zero claims among 4 observed risks. Find the predictive mean number of claims among 5 future risks and the probability of zero future claims. (Answer: mean \(=5/14=0.3571\); \(P(K=0)=13/18=0.7222\).)
- Two prior claim rates have distributions \(\operatorname{Gamma}(2,4)\) and \(\operatorname{Gamma}(8,16)\). Each observes zero claims over 2 exposure units. Find each posterior predictive mean over 4 future exposure units. (Answer: \(4/3=1.3333\) and \(16/9=1.7778\), respectively.)
- A claim rate has prior \(\operatorname{Gamma}(1,2)\). After zero claims over 1 exposure unit, find the probabilities of zero claims, exactly one claim, and at least two claims over the next 3 exposure units. (Answer: \(0.5000,\ 0.2500,\ 0.2500\), respectively.)