Why Forecast Methods Can Disagree
A forecast may concern an average, such as the expected number of claims, or a nonlinear quantity, such as the probability of no claims or the probability that a loss exceeds a threshold. These are different forecasting targets. Averaging a parameter-dependent expectation over its posterior can give the same answer as evaluating it at the posterior mean when the expectation is linear in the parameter. For a nonlinear probability, the two calculations generally differ.
This distinction matters in ratemaking because claim probabilities often involve nonlinear functions of an unknown frequency or severity parameter. A point estimate can be a useful summary, but substituting it into a nonlinear formula does not generally account for the range of parameter values still supported by the data.
These definitions compare forecasts of the same target. For example, when the target is the probability of zero claims, \(q(\theta)\) is the conditional probability of zero claims. The posterior predictive answer averages that probability over plausible parameter values; the plug-in answer calculates it at one selected parameter value. Unless \(q\) is linear or another special equality applies, they need not agree.
Linearity and the Difference Between the Forecasts
For a future Poisson claim count \(N\) over one unit of exposure, conditional on rate \(\lambda\), \(E[N\mid\lambda]=\lambda\). Thus the posterior predictive mean count is \(E[\lambda\mid D]\), which equals the plug-in mean count when the plug-in value is the posterior mean. But the conditional probability of no claims is \(q(\lambda)=e^{-\lambda}\), a nonlinear function. Its posterior average is not generally \(e^{-E[\lambda\mid D]}\).
A useful justification follows directly from the definition of expectation. If \(q(\theta)=a+b\theta\), where \(a\) and \(b\) are constants, then \(E[q(\theta)\mid D]=a+bE[\theta\mid D]=q(E[\theta\mid D])\). For a nonlinear \(q\), that algebra does not hold. The direction of the difference depends on the shape of \(q\) over the posterior's range; one should not assume that a posterior predictive probability is always larger or always smaller.
The following examples use the posterior mean as the plug-in estimate so the comparison isolates the effect of averaging over parameter uncertainty. All scenarios are hypothetical. There is no interest discounting or mortality component in these examples; exposures and claim outcomes are treated as specified.
Poisson Frequency: Claim-Free Probability
Worked Example: Averaging a Poisson Zero-Claim Probability
A hypothetical portfolio has an unknown claim rate \(\lambda\) per unit of exposure. Before observing experience, \(\lambda\sim\operatorname{Gamma}(2,\text{ rate }1)\), using the shape-rate convention. One claim is observed over one unit of exposure. Forecast the probability of zero claims over one future unit in two ways: by integrating over the posterior, and by plugging in the posterior mean. Also compare the forecasts of the future mean count.
Method and posterior. Under a Poisson likelihood, one observed claim over one exposure unit contributes a factor proportional to \(\lambda e^{-\lambda}\). Multiplying this by the prior density, proportional to \(\lambda^{2-1}e^{-\lambda}\), gives a posterior density proportional to \(\lambda^2e^{-2\lambda}\). Therefore \(\lambda\mid D\sim\operatorname{Gamma}(3,\text{ rate }2)\). Its posterior mean is \(3/2=1.5\).
Posterior predictive zero-claim probability. Conditional on \(\lambda\), the future count is Poisson with mean \(\lambda\), so \(P(N=0\mid\lambda)=e^{-\lambda}\). For a Gamma variable with shape \(\alpha\) and rate \(\beta\), integrating its density gives \(E[e^{-\lambda}]=(\beta/(\beta+1))^\alpha\). Here this is
Plug-in zero-claim probability. Substitute the posterior mean rate, \(1.5\), into the conditional Poisson probability:
Future mean count. The posterior predictive mean is \(E[E(N\mid\lambda)\mid D]=E[\lambda\mid D]=1.5\). The plug-in mean is also \(E[N\mid\lambda=1.5]=1.5\). The mean counts agree because the conditional mean is linear in \(\lambda\), while the zero-claim probability is nonlinear.
Check and common mistake. The predictive zero probability, \(0.2963\), exceeds the plug-in value, \(0.2231\), in this case. A common wrong calculation is to report \(e^{-1.5}=0.2231\) as the posterior predictive probability. That is only the probability at the posterior mean rate; the correct posterior predictive probability averages \(e^{-\lambda}\) across the Gamma posterior. The computed probability is between zero and one, as required.
Binary Claim Incidence: Probability of a Threshold
Worked Example: At Least Two Claims in a Small Cohort
A hypothetical rating class has an unknown probability \(p\) of a claim for each exposure. The prior is \(p\sim\operatorname{Beta}(2,3)\). In three observed exposures, one claim occurs. For four future independent exposures from the same class, calculate the probability of at least two claims using the posterior predictive distribution and using the posterior mean as a plug-in estimate.
Method and posterior. The observed data contain one claim and two nonclaims. The Beta-Binomial update gives \(p\mid D\sim\operatorname{Beta}(2+1,3+2)=\operatorname{Beta}(3,5)\), with posterior mean \(3/(3+5)=3/8=0.375\). Let \(K\) be the future claim count among four exposures.
Posterior predictive calculation. For \(k\) claims among \(n\) future exposures, integrating a Binomial likelihood over a \(\operatorname{Beta}(a,b)\) posterior gives \(P(K=k\mid D)=\binom{n}{k}B(a+k,b+n-k)/B(a,b)\). Here it is simpler to calculate the complement \(P(K\ge2)=1-P(K=0)-P(K=1)\). For \(\operatorname{Beta}(3,5)\), the probabilities are
Thus \(P(K\ge2\mid D)=1-7/33-7/22=31/66=0.4697\).
Plug-in calculation. Use \(p=3/8\) in a \(\operatorname{Binomial}(4,p)\) model. Then \(P(K=0)=(5/8)^4=625/4096=0.1526\), and \(P(K=1)=4(3/8)(5/8)^3=1500/4096=0.3662\). Therefore,
Check and common mistake. The answers differ by about \(0.0115\). A common error is to calculate the threshold probability at \(p=0.375\) and call it posterior predictive; that leaves out uncertainty in \(p\). The predictive calculation instead averages the event probability over \(\operatorname{Beta}(3,5)\). The two methods agree on the expected future count, \(4(3/8)=1.5\), but not on this nonlinear threshold probability.
Severity: Probability a Loss Exceeds a Threshold
Worked Example: Predictive Exceedance Under an Uncertain Exponential Rate
For a hypothetical coverage, individual claim severity \(X\), conditional on rate \(\lambda\), has an exponential distribution with density \(f_X(x\mid\lambda)=\lambda e^{-\lambda x}\) for \(x\ge0\). The prior for \(\lambda\) is \(\operatorname{Gamma}(1,\text{ rate }1)\). One observed claim has severity \(1\) dollar in the chosen measurement units. Calculate the predictive probability that a future claim exceeds \(1\) dollar, and compare it with the plug-in probability. Severity observations are conditionally independent given \(\lambda\).
Posterior. The one-observation likelihood is proportional to \(\lambda e^{-\lambda}\). Multiplying by the prior density, proportional to \(e^{-\lambda}\), yields a posterior density proportional to \(\lambda e^{-2\lambda}\). Hence \(\lambda\mid D\sim\operatorname{Gamma}(2,\text{ rate }2)\), and \(E[\lambda\mid D]=2/2=1\).
Posterior predictive exceedance probability. Conditional on \(\lambda\), the exponential survival probability at threshold \(c\) is \(P(X>c\mid\lambda)=e^{-c\lambda}\). Integrating over a Gamma posterior gives \(E[e^{-c\lambda}\mid D]=(\beta/(\beta+c))^\alpha\). With shape \(2\), rate \(2\), and \(c=1\),
Plug-in exceedance probability. At the posterior mean rate \(\lambda=1\), the conditional survival probability is
Check and common mistake. The posterior predictive exceedance probability is higher here because averaging \(e^{-\lambda}\) over the posterior gives \(4/9\), not \(e^{-E[\lambda\mid D]}\). Substituting \(\lambda=1\) directly into the survival function is a valid plug-in calculation, but it is not the posterior predictive calculation. Both results are valid probabilities, and neither calculation changes the severity units or the stated threshold.
Interpreting the Comparison
The examples show that the choice between a posterior predictive forecast and a plug-in forecast depends on the target, not merely on the posterior distribution. A linear target, such as the conditional expected Poisson count, can be forecast by the posterior mean parameter without changing the result. For a nonlinear event probability, integration and substitution are different operations.
The discrepancy can matter for decisions involving claim-free discounts, threshold-based rating rules, attachment points, or other decisions tied to probabilities rather than means. It does not establish that one method is universally preferable for every purpose: the posterior predictive forecast explicitly averages parameter uncertainty, while the plug-in forecast answers a point-estimate-based question. State which forecast is being requested, and do not label one as the other.
Write the conditional mean or probability as a function \(q(\theta)\), including the exposure, threshold, or event of interest.
Use the observed data and the stated likelihood and prior to obtain \(\pi(\theta\mid D)\), keeping parameter conventions explicit.
Integrate \(q(\theta)\) over the posterior for the posterior predictive result; separately calculate \(q(E[\theta\mid D])\) for the posterior-mean plug-in result.
Confirm that each result is a probability or mean for the same future exposure, and explain whether equality follows from linearity or is merely numerical.
A frequent exam trap is to compute the posterior mean parameter correctly and then assume that every forecast can be obtained by substituting that mean. That shortcut is justified for a linear conditional mean, but not for a nonlinear claim probability without a specific mathematical reason for equality. This distinction is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.
Check Your Understanding
For each problem, compare the posterior predictive forecast with the forecast obtained by plugging in the posterior mean parameter.
- A claim rate has posterior \(\lambda\mid D\sim\operatorname{Gamma}(3,\text{ rate }2)\). For one future exposure unit with a conditional Poisson model, find the posterior predictive probability of zero claims and the plug-in probability. (Answer: posterior predictive \(8/27=0.2963\); plug-in \(e^{-1.5}=0.2231\).)
- A claim probability has posterior \(p\mid D\sim\operatorname{Beta}(2,4)\). For two future independent exposures, find the probability of at least one claim by both methods. (Answer: posterior predictive \(11/21=0.5238\); plug-in \(5/9=0.5556\).)
- A future severity is exponential conditional on an uncertain rate, with posterior \(\lambda\mid D\sim\operatorname{Gamma}(2,\text{ rate }3)\). Find the probability that severity exceeds \(2\) measurement units by both methods. (Answer: posterior predictive \(9/25=0.3600\); plug-in \(e^{-4/3}=0.2636\).)
- A Poisson claim rate has posterior \(\lambda\mid D\sim\operatorname{Gamma}(4,\text{ rate }2)\). Over two future exposure units, calculate the probability of zero claims using posterior integration and the posterior-mean plug-in rate. (Answer: posterior predictive \(1/16=0.0625\); plug-in \(e^{-4}=0.0183\).)