Building Predictive Reserve Distributions From Credible Estimates

Why a Predictive Reserve Is a Distribution

A credible estimate of ultimate losses is not necessarily a complete description of future reserve outcomes. Even if an estimate is well calibrated, actual future payments vary because claims emerge randomly. In addition, the ultimate loss level itself may be uncertain: another credible estimate, posterior draw, or bootstrap replicate could produce a different ultimate. A predictive reserve distribution brings these two sources of variation together.

The distinction matters whenever the question concerns a probability, such as the chance that future payments exceed a threshold, or a measure of variability. Substituting a single credible ultimate estimate into a process model can produce a useful conditional distribution, but it does not include uncertainty in that estimate. A predictive distribution averages over both.

Definition: Let $D$ denote the observed data and $\Theta$ an uncertain ultimate-loss parameter or ultimate estimate. A predictive reserve $R$ has conditional distribution $P(R\le r\mid D)=\int P(R\le r\mid \theta,D)\,\pi(d\theta\mid D)$, where $\pi(d\theta\mid D)$ describes uncertainty in $\Theta$ after using the data. The conditional distribution describes future process variation at a given $\theta$; the mixing distribution describes uncertainty about $\theta$.

In practice, the distribution for $\Theta$ might be obtained from a posterior model, a credibility model with an explicit uncertainty distribution, or a resampling procedure. A credibility point estimate alone does not specify this distribution. If only the point estimate is available, parameter uncertainty cannot be included without additional assumptions or information.

Unless stated otherwise, the examples below use hypothetical insurance portfolios, undiscounted payments in constant dollars, and an effective annual interest rate of $i=0\%$. Mortality is not part of these claim-emergence models. No mortality assumption is needed.

Mixing Process and Parameter Uncertainty

The law of total expectation and the law of total variance provide a direct way to calculate predictive summaries. Conditional on $\Theta$, first calculate the mean and variance of the future reserve. Then average over the uncertainty in $\Theta$.

E[R\mid D]=E\!\left[E[R\mid\Theta,D]\mid D\right]

\operatorname{Var}(R\mid D) = E\!\left[\operatorname{Var}(R\mid\Theta,D)\mid D\right] + \operatorname{Var}\!\left(E[R\mid\Theta,D]\mid D\right)

The first variance term is the average process variance: outcomes vary even if the ultimate parameter were known. The second is the variance of the conditional means: the expected future reserve changes with the uncertain ultimate parameter. This separation prevents a common error—reporting only process variance and thereby omitting parameter uncertainty.

When this result applies: The conditional distributions and the distribution of $\Theta$ must describe the same model and data basis. Any dependence between the estimate and future process must be represented in $P(R\mid\Theta,D)$. Do not add a parameter-variance term a second time if it is already included in a simulated predictive distribution.

Worked Example: Uncertain Ultimate Count and Binomial Emergence

A hypothetical portfolio has one claim already reported. A credibility analysis represents the ultimate claim count $N$ as 3 with probability 0.40 and 5 with probability 0.60. Each ultimate claim has a fixed payment of $1,000. Each claim not yet reported emerges during the next year independently with probability $p=0.50$. Find the predictive distribution of next-year payments, its mean and variance, and the probability of at least $3,000 of payments.

Method and assumptions. Condition on $N$. With one claim already reported, the number of potential future claims is $N-1$, and the number emerging is conditionally binomial with parameters $N-1$ and 0.50. Each payment is $1,000. The distribution for $N$ is the stated two-point credibility uncertainty distribution; no additional estimate uncertainty is assumed. The common trap is to treat the expected ultimate count as known and omit variation in $N$.

If $N=3$, two claims remain, so the future claim count has probabilities $0.25, 0.50, 0.25$ for 0, 1, and 2 claims. If $N=5$, four claims remain, so its probabilities for 0 through 4 claims are $0.0625, 0.25, 0.375, 0.25, 0.0625$. Weight each conditional probability by the corresponding probability for $N$:

Future claims $k$	$P(K=k\mid N=3)$	$P(K=k\mid N=5)$	$P(K=k)$
0	0.25	0.0625	$0.40(0.25)+0.60(0.0625)=0.1375$
1	0.50	0.25	$0.40(0.50)+0.60(0.25)=0.3500$
2	0.25	0.375	$0.40(0.25)+0.60(0.375)=0.3250$
3	0	0.25	$0.1500$
4	0	0.0625	$0.0375$

The probabilities sum to $0.1375+0.3500+0.3250+0.1500+0.0375=1.0000$, as required. The reserve is $R=1{,}000K$, so the probability of at least $3,000 is $P(K\ge 3)=0.1500+0.0375=0.1875$.

For the mean, the conditional expected future claim counts are $2(0.50)=1$ when $N=3$ and $4(0.50)=2$ when $N=5$. Thus $E[K]=0.40(1)+0.60(2)=1.60$, giving $E[R]=\$1{,}600$.

For the variance, the conditional variances are $2(0.50)(0.50)=0.50$ and $4(0.50)(0.50)=1.00$. The mean conditional variance is $0.40(0.50)+0.60(1.00)=0.80$. The conditional means, 1 and 2, have variance $0.40(1-1.60)^2+0.60(2-1.60)^2=0.24$. Therefore $\operatorname{Var}(K)=0.80+0.24=1.04$, and $\operatorname{Var}(R)=1{,}000^2(1.04)=\$1{,}040{,}000$. The variance is in dollars squared.

Check. The probability table sums to one; the tabulated mean is $0.3500+2(0.3250)+3(0.1500)+4(0.0375)=1.60$. The variance calculation includes both process and ultimate-count uncertainty.

When Claim Severities Also Vary

A predictive reserve can include random payment sizes as well as uncertain ultimate counts and emergence. For a simple independent-claim model, let $N$ be the uncertain number of outstanding claims. For each claim, let $I$ equal 1 if it emerges during the horizon and 0 otherwise, and let $X$ be its payment if it emerges. The payment contribution is $Z=IX$. If emergence is independent of severity, then $E[Z]=pE[X]$ and $E[Z^2]=pE[X^2]$. Consequently,

\operatorname{Var}(Z)=pE[X^2]-p^2(E[X])^2

For conditionally independent claims, $R=\sum_{j=1}^{N} Z_j$ has conditional mean $N E[Z]$ and conditional variance $N\operatorname{Var}(Z)$. If $N$ is uncertain, the law of total variance then accounts for variation in $N$ as well. The independence assumptions are important: claim severities, emergence, or the ultimate count may be dependent in a more detailed model.

Worked Example: Count, Severity, and Emergence Uncertainty

A hypothetical book has $N=1$ outstanding claim with probability 0.50 or $N=3$ outstanding claims with probability 0.50. Each claim emerges during the next year independently with probability $p=0.50$. Conditional on emergence, its payment $X$ is $1,000 or $3,000 with equal probability, independently across claims. Find the mean and variance of next-year payments and the probability that payments are at least $3,000.

Method and assumptions. Work in thousands of dollars. For one outstanding claim, $Z=IX$ is 0 with probability 0.50, 1 with probability 0.25, and 3 with probability 0.25. Thus $E[Z]=0.25(1)+0.25(3)=1$ and $E[Z^2]=0.25(1)+0.25(9)=2.5$. Hence $\operatorname{Var}(Z)=2.5-1^2=1.5$. The common trap is to calculate severity variance alone: the zero payment when a claim does not emerge must also be included.

The count has $E[N]=0.50(1)+0.50(3)=2$ and $\operatorname{Var}(N)=0.50(1-2)^2+0.50(3-2)^2=1$. Conditional on $N$, $E[R\mid N]=N$ and $\operatorname{Var}(R\mid N)=1.5N$. Therefore,

$E[R]=E[N]=2$ thousand dollars, or $2,000. Also, $\operatorname{Var}(R)=E[1.5N]+\operatorname{Var}(N)=1.5(2)+1=4$ thousand-dollars squared, or $4,000,000.

If $N=1$, the largest possible payment is $3,000, and the probability of reaching that amount is 0.25. If $N=3$, the reserve is the sum of three independent $Z$ values. Its probability of being less than 3 (thousand dollars) comes from: all three zero, $0.50^3=0.125$; exactly one payment of 1 and two zeros, $3(0.25)(0.50^2)=0.1875$; or exactly two payments of 1 and one zero, $\binom{3}{2}(0.25^2)(0.50)=0.09375$. These are the only totals below 3. Thus $P(R\ge3\mid N=3)=1-(0.125+0.1875+0.09375)=0.59375$. Mixing over $N$, $P(R\ge\$3{,}000)=0.50(0.25)+0.50(0.59375)=0.421875$, or 0.4219 to four decimals.

Check. The mean and variance use the same per-claim distribution that produced the threshold calculation. In particular, the count uncertainty adds $1$ thousand-dollars squared to the expected conditional process variance of $3$ thousand-dollars squared.

A Continuous Parameter Example: Gamma-Poisson Mixing

A useful model for uncertain claim counts has a Poisson future count conditional on an uncertain intensity. Suppose $\Theta$ is an ultimate claim intensity with a Gamma distribution using shape $\alpha$ and rate $\beta$, so its density is proportional to $\theta^{\alpha-1}e^{-\beta\theta}$ for $\theta>0$. Conditional on $\Theta=\theta$, suppose future emergence over the specified horizon is Poisson with mean $q\theta$, where $q$ is the fraction of ultimate intensity expected to emerge during that horizon. This model assumes conditional Poisson emergence, a common $q$, and a Gamma uncertainty distribution calibrated to the available data.

K\mid\Theta=\theta\sim\operatorname{Poisson}(q\theta),\qquad E[K]=\frac{q\alpha}{\beta},\qquad \operatorname{Var}(K)=\frac{q\alpha}{\beta}+\frac{q^2\alpha}{\beta^2}

The first variance term is conditional Poisson process variance. The second is the contribution from uncertainty in $\Theta$. A fixed-intensity Poisson model with $\theta=E[\Theta]$ has the same mean but lacks that second variance term.

Justification: Conditional on $\Theta$, $E[K\mid\Theta]=q\Theta$ and $\operatorname{Var}(K\mid\Theta)=q\Theta$. For a Gamma distribution with shape $\alpha$ and rate $\beta$, $E[\Theta]=\alpha/\beta$ and $\operatorname{Var}(\Theta)=\alpha/\beta^2$. Applying the laws of total expectation and variance gives $E[K]=q\alpha/\beta$ and $\operatorname{Var}(K)=E[q\Theta]+\operatorname{Var}(q\Theta)=q\alpha/\beta+q^2\alpha/\beta^2$.

Worked Example: Predictive Count Probabilities Under Gamma Uncertainty

A hypothetical portfolio has posterior ultimate claim intensity $\Theta\sim\operatorname{Gamma}(\alpha=4,\beta=2)$, using the rate parameterization. The next-year emergence fraction is $q=0.50$. Conditional on $\Theta$, the number of emerging claims $K$ is Poisson with mean $0.50\Theta$. Each claim pays exactly $2,000. Find the predictive mean and variance of payments, the probability of no payments, and the probability of payments of at least $4,000.

Method and assumptions. Mix the conditional Poisson model over the stated Gamma distribution. The claims emerge conditionally according to a Poisson model, and each has fixed severity $2,000. The common trap is to substitute only the posterior mean intensity and thereby discard the additional variance and change the predictive probabilities.

The intensity has $E[\Theta]=4/2=2$ and $\operatorname{Var}(\Theta)=4/2^2=1$. Therefore $E[K]=qE[\Theta]=0.50(2)=1$, while $\operatorname{Var}(K)=qE[\Theta]+q^2\operatorname{Var}(\Theta)=0.50(2)+0.50^2(1)=1.25$. Payments are $R=2{,}000K$, so $E[R]=\$2{,}000$ and $\operatorname{Var}(R)=2{,}000^2(1.25)=\$5{,}000{,}000$.

For the probability calculation, the predictive probability of $K=0$ is the Gamma average of the conditional Poisson probability: $P(K=0)=E[e^{-q\Theta}]=(\beta/(\beta+q))^\alpha=(2/2.5)^4=0.4096$. The predictive probability of one claim is $P(K=1)=0.4096(4)(0.50/2.5)=0.32768$. At least $4,000 means at least two claims, so $P(R\ge\$4{,}000)=P(K\ge2)=1-0.4096-0.32768=0.26272$, or 0.2627 to four decimals.

Check. The mean count is one. The predictive probabilities for zero and one claim sum to 0.73728, leaving 0.26272 for two or more. A Poisson model fixed at $\Theta=2$ would instead have variance 1 and probability of no claims $e^{-1}$, illustrating the effect of parameter uncertainty.

Practical Checks Before Reporting a Predictive Reserve

A predictive distribution is only as meaningful as its conditioning and units. Before reporting results, check the following:

Define the horizon. State whether the reserve is payments in a particular future period or all future unpaid losses. An emergence probability for one year is not automatically the probability for ultimate development.
Use the correct outstanding base. If the ultimate count includes already reported claims, subtract those claims before modeling future emergence, unless the model explicitly includes them.
Keep uncertainty sources distinct. State which distribution represents ultimate-estimate uncertainty and which represents future process variation. Do not count either source twice.
Check conditional assumptions. Conditional independence, constant emergence probabilities, and a particular severity distribution are modeling assumptions, not consequences of credibility weighting.
Interpret the variance units correctly. If payments are measured in thousands of dollars, a variance calculated in those units is in thousands of dollars squared; multiply by $1{,}000^2$ to express it in dollars squared.
Do not infer a full distribution from a point estimate. A credibility mean, without a model for its uncertainty, does not determine predictive probabilities or reserve quantiles.

Key takeaway: A predictive reserve distribution averages the future process distribution over uncertainty in the ultimate estimate. The law of total variance makes the two sources explicit: expected process variance plus variance in the conditional reserve mean. This distinction is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Check Your Understanding

For each problem, treat payments as undiscounted and use the stated assumptions. Check both the conditioning and the units.

A hypothetical book has $N=2$ outstanding claims with probability 0.25 or $N=4$ with probability 0.75. Each claim emerges during the next year independently with probability 0.25 and pays $1,000 if it emerges. Find the predictive mean and variance of next-year payments. (Answer: mean $875; variance $703,125 dollars squared.)
Suppose $\Theta\sim\operatorname{Gamma}(\alpha=3,\beta=1.5)$ with rate parameterization, and $K\mid\Theta\sim\operatorname{Poisson}(0.50\Theta)$. Each claim pays $1,000. Find the predictive mean and variance of payments. (Answer: mean $1,000; variance $1,333,333.33 dollars squared.)
A hypothetical book has $N=2$ or $N=4$ outstanding claims, each with probability 0.50. Each independently emerges with probability 0.50 and pays $1,000 or $3,000 with equal conditional probabilities. Find the predictive mean and variance. (Answer: mean $3,000; variance $5,500,000 dollars squared.)
Let $\Theta\sim\operatorname{Gamma}(\alpha=2,\beta=1)$ with rate parameterization and $K\mid\Theta\sim\operatorname{Poisson}(0.50\Theta)$. Find the probability of at least two future claims. (Answer: $7/27$, or 0.2593 to four decimals.)
A credibility analysis supplies only a point estimate of 12 outstanding ultimate claims. Explain what additional assumption is needed to calculate a predictive distribution for next-year payments, and name the two sources of variation that distribution should represent. (Answer: Specify a future emergence and payment model and, to include estimate uncertainty, an uncertainty distribution for the ultimate count or estimate; represent process variation and ultimate-estimate uncertainty.)

Future claims \(k\)	\(P(K=k\mid N=3)\)	\(P(K=k\mid N=5)\)	\(P(K=k)\)
0	0.25	0.0625	\(0.40(0.25)+0.60(0.0625)=0.1375\)
1	0.50	0.25	\(0.40(0.50)+0.60(0.25)=0.3500\)
2	0.25	0.375	\(0.40(0.25)+0.60(0.375)=0.3250\)
3	0	0.25	\(0.1500\)
4	0	0.0625	\(0.0375\)

What You'll Learn

Why a Predictive Reserve Is a Distribution

Mixing Process and Parameter Uncertainty

Worked Example: Uncertain Ultimate Count and Binomial Emergence

Worked Example: Uncertain Ultimate Count and Binomial Emergence

When Claim Severities Also Vary

Worked Example: Count, Severity, and Emergence Uncertainty

A Continuous Parameter Example: Gamma-Poisson Mixing

Worked Example: Predictive Count Probabilities Under Gamma Uncertainty

Practical Checks Before Reporting a Predictive Reserve

Check Your Understanding

Previous

Next