Simulating Reserve Outcomes After Credibility Estimation

From a Credible Estimate to Simulated Reserve Outcomes

A credibility-weighted ultimate estimate gives a central estimate of total losses, but it does not describe the range of future payments that could produce that ultimate. To study that range, first reconcile the credible ultimate with losses already paid, then simulate the remaining development under an explicitly stated model. Each simulation produces one possible reserve outcome.

The central estimate and the simulated outcomes have different roles. The credible ultimate anchors the expected total loss; the development model describes variation around that estimate. A prediction interval summarizes the simulated reserve outcomes, not the uncertainty in the credibility weight or in the model parameters.

Definition: Let $\widehat U_{\mathrm{cred}}$ be a credibility-weighted estimate of ultimate losses, and let $P$ be cumulative paid losses at the valuation date. The indicated reserve is $\widehat R=\widehat U_{\mathrm{cred}}-P$. A simulated reserve outcome $R^{(b)}$ is the total future payment generated in simulation $b$. If the future payments are the only amounts remaining to be paid, $R^{(b)}$ is their sum.

For a credibility blend of a chain-ladder ultimate estimate $\widehat U_{\mathrm{CL}}$ and an a priori ultimate estimate $U_0$, the selected ultimate is $\widehat U_{\mathrm{cred}}=Z\widehat U_{\mathrm{CL}}+(1-Z)U_0$, where $0\leq Z\leq1$. The corresponding indicated reserve subtracts paid-to-date losses once. In a simulation, the future increments should have expected total equal to that reserve if the purpose is to represent development around the selected estimate.

\widehat R = Z\widehat U_{\mathrm{CL}}+(1-Z)U_0-P

The simulated reserve is not automatically the credible ultimate itself. The ultimate includes both paid and unpaid loss; the reserve includes only what remains unpaid. When the model generates future increments $Y_1,\ldots,Y_m$, the simulated outcome is $R=\sum_{j=1}^{m}Y_j$. Its mean should be reconciled to $\widehat R$, unless the modeling objective explicitly uses a different mean.

Why the Simulated Mean Must Reconcile

The reconciliation follows from the definition of ultimate loss. If the future increments complete the loss total, then ultimate loss equals paid-to-date loss plus future payments. Taking expectations, and treating the current paid amount as fixed at the valuation date, gives the expected future total required by the selected ultimate.

\widehat U_{\mathrm{cred}} = P+\sum_{j=1}^{m}E[Y_j] \quad\Longrightarrow\quad \sum_{j=1}^{m}E[Y_j] = \widehat U_{\mathrm{cred}}-P = \widehat R

This is a reconciliation condition, not a claim that every simulated outcome equals the indicated reserve. Individual outcomes vary; their model-based mean should agree with it. If future increments are independent, their variances add. If they are dependent, covariances must also be included:

E[R]=\sum_{j=1}^{m}E[Y_j], \qquad \operatorname{Var}(R) = \sum_{j=1}^{m}\operatorname{Var}(Y_j) + 2\sum_{j<k}\operatorname{Cov}(Y_j,Y_k)

A simulation implements this model repeatedly. In each repetition, draw the specified future increments, add them to obtain one reserve outcome, and retain that outcome. The mean of the simulated outcomes estimates the model-based reserve mean. Their empirical quantiles estimate prediction limits. With a finite number of simulations, those summaries may differ slightly from their theoretical values.

Set the selected ultimate and paid amount.
Calculate the credible ultimate using the stated blend, and subtract cumulative paid losses at the valuation date to obtain the indicated reserve.

Specify future development.
State the distribution of each future increment, its parameters, and any dependence between increments. Check that their expected total reconciles to the indicated reserve.

Generate reserve outcomes.
For each simulation, draw the future increments under the specified model and sum them. Do not add amounts already paid.

Summarize the outcomes.
Calculate or estimate the reserve mean and the requested prediction limits. State whether an interval is equal-tail and how discrete quantiles are defined.

Worked Simulation Examples

Worked Example: Normal Future Increments

A hypothetical book has a chain-ladder ultimate estimate of $\$1{,}200$ thousand and an a priori ultimate estimate of $\$1{,}000$ thousand. The credibility weight on chain ladder is $Z=0.40$. Paid losses are $\$700$ thousand. Model the next two future increments as independent normal random variables: $Y_1\sim N(250,40^2)$ and $Y_2\sim N(130,30^2)$, with amounts in thousands of dollars. Find the indicated reserve, the simulated reserve distribution, and its central 90% prediction interval. Use the standard normal 95th percentile $1.6449$.

Method. First form the credibility-weighted ultimate, then subtract paid losses. Simulate the two independent increments and sum them. The common mistake to avoid is treating the ultimate estimate as the reserve or adding the paid amount again.

Step 1: credible ultimate. The weight on the a priori estimate is $1-Z=0.60$, so $\widehat U_{\mathrm{cred}}=0.40(1{,}200)+0.60(1{,}000)=480+600=1{,}080$ thousand dollars.

Step 2: indicated reserve. Subtract the $\$700$ thousand already paid: $\widehat R=1{,}080-700=380$ thousand dollars.

Step 3: reserve distribution. The future total is $R=Y_1+Y_2$. Independent normal variables sum to a normal variable, with means and variances adding. Thus $E[R]=250+130=380$, and $\operatorname{Var}(R)=40^2+30^2=1{,}600+900=2{,}500$. The standard deviation is $\sqrt{2{,}500}=50$ thousand dollars. Therefore $R\sim N(380,2{,}500)$.

Step 4: central 90% prediction interval. A central 90% normal interval uses the 5th and 95th percentiles: $380\pm1.6449(50)=380\pm82.245$. The limits are $297.755$ and $462.245$ thousand dollars. Rounded to one decimal place, the interval is $\$297.8$ thousand to $\$462.2$ thousand.

Check. The modeled expected future total, $380$ thousand dollars, equals the indicated reserve. The interval is centered on that mean, as expected for this normal model.

Worked Example: Independent Discrete Development Shocks

For a hypothetical portfolio, the chain-ladder ultimate is $\$1{,}080$ thousand, the a priori ultimate is $\$980$ thousand, and $Z=0.50$. Paid losses are $\$725$ thousand. Future development consists of a first-period payment of $240+100B_1$ and a second-period payment of $200B_2$, in thousands of dollars. Here $B_1$ and $B_2$ are independent Bernoulli variables, with $P(B_1=1)=0.25$ and $P(B_2=1)=0.20$. Find the reserve mean and central 80% equal-tail prediction interval.

Method. Calculate the credible reserve, enumerate the possible future totals and their probabilities, then apply the discrete quantile definition $Q_p=\inf\{r:F_R(r)\geq p\}$. The trap is to treat the two shocks as mutually exclusive, even though they are specified as independent.

Step 1: indicated reserve. The credible ultimate is $0.50(1{,}080)+0.50(980)=540+490=1{,}030$ thousand dollars. Subtracting paid losses gives $\widehat R=1{,}030-725=305$ thousand dollars.

Step 2: check the expected future total. The first-period expected payment is $240+100(0.25)=265$. The second-period expected payment is $200(0.20)=40$. Their total expectation is $265+40=305$ thousand dollars, matching the indicated reserve.

Step 3: enumerate the outcomes. If neither shock occurs, the reserve is $240$, with probability $(0.75)(0.80)=0.60$. If only $B_1=1$, it is $340$, with probability $(0.25)(0.80)=0.20$. If only $B_2=1$, it is $440$, with probability $(0.75)(0.20)=0.15$. If both occur, it is $540$, with probability $(0.25)(0.20)=0.05$. Amounts are in thousands of dollars, and the probabilities sum to $1.00$.

Step 4: mean and interval. The mean is $240(0.60)+340(0.20)+440(0.15)+540(0.05)=144+68+66+27=305$ thousand dollars. For an 80% equal-tail interval, use $Q_{0.10}$ and $Q_{0.90}$. The cumulative probabilities at $240,340,440,540$ are $0.60,0.80,0.95,1.00$. Hence $Q_{0.10}=240$, and $Q_{0.90}=440$. The interval is $\$240$ thousand to $\$440$ thousand.

Check. The discrete distribution's mean is the indicated reserve. Its interval is not symmetric around the mean; equal-tail limits need not be symmetric for a discrete distribution.

Worked Example: A Shared Shock Across Development Periods

A hypothetical portfolio has a chain-ladder ultimate estimate of $\$1{,}500$ thousand, an a priori ultimate of $\$1{,}100$ thousand, and $Z=0.50$. Paid losses are $\$920$ thousand. Future period 1 payment is $100+100B$, and future period 2 payment is $200+300B$, in thousands of dollars, where the same Bernoulli variable $B$ is used in both periods and $P(B=1)=0.20$. Find the expected reserve, its variance, and its central 90% equal-tail prediction interval.

Method. Preserve the shared shock when forming each simulated outcome. The common mistake is to replace it with two independent shocks, which changes the distribution and understates the variance.

Step 1: credible reserve. The credible ultimate is $0.50(1{,}500)+0.50(1{,}100)=750+550=1{,}300$ thousand dollars. After subtracting the $\$920$ thousand paid, the indicated reserve is $1{,}300-920=380$ thousand dollars.

Step 2: combine the future increments. Since both periods use the same $B$, their total is $R=(100+100B)+(200+300B)=300+400B$. With probability $0.80$, $B=0$ and $R=300$; with probability $0.20$, $B=1$ and $R=700$, in thousands of dollars.

Step 3: mean and variance. The mean is $300(0.80)+700(0.20)=240+140=380$ thousand dollars. Equivalently, $E[R]=300+400(0.20)=380$. Since $\operatorname{Var}(B)=0.20(0.80)=0.16$, the reserve variance is $400^2(0.16)=25{,}600$ in squared thousands of dollars; the standard deviation is $160$ thousand dollars.

Step 4: equal-tail interval. The cumulative probability is $0.80$ at $300$ and $1.00$ at $700$. Therefore $Q_{0.05}=300$ and $Q_{0.95}=700$, giving a central 90% prediction interval of $\$300$ thousand to $\$700$ thousand.

Check and dependence warning. The mean agrees with the indicated reserve. If the two period-specific shock terms were incorrectly treated as independent Bernoulli shocks, the variance of the shock portion would be $100^2(0.16)+300^2(0.16)=16{,}000$, rather than the correct $400^2(0.16)=25{,}600$. The stipulated shared shock creates positive dependence, so the independent model would not reproduce the stated outcomes.

Interpreting the Results and Avoiding Common Traps

A reserve mean answers a different question from a prediction interval. The mean is the probability-weighted average of future payments under the model. A prediction interval describes a range of outcomes with a stated coverage level under that same model. It is not a guarantee that a future reserve will fall within the limits, nor is it a confidence interval for the credibility estimate.

For a simulated sample of $B$ outcomes $R^{(1)},\ldots,R^{(B)}$, the empirical mean is $\overline R=B^{-1}\sum_{b=1}^{B}R^{(b)}$. To estimate equal-tail limits with coverage $1-\alpha$, use the empirical quantiles at probabilities $\alpha/2$ and $1-\alpha/2$. For a discrete model, define the quantile consistently, for example as the smallest outcome whose cumulative probability reaches or exceeds the target probability.

Key takeaway: A sound reserve simulation has three aligned parts: the credible ultimate, paid losses at the valuation date, and a future-development model whose expected total matches the resulting reserve. The mean checks the reconciliation; the prediction interval reflects the modeled spread and shape of future outcomes.

Before reporting results, check the valuation basis and units, confirm that paid losses have not been included twice, and preserve any dependence specified between future increments. A model with the right mean but the wrong dependence can still give materially different reserve variability and prediction limits. The conclusions are conditional on the selected credibility estimate and development assumptions; this calculation does not add uncertainty in those selections unless that uncertainty is explicitly modeled.

Check Your Understanding

For each problem, use the stated model and report monetary amounts in thousands of dollars.

A hypothetical portfolio has chain-ladder ultimate $760$, a priori ultimate $700$, credibility weight $Z=0.40$, and paid losses $500$. Find the indicated reserve. (Answer: $224$)
Two independent future increments are normal with means $80$ and $120$, and variances $25$ and $144$. Find the reserve mean and central 95% prediction interval, using $1.96$ as the standard normal 97.5th percentile. (Answer: Mean $200$; interval $174.52$ to $225.48$)
A reserve outcome is $100+50B+150C$, where $B$ and $C$ are independent Bernoulli variables with probabilities $0.40$ and $0.20$. Find its mean and central 80% equal-tail prediction interval. (Answer: Mean $150$; interval $100$ to $250$)
A credible ultimate is $900$, paid losses are $650$, and future payments total $150+200B+100C$, where $B$ and $C$ are independent Bernoulli variables with probability $0.50$ each. Find the reserve mean and central 90% equal-tail interval. (Answer: Mean $300$; interval $150$ to $450$)
A chain-ladder ultimate of $1{,}400$ and a priori ultimate of $1{,}200$ are blended with $Z=0.25$; paid losses are $900$. Future payments total $250+200B$, where $P(B=1)=0.50$. Find the indicated reserve, the central 80% equal-tail interval, and the probability that the outcome exceeds $400$. (Answer: Indicated reserve $350$; interval $250$ to $450$; probability $0.5000$)

What You'll Learn