Exam Workshop: Credibility, Simulation, And Reserve Selection

From Reserve Indications to a Defensible Selection

A reserve problem often supplies more than one reasonable indication. Chain ladder extrapolates observed claims development; Bornhuetter-Ferguson (BF) combines an expected ultimate with the portion of that ultimate not yet observed; a credibility-weighted estimate can balance these sources of information. The arithmetic matters, but so does the basis of comparison: a method that estimates an ultimate loss must be converted consistently before it is compared with an unpaid reserve.

This tutorial follows a hypothetical two-line portfolio through the calculations and then examines simulation uncertainty. The purpose is not to declare one method universally best. A defensible selection reflects the maturity and reliability of the data, the credibility assigned to each source, and the uncertainty relevant to the decision. The methods and calculations are aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Definition: Let $C$ denote cumulative paid losses at the current valuation age, and let $U$ denote ultimate losses. The unpaid reserve is $R=U-C$. A cumulative loss development factor (LDF) $L$ projects the current cumulative amount to ultimate, so the chain ladder reserve is $C(L-1)$. If the expected ultimate is $E$, and the same selected development pattern implies that the proportion paid by the current age is $1/L$, the BF reserve is $E(1-1/L)$.

The development factors used below are selected assumptions, not universal constants. Their use requires that the observed development pattern be relevant to the claims being projected. The BF formula also assumes that the expected ultimate is on the same loss, exposure, and valuation basis as the claims being reserved. Neither formula by itself allows for a change in the underlying claim process or a mismatch in the data basis.

Why the Bornhuetter-Ferguson Unpaid Proportion Has This Form

Suppose the selected LDF from the current age to ultimate is $L$. Under that development pattern, current cumulative paid losses represent $1/L$ of ultimate: multiplying current paid losses by $L$ projects ultimate. Therefore, the fraction of ultimate not yet paid is $1-1/L$. Applying that fraction to an expected ultimate $E$ gives the BF estimate of unpaid losses.

\text{Expected paid proportion}=\frac{1}{L}, \qquad \text{Expected unpaid proportion}=1-\frac{1}{L}, \qquad R_{\mathrm{BF}}=E\left(1-\frac{1}{L}\right).

The derivation is a proportion argument: if $L=1.8$, the current age represents $1/1.8$ of ultimate, and the remaining fraction is $1-1/1.8$. The BF reserve applies that remaining fraction to the expected ultimate rather than extrapolating the observed paid amount. In contrast, chain ladder applies the factor to the observed amount. These are different sources of information, even when they use the same development pattern.

Worked Example: Compare Chain Ladder, BF, and a Credibility Blend

A hypothetical line has cumulative paid losses of $600,000 at the current age. The selected LDF to ultimate is 1.8. Earned premium is $1,000,000, and the selected expected loss ratio is 60%. A reserving team assigns credibility weight $Z=0.40$ to the chain ladder indication and $1-Z$ to BF. Calculate the chain ladder reserve, BF reserve, and blended reserve.

Assumptions. All amounts are undiscounted, with an effective annual interest rate of $i=0\%$. Mortality is not applicable because this is a non-life reserve calculation. No probability distribution is assumed for these point estimates; the selected LDF of 1.8, earned premium, loss ratio, and credibility weight are fixed inputs. The expected ultimate is on the same basis as the cumulative paid losses.

Step 1: Calculate the expected ultimate. The expected loss ratio is 0.60, so

E=1{,}000{,}000(0.60)=\$600{,}000.

Step 2: Calculate the chain ladder reserve. The projected ultimate is current paid losses multiplied by the LDF:

U_{\mathrm{CL}}=600{,}000(1.8)=\$1{,}080{,}000, \qquad R_{\mathrm{CL}}=1{,}080{,}000-600{,}000=\$480{,}000.

Step 3: Calculate the BF reserve. The unpaid proportion is $1-1/1.8=1-0.5556=0.4444$, using the exact fraction $4/9$:

R_{\mathrm{BF}}=600{,}000\left(1-\frac{1}{1.8}\right) =600{,}000\left(\frac{4}{9}\right) =\$266{,}666.67.

Step 4: Apply the stated credibility weights to reserves.

R_{\mathrm{blend}} =0.40(480{,}000)+0.60(266{,}666.67) =192{,}000+160{,}000 =\$352{,}000.00.

The selected blended reserve is $352,000.00. As a check, the BF ultimate is current paid plus BF unpaid, or $600,000 + $266,666.67 = $866,666.67. The blended ultimate is $0.40(1{,}080{,}000)+0.60(866{,}666.67)=\$952{,}000$; subtracting current paid of $600,000 gives the same blended reserve of $352,000. The method avoids treating the BF expected ultimate itself as the BF reserve.

Building a Multi-Line Selection

A portfolio-level result should preserve the basis of each line. A line with mature, stable experience may support a larger data-driven weight; a line with sparse or changing experience may rely more heavily on an expected ultimate. A credibility weight is not automatically determined by the LDF, and it should not be mistaken for a probability that one method is correct. Here, each weight is supplied as an assumption; the examples illustrate its application, not how to estimate it.

Worked Example: Select a Reserve Across Two Lines

Consider two hypothetical lines. Line A has $600,000 paid, LDF 1.8, expected ultimate $600,000, and credibility weight $Z_A=0.40$ on chain ladder. Line B has $300,000 paid, LDF 2.0, expected ultimate $900,000, and weight $Z_B=0.25$ on chain ladder. For each line, blend the chain ladder and BF reserves, then calculate the portfolio reserve.

Assumptions. All amounts are undiscounted at $i=0\%$; mortality is not applicable. No probability distribution is assumed for these point estimates. The two LDFs, expected ultimates, and line-specific weights are fixed inputs. Each line's expected ultimate and paid amount are stated on a consistent basis.

Line A. Its chain ladder reserve is $600{,}000(1.8-1)=\$480{,}000$. Its BF reserve is $600{,}000(1-1/1.8)=\$266{,}666.67$. Therefore,

R_A=0.40(480{,}000)+0.60(266{,}666.67) =192{,}000+160{,}000 =\$352{,}000.00.

Line B. Its chain ladder reserve is $300{,}000(2.0-1)=\$300{,}000$. Its BF unpaid proportion is $1-1/2=0.5$, so the BF reserve is $900{,}000(0.5)=\$450{,}000$. Applying $Z_B=0.25$ to chain ladder gives

R_B=0.25(300{,}000)+0.75(450{,}000) =75{,}000+337{,}500 =\$412{,}500.00.

Portfolio total. Add the line-level reserves:

R_{\mathrm{portfolio}}=352{,}000+412{,}500=\$764{,}500.00.

As a consistency check, the portfolio chain ladder reserve is $480,000 + $300,000 = $780,000. The portfolio BF reserve is $266,666.67 + $450,000 = $716,666.67. The blended total, $764,500, lies between those totals because each line's estimate is a convex weighted average. This check would fail if one line were accidentally weighted on an ultimate basis while the other were weighted on an unpaid-reserve basis.

Simulation: Reserve Variability Versus Monte Carlo Error

A point estimate answers a question about the selected central reserve, not how much future outcomes may vary around it. A predictive simulation can represent variation in future unpaid losses under an assumed model. Separately, a finite number of simulation runs creates Monte Carlo error in estimated summaries such as the simulated mean. These quantities are not interchangeable: predictive variability describes the modelled reserve outcomes, while Monte Carlo error describes numerical noise from using a finite simulation sample.

Definition: If simulated reserve outcomes $R_1,\ldots,R_n$ are independent draws from a predictive reserve distribution with standard deviation $\sigma_R$, the standard error of their sample mean is $\sigma_R/\sqrt{n}$. This standard error measures simulation error in estimating the mean; it is not the predictive standard deviation of a single reserve outcome.

Worked Example: Quantify Predictive and Simulation Uncertainty

For a hypothetical reserve model, future unpaid loss $R$, in thousands of dollars, has the following discrete predictive distribution: $182 with probability 0.10, $282 with probability 0.30, $382 with probability 0.40, and $482 with probability 0.20. A simulation uses 10,000 independent draws. An illustrative run produces counts 1,000, 3,000, 4,000, and 2,000 for those outcomes, respectively. Calculate the model mean and standard deviation, the mean from the stated run, its Monte Carlo standard error, and $P(R>400)$.

Assumptions. The model is the stated discrete distribution with the four outcomes and probabilities given. Simulation draws are independent from that distribution. Amounts are undiscounted at $i=0\%$; mortality is not applicable. The counts are illustrative simulation output, not observed industry data.

Step 1: Calculate the predictive mean. In thousands of dollars,

$$ E[R]=0.10(182)+0.30(282)+0.40(382)+0.20(482) =18.2+84.6+152.8+96.4=352. $$

Step 2: Calculate the predictive variance. The deviations from 352 are $-170,-70,30,130$, respectively. Thus,

\begin{aligned} \operatorname{Var}(R) &=0.10(170^2)+0.30(70^2)+0.40(30^2)+0.20(130^2)\\ &=2{,}890+1{,}470+360+3{,}380\\ &=8{,}100\ \text{(thousand dollars)}^2. \end{aligned}

The predictive standard deviation is $\sqrt{8{,}100}=90$ thousand dollars, or $90,000.

Step 3: Calculate the simulated mean. The run's weighted mean, in thousands, is

\frac{1{,}000(182)+3{,}000(282)+4{,}000(382)+2{,}000(482)}{10{,}000} =\frac{3{,}520{,}000}{10{,}000} =352.

The simulated mean is $352,000. The match with the model mean occurs because the illustrative counts reproduce the stated probabilities exactly; a real finite run need not do so.

Step 4: Calculate Monte Carlo standard error and exceedance probability.

\operatorname{SE}(\bar R)=\frac{90{,}000}{\sqrt{10{,}000}}=\$900, \qquad P(R>400)=P(R=482)=0.20.

The $90,000 standard deviation describes variation in an individual predictive reserve outcome. The $900 standard error describes the precision of the simulated mean under the stated independent-draw assumption. The exceedance probability is 0.2000 to four decimal places.

Interpreting and Selecting the Estimate

Method selection should follow the data and model assumptions rather than a preference for a particular formula. Chain ladder may be more responsive to the most recent observed experience, but that responsiveness can be undesirable when the latest data are sparse or distorted. BF can reduce dependence on immature observed losses, but it depends on a credible expected ultimate and a suitable emergence pattern. A credibility blend makes the trade-off explicit only if its weights have a defensible basis.

Simulation can add information about the range of outcomes, but it does not repair a poorly specified model. For example, a narrow simulated distribution can result from omitting parameter uncertainty or dependence, rather than from genuinely low reserve risk. The interpretation must state which sources of uncertainty are included. If simulation varies only future claim outcomes while holding development factors and expected loss assumptions fixed, its predictive spread is conditional on those fixed assumptions.

Align the bases.
Confirm that paid losses, expected ultimate losses, exposure, and valuation age refer to the same coverage, currency, and loss basis.

Compute comparable unpaid amounts.
Use $C(L-1)$ for chain ladder and $E(1-1/L)$ for BF, then apply any credibility weights to quantities on the same reserve basis.

Assess the sources of uncertainty.
Identify whether a simulation includes process variation, parameter uncertainty, or both, and state what is held fixed.

Make and explain the selection.
Relate the chosen estimate to maturity, data quality, expected-loss assumptions, and the decision purpose; do not treat a weighted average as automatically superior.

Common Exam Traps

Calling expected ultimate the BF reserve. BF reserve is expected ultimate multiplied by the expected unpaid proportion. In Worked Example 1, using $600,000 directly as the reserve is wrong; the BF reserve is $266,666.67.
Mixing ultimate and unpaid-reserve estimates. If ultimate estimates are blended, subtract current paid once after blending. Alternatively, blend unpaid reserves directly. Do not subtract paid from one component but not the other.
Using a credibility weight without stating what it weights. A weight on chain ladder ultimate estimates is not automatically a weight on chain ladder reserves unless both estimates are converted consistently.
Calling Monte Carlo error reserve risk. Increasing the number of runs reduces the standard error of a simulated mean at the rate $1/\sqrt{n}$, but it does not reduce the predictive variability of future claims.
Assuming a simulated mean validates a selected reserve. Simulation reproduces the assumptions supplied to it. Agreement between a simulated mean and a point estimate is not evidence that the assumptions are appropriate.

Key takeaway: Calculate each method on a consistent unpaid-loss basis, make credibility weights explicit, and distinguish the predictive distribution of reserves from the numerical error of a finite simulation. A reserve selection is defensible when its data basis and assumptions are clear.

Check Your Understanding

Use the stated assumptions in each problem. Amounts are undiscounted, the effective annual interest rate is $i=0\%$, and mortality is not applicable. Round money to the nearest cent where needed.

A hypothetical portfolio has $420,000 paid and an LDF of 1.6. Its expected ultimate is $700,000. Calculate the chain ladder and BF reserves. (Answer: chain ladder $252,000.00; BF $262,500.00)
A line has $250,000 paid, LDF 2.4, and expected ultimate $720,000. A stated credibility weight of 0.30 applies to chain ladder reserve and 0.70 to BF reserve. Calculate the blended reserve. (Answer: $399,000.00)
Two lines have chain ladder reserves of $180,000 and $260,000, and BF reserves of $220,000 and $200,000. Apply chain ladder weights of 0.50 and 0.25, respectively, to form each line's credibility-weighted reserve. Find the total. (Answer: $415,000.00)
A predictive reserve distribution, in thousands of dollars, is $100 with probability 0.20, $300 with probability 0.50, and $500 with probability 0.30. Find its mean and standard deviation. (Answer: mean $320,000; standard deviation $140,000)
A simulation uses 4,900 independent draws from a predictive reserve model whose standard deviation is $140,000. Calculate the Monte Carlo standard error of the simulated mean, and state whether it is the predictive standard deviation or the error in estimating the mean. (Answer: $2,000; it is the standard error of the simulated mean, not the predictive standard deviation)

What You'll Learn