Why Competing Tail Estimates Need Reconciliation
A tail estimate represents development beyond the latest age for which losses are evaluated. It can materially affect an ultimate projection, yet the information available for estimating it is often sparse. An extrapolation of late link ratios, an external benchmark, and an indication from the subject book’s historical experience may therefore give different answers. Reconciliation means investigating those differences and selecting an estimate that is supported by comparable evidence—not simply averaging every number presented.
This tutorial is aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus. The emphasis is on comparing estimates and documenting a selection. It assumes familiarity with development ages, link ratios, and tail factors.
The starting age matters because a factor from age 12 to ultimate includes development between ages 12 and 24, while a factor from age 24 to ultimate does not. The loss basis and scope matter as well: two estimates may differ because they apply to different coverage, claim populations, or definitions of ultimate rather than because one method is intrinsically better. Before comparing the numerical values, establish what each one measures.
A Consistent Comparison Process
The main sources of an indication may include extrapolation of late-age development, an external benchmark, and historical experience for the subject business. Reconciliation begins by recording each source’s age, scope, assumptions, and limitations. If an estimate is not on the target basis, convert it when the required information is available; otherwise, do not treat it as directly comparable.
Specify the starting development age, loss basis, scope, and definition of ultimate for the factor being selected.
Record how each estimate was developed, the information it uses, and the assumptions or limitations that could affect its relevance.
Convert factors to the target starting age when possible. Do not blend a factor on one age basis with factors on another.
Check whether disagreement is consistent with differences in scope, data quality, time period, or model assumptions. A large difference is a reason to investigate, not by itself a reason to discard an estimate.
Use the evidence to support a point estimate or range. State the rationale, any weights used, and what uncertainty or unresolved basis issue remains.
A weighted average is one possible way to combine estimates already judged comparable. If the estimates are \(T_1,\ldots,T_n\), assign nonnegative weights \(w_1,\ldots,w_n\) that sum to 1. The resulting selection is
The weights express the relative importance assigned to the sources in this selection. They are not automatically credibility weights, and using a weighted average does not resolve a mismatch in age or scope. The analyst must support the weights by considering relevance and reliability, not choose them merely to produce a desired result.
Why a Weighted Average Can Be a Reasonable Summary
There is a simple mathematical justification for a weighted average when all estimates address the same quantity. Suppose \(T_j\) is the estimate from source \(j\), and the chosen weights are nonnegative. Consider selecting a value \(T\) that minimizes the weighted squared distance from the source estimates:
Differentiating with respect to \(T\) gives
At a minimum, this derivative is zero. If \(\sum_j w_j>0\), solving the resulting equation gives
When the weights sum to 1, this is exactly the weighted average above. The second derivative is \(2\sum_j w_j>0\), so the solution is a minimum. This derivation explains how a weighted average summarizes estimates under a stated squared-distance criterion; it does not establish that the estimates are unbiased, independent, or equally well supported. Those judgments still require actuarial review.
Worked Examples
Worked Example: Combine Extrapolation, Benchmark, and Historical Evidence
A hypothetical liability portfolio is being evaluated at age 60. A tail extrapolation supplies four successive remaining link ratios of 1.020, 1.014, 1.010, and 1.007. Assume these are all the modeled remaining periods and that no material development remains after the last period. The external benchmark and subject-book historical indication, already stated on the same age-60-to-ultimate basis and scope, are 1.060 and 1.048. The analyst assigns weights of 50%, 20%, and 30% to extrapolation, benchmark, and historical experience, respectively. Calculate the extrapolated tail and the weighted selection. The weights are illustrative professional judgments, not formal credibility estimates.
Method. Multiply the successive link ratios to obtain the extrapolated tail-to-ultimate factor. Then combine the three same-basis estimates using the stated weights.
Step 1: Calculate the extrapolated factor. Multiply the first two ratios: \(1.020\times1.014=1.034280\). Multiply by the third ratio: \(1.034280\times1.010=1.0446228\). Multiply by the fourth ratio: \(1.0446228\times1.007=1.0519351596\). Thus the extrapolated estimate is 1.0519351596, or 1.0519 to four decimal places.
Step 2: Check the weights. The weights sum to \(0.50+0.20+0.30=1.00\), so they are normalized.
Step 3: Calculate the weighted selection. The weighted contributions are \(0.50\times1.0519351596=0.5259675798\), \(0.20\times1.060=0.2120000000\), and \(0.30\times1.048=0.3144000000\). Their sum is \(1.0523675798\). The selected factor is therefore 1.0524 to four decimal places.
Check and mistake avoided. The result, 1.0524, lies between the smallest and largest input estimates, 1.048 and 1.060, as a normalized weighted average must. A common mistake is to add the successive link ratios instead of multiplying them; these are sequential multiplicative factors, so addition would not produce a tail-to-ultimate factor.
Worked Example: Convert a Benchmark Before Blending
A hypothetical external benchmark gives a factor of 1.072 from age 36 to ultimate. For the subject portfolio, the selected factor from age 36 to age 48 is 1.030. At age 48, the subject’s extrapolated tail estimate is 1.038 and its historical indication is 1.042. Assign weights of 40% to the converted benchmark, 35% to extrapolation, and 25% to historical experience. Calculate the age-48 selected factor.
Method. First convert the age-36 benchmark to an age-48-to-ultimate factor using the multiplicative relationship between the two consecutive development periods. Then blend only the age-48 estimates.
Step 1: State the relationship. The factor from age 36 to ultimate is the product of the factor from age 36 to age 48 and the factor from age 48 to ultimate:
Step 2: Convert the benchmark. Solve for the factor from age 48 to ultimate: \(T_{48}=T_{36}/F_{36,48}=1.072/1.030=1.0407766990\ldots\). The converted benchmark is 1.0408 to four decimal places.
Step 3: Check and apply the weights. The weights sum to \(0.40+0.35+0.25=1.00\). Using the unrounded converted benchmark, the weighted contributions are \(0.40\times1.0407766990=0.4163106796\), \(0.35\times1.038=0.3633000000\), and \(0.25\times1.042=0.2605000000\). Their sum is \(1.0401106796\), so the selected factor is 1.0401 to four decimal places.
Check and mistake avoided. The converted benchmark is smaller than 1.072 because the age-36-to-age-48 factor of 1.030 accounts for development already included in the broader benchmark. Multiplying \(1.072\times1.030\) would count that development again and would incorrectly give 1.10416. Rounding the converted benchmark before weighting can also cause avoidable differences; keep full precision until the final step.
Worked Example: Reconcile a High Estimate That Is Not Comparable
A hypothetical portfolio is being evaluated at age 72. Three estimates on the same age-72-to-ultimate basis and scope are: extrapolation 1.055, a current external benchmark 1.060, and historical experience 1.058. A fourth, older benchmark is 1.090, but it runs from age 48 to ultimate and no factor is available to convert it to age 72. The analyst assigns weights of 40%, 35%, and 25% to the first three estimates, respectively. Select a point estimate and a defensible range from the comparable indications.
Method. Keep the older benchmark separate because its starting age does not match and the data needed to convert it are unavailable. Calculate the weighted selection from the three comparable estimates, and use their spread to describe a simple sensitivity range.
Step 1: Confirm the usable basis. The estimates 1.055, 1.060, and 1.058 all refer to age 72, the same scope, and the same definition of ultimate. The 1.090 estimate does not: it includes development from age 48 onward. Without a conversion factor, it cannot be treated as a direct age-72 indication.
Step 2: Check the weights. \(0.40+0.35+0.25=1.00\).
Step 3: Calculate the selected point estimate. The weighted contributions are \(0.40\times1.055=0.4220\), \(0.35\times1.060=0.3710\), and \(0.25\times1.058=0.2645\). The sum is \(0.4220+0.3710+0.2645=1.0575\). The point estimate is 1.0575, or 1.058 to three decimal places.
Step 4: State the range. The lowest and highest of the three comparable estimates are 1.055 and 1.060. A simple supported sensitivity range is therefore 1.055 to 1.060. This range describes the spread of these indications; it is not a statistical confidence interval.
Check and mistake avoided. The point estimate lies within the range of the comparable inputs. Excluding the 1.090 benchmark is justified here by its unresolved age-basis mismatch, not merely by its being high. If conversion information becomes available, the benchmark can be evaluated on the target basis and reconsidered.
Interpreting Disagreement and Avoiding Traps
Disagreement between tail estimates is useful diagnostic information. Before making a selection, determine whether it may reflect different starting ages, scopes, data periods, or assumptions about how development continues. Consider the quality and relevance of each source and whether the extrapolation is sensitive to its chosen pattern. A benchmark may be valuable context but still require conversion; subject-book history may be relevant but limited by the amount or consistency of its experience.
Do not assume that one method should always dominate. Nor does a simple average resolve a meaningful discrepancy: it can obscure a basis error or give too much influence to a weakly supported source. If an estimate cannot be converted, identify that limitation and explain whether it is used only as context or excluded from the numerical selection. If the evidence supports several plausible values rather than one clearly preferred value, report a range and explain how it was chosen.
| Issue | Why it matters | Appropriate response |
|---|---|---|
| Different starting ages | The factors cover different amounts of development. | Convert to a common age when required factors are available; otherwise, do not blend directly. |
| Different scope or ultimate definition | The estimates may not describe the same target quantity. | Clarify the scope and determine whether a defensible adjustment is possible. |
| Large disagreement between comparable estimates | It may indicate sensitivity, data limitations, or different assumptions. | Investigate the sources and reflect the remaining uncertainty in the selection or range. |
| Weights that do not sum to 1 | The result is not the stated weighted average. | Normalize the weights or divide the weighted sum by their total. |
A frequent exam trap is to treat every supplied factor as if it were already on the same basis. Another is to multiply when converting a broader benchmark to a later development age, thereby counting earlier development twice. State the factor relationship before converting, verify that the weights sum to 1, and retain precision until the final rounding. A defensible reconciliation is transparent about both the selected value and the evidence that does not support a direct comparison.
Check Your Understanding
For each problem, compare like with like before calculating, and round only as requested.
- A hypothetical portfolio’s remaining modeled link ratios are 1.018, 1.011, and 1.006, with no material development assumed afterward. Calculate the extrapolated tail-to-ultimate factor to four decimal places. (Answer: 1.0354)
- An external benchmark is 1.084 from age 12 to ultimate. The subject portfolio’s age-12-to-age-24 factor is 1.040. Convert the benchmark to an age-24-to-ultimate factor, to four decimal places. (Answer: 1.0423)
- Three comparable estimates are 1.047, 1.052, and 1.060, with respective weights 50%, 30%, and 20%. Verify the weights and calculate the weighted selection to four decimal places. (Answer: 1.0511)
- Three estimates on the same target basis are 1.062, 1.067, and 1.081. A documented selection assigns respective weights of 40%, 40%, and 20%. Calculate the point estimate to four decimal places and give the range spanning the three indications. (Answer: 1.0678; range 1.062 to 1.081)