CAS Exam 5 · Chain Ladder Estimation · Tutorial 693 of 989

Chain Ladder Sensitivity and Reasonableness Workshop

Learn how factor choices affect chain ladder ultimates and reserves, and how triangle diagnostics can distinguish a plausible selection from a mechanically calculated one.

Intermediate 12 min read

What You'll Learn

Calculate chain ladder ultimates and reserves under alternative factor selections
Compare simple averages with volume-weighted age-to-age factors
Quantify the effect of factor changes on accident-year and total reserves
Check selected factors against observed triangle patterns and data limitations
Apply the tower property using nested information sets in a chain ladder justification

Why Chain Ladder Estimates Need a Reasonableness Check

A chain ladder calculation can be arithmetically correct while its selected development factors are difficult to defend. A small change in a factor can materially change an immature accident year’s projected ultimate, especially when several future factors compound. Sensitivity analysis measures that effect; a reasonableness review asks whether the selected factors are consistent with the observed triangle and whether the method’s assumptions remain credible.

This workshop treats a chain ladder estimate as a model-based projection, not as a guarantee of future claims. All examples use hypothetical cumulative losses. Interest, mortality, and a claim severity distribution are not used; the calculations are deterministic projections from selected factors.

Definition: Let $C_{i,j}$ be cumulative losses for accident year $i$ through development age $j$. An age-to-age factor $f_j$ projects cumulative losses from age $j$ to age $j+1$. If the last modeled age is $J$, and $T$ is a tail factor beyond that age, the cumulative development factor from age $j$ to ultimate is $$ L_j=\left(\prod_{r=j}^{J-1} f_r\right)T. $$ For an accident year observed through age $k$, the chain ladder projected ultimate is $\hat U_i=C_{i,k}L_k$, and the estimated unpaid reserve is $\hat R_i=\hat U_i-C_{i,k}$. A tail factor is set to 1.00 when no additional development beyond the last modeled age is assumed.

Factor selection is a judgment informed by the observed age-to-age ratios, the amount and quality of data, and whether the historical development pattern is relevant to the period being projected. Common selection bases include a simple average of observed ratios and a volume-weighted average. Neither is automatically appropriate in every triangle. Sensitivity analysis compares plausible alternatives; it does not decide which alternative is best without considering the underlying data.

How Development Factors Compound

A projected ultimate at an early age depends on every selected factor between the current age and ultimate. Consequently, the effect of changing one factor depends on how much of the triangle remains to develop. A change to a factor that applies to several immature years may affect the total reserve much more than the same-sized change to a factor that applies to only one year.

Justification: The multiplication of successive factors can be justified as a conditional-expectation result when the information sets are nested. Suppose $\mathcal F_j$ contains the information available through development age $j$, with $\mathcal F_j\subseteq\mathcal F_{j+1}$, and the one-step assumptions are $$ E[C_{i,j+1}\mid\mathcal F_j]=f_jC_{i,j}, \qquad E[C_{i,j+2}\mid\mathcal F_{j+1}]=f_{j+1}C_{i,j+1}, $$ where the selected factors are treated as fixed. By the tower property, $$ \begin{aligned} E[C_{i,j+2}\mid\mathcal F_j] &=E\!\left[E[C_{i,j+2}\mid\mathcal F_{j+1}]\mid\mathcal F_j\right]\\ &=E[f_{j+1}C_{i,j+1}\mid\mathcal F_j]\\ &=f_{j+1}f_jC_{i,j}. \end{aligned} $$ The nested conditioning is essential: separate one-step statements without a specified relationship between their conditioning information do not, by themselves, establish this two-step result. In practical chain ladder work, factors are estimated from triangle data, so this calculation explains the projection structure under the stated assumptions rather than proving that the selected factors are correct.

Worked Example: Measuring the Effect of Alternative Factors

A hypothetical portfolio has the following cumulative losses, in thousands of dollars. The selected factors are 1.50 from 12 to 24 months, 1.20 from 24 to 36 months, and 1.10 from 36 to 48 months. Assume a tail factor of 1.00 after 48 months. Compare the resulting reserve with an alternative selection of 1.25 from 24 to 36 months and 1.15 from 36 to 48 months, holding the 12-to-24 factor fixed.

Accident year	12 months	24 months	36 months	48 months
Year 1	100	150	180	198
Year 2	120	180	216	—
Year 3	140	210	—	—

Method and assumptions. Apply the selected factors to each year’s latest observed cumulative loss. All amounts are in thousands of dollars. This is a deterministic hypothetical calculation; no interest, mortality, or claim distribution is assumed.

Base selection. Year 1 has reached 48 months, so its projected ultimate is 198 and its estimated reserve is $198-198=0$. Year 2 is at 36 months, so its age-36-to-ultimate factor is $1.10$: $$ \hat U_2=216(1.10)=237.60,\qquad \hat R_2=237.60-216=21.60. $$ Year 3 is at 24 months. Its cumulative factor is $1.20(1.10)=1.32$, giving $$ \hat U_3=210(1.32)=277.20,\qquad \hat R_3=277.20-210=67.20. $$ The total base reserve is $0+21.60+67.20=88.80$ thousand dollars.

Alternative selection. Year 2’s revised age-36-to-ultimate factor is 1.15: $$ \hat U_2=216(1.15)=248.40,\qquad \hat R_2=248.40-216=32.40. $$ For Year 3, the revised cumulative factor is $1.25(1.15)=1.4375$: $$ \hat U_3=210(1.4375)=301.875,\qquad \hat R_3=301.875-210=91.875. $$ The alternative total reserve is $32.40+91.875=124.275$ thousand dollars. The increase over the base selection is $124.275-88.80=35.475$ thousand dollars.

Check and common mistake. Each projected ultimate is at least its current cumulative loss, and subtracting the current losses gives the stated reserves. A common error is to apply only the 24-to-36 factor to Year 3, omitting the subsequent 36-to-48 factor. That would give $210(1.25)=262.50$, not the ultimate under the stated 48-month endpoint. The factor sequence must continue to the assumed ultimate age.

Comparing Factor Selection Methods

For observed ratios $r_m=C_{m,j+1}/C_{m,j}$ at a particular development age, the simple average gives each observed ratio equal weight. The volume-weighted average, often called the ratio of sums, gives more weight to observations with larger cumulative losses: $$ \bar r_{\text{simple}}=\frac{1}{n}\sum_{m=1}^{n}r_m, \qquad \bar r_{\text{weighted}}= \frac{\sum_{m=1}^{n}C_{m,j+1}}{\sum_{m=1}^{n}C_{m,j}}. $$

These methods answer different practical questions. A simple average gives each included accident year the same weight, while a ratio of sums weights years in proportion to their denominator cumulative losses. A large year can therefore have substantial influence on the weighted factor. Inspect both the calculated factor and the individual ratios; a summary statistic can conceal a pattern or an unusual observation.

Worked Example: Simple Average Versus Volume-Weighted Factor

A hypothetical triangle has 12-to-24-month cumulative losses of 100 and 150 thousand dollars, developing respectively to 140 and 240 thousand dollars. At 24-to-36 months, the first year develops from 168 to 201.6 thousand dollars. The latest accident year is at 12 months with cumulative losses of 200 thousand dollars. Select the 12-to-24 factor by each method, use 1.20 for the 24-to-36 factor, and compare the latest year’s reserve. Assume no tail beyond 36 months.

Method and assumptions. Calculate the two observed 12-to-24 ratios and then project the latest year through both development ages. Amounts are in thousands of dollars; the hypothetical calculation uses no interest, mortality, or claim distribution.

The observed ratios are $140/100=1.40$ and $240/150=1.60$. The simple average is $$ (1.40+1.60)/2=1.50. $$ The volume-weighted factor is $$ \frac{140+240}{100+150}=\frac{380}{250}=1.52. $$ Using the simple average, the latest year’s projected ultimate is $200(1.50)(1.20)=360$, so its reserve is $360-200=160$ thousand dollars. Using the weighted factor, the projected ultimate is $200(1.52)(1.20)=364.80$, so its reserve is $364.80-200=164.80$ thousand dollars. The weighted selection produces a reserve $164.80-160=4.80$ thousand dollars higher.

Check and common mistake. The weighted factor is $380/250$, not the average of 140 and 240 divided by the average of 100 and 150 by some additional weighting step; those averages happen to produce the same ratio here, but the ratio-of-sums definition is the method. Do not assume the weighted answer is better simply because it uses more dollars of experience. The larger observation may represent a different risk mix or an unusual year, so the underlying ratios still need review.

Reasonableness Checks and Model Limitations

A useful review connects the selected factors to the triangle rather than treating them as inputs that need no explanation. A factor selection should be compared with the observed ratios at that age, the number of observations, and whether the pattern is stable enough to support projection. In particular, an apparently precise factor based on very few observations may have substantial uncertainty.

Recalculate the observed ratios.
For each available accident year, divide the next-age cumulative amount by the current-age amount. Check the age labels and confirm that the numerator and denominator refer to the same accident year.

Compare candidate selections.
Calculate the selected factor using the stated method. Compare it with individual ratios and reasonable alternatives, such as a simple average versus a volume-weighted average, when the data support that comparison.

Trace the effect to ultimate and reserve.
Apply each factor only to the accident years that have not yet reached that development age. Recalculate the full remaining factor product before subtracting current cumulative losses.

Ask whether the historical pattern remains relevant.
Consider whether changes in claim handling, case reserve practices, settlement speed, claim mix, or calendar-year conditions could make historical development a poor guide to future development.

Worked Example: Investigating a High Recent Link Ratio

A hypothetical triangle, in thousands of dollars, has these cumulative losses: Year 1: 100, 120, 132; Year 2: 100, 130, 143; Year 3: 100, 180; and Year 4: 100. The entries are at 12, 24, and 36 months, where available. For 24-to-36 months, select 1.10. Compare projecting Year 4 using the simple average of all 12-to-24 ratios with using the average of the two most recent observed ratios. Assume no tail beyond 36 months.

Method and assumptions. First calculate each observed 12-to-24 ratio, then calculate each candidate ultimate and the combined reserve for the two immature years. Amounts are in thousands of dollars; no interest, mortality, or claim distribution is used.

The observed ratios are $120/100=1.20$, $130/100=1.30$, and $180/100=1.80$. The simple average of all three is $$ (1.20+1.30+1.80)/3=4.30/3=1.4333\ldots. $$ Year 3 is at 24 months, so its ultimate is $180(1.10)=198$, and its reserve is $198-180=18$. Year 4’s ultimate under the all-year average is $$ 100(1.4333\ldots)(1.10)=157.6666\ldots, $$ so its reserve is $157.6666\ldots-100=57.6666\ldots$. The combined reserve is $18+57.6666\ldots=75.6666\ldots$ thousand dollars.

The average of the two most recent ratios is $(1.30+1.80)/2=1.55$. Year 4’s projected ultimate becomes $100(1.55)(1.10)=170.50$, and its reserve is $170.50-100=70.50$. Year 3’s estimate remains unchanged, so the combined reserve is $18+70.50=88.50$ thousand dollars. The recent-two selection is higher by $88.50-75.6666\ldots=12.8333\ldots$ thousand dollars. Rounded only at the final step to two decimal places, the respective combined reserves are 75.67 and 88.50 thousand dollars.

Check and common mistake. The higher factor changes Year 4, which is at 12 months; it does not change Year 3’s age-24 projection. A common mistake is to apply the 12-to-24 factor to Year 3 as well, even though Year 3 has already emerged through 24 months. The factor must match the current development age.

The high 1.80 ratio deserves investigation, not automatic acceptance or rejection. It could reflect a large claim, a change in claim mix, a calendar-year influence, or a genuine shift in development. The triangle alone does not identify the cause. Choosing the recent-two average solely because it is more recent would be as mechanical as retaining the all-year average without examining the observation.

Common Exam Traps

The most frequent error is to report a projected ultimate as if it were a reserve. The reserve is the projected ultimate less the cumulative losses already observed. Another is to apply a factor to an accident year already developed beyond the factor’s starting age, or to stop a factor product before the assumed ultimate age. Always label the current age, write the remaining factor product, and then subtract the current cumulative value.

Key takeaway: A defensible chain ladder estimate requires both correct arithmetic and a reasoned factor selection. Quantify the reserve’s sensitivity to plausible alternatives, verify that each factor applies to the correct age, and assess whether the observed development pattern supports the projection. The chain ladder result is limited when development patterns are unstable, data are sparse or distorted, or historical conditions do not represent future claim emergence.

Check Your Understanding

Work each problem before checking the answer. All values are hypothetical and in thousands of dollars unless stated otherwise.

A latest accident year has cumulative losses of 80 at 12 months. The selected 12-to-24 and 24-to-36 factors are 1.25 and 1.10, with no tail beyond 36 months. Find the projected ultimate and reserve. (Answer: Ultimate 110; reserve 30.)
Two observed 12-to-24 ratios are 1.40 on a denominator of 100 and 1.60 on a denominator of 200. Use the volume-weighted factor, then a 24-to-36 factor of 1.10, to project a latest 12-month cumulative value of 150. What are the ultimate and reserve? (Answer: Weighted factor 1.5333; ultimate 253; reserve 103.)
A year has cumulative losses of 500 at 24 months. Compare the ultimate and reserve using either 1.20 or 1.30 for 24-to-36 months, followed in both cases by 1.10 for 36-to-48 months. Assume no tail. (Answer: With 1.20, ultimate 660 and reserve 160; with 1.30, ultimate 715 and reserve 215.)
Observed 12-to-24 ratios are 1.10, 1.20, and 1.50. The latest year is at 12 months. What is the all-year simple average, and why is it not automatically correct to select only the latest ratio? (Answer: Simple average 1.2667; the latest ratio may be unusual, and its cause and relevance to future development must be assessed.)
Explain the conditioning requirement for multiplying two successive chain ladder factors in an expected-value justification. (Answer: The information sets must be nested, with $\mathcal F_j\subseteq\mathcal F_{j+1}$, and the one-step conditional expectations specified with respect to those sets; then the tower property gives the product of the two factors.)

Back to Tutorials