Selecting Factors When Development Data Are Sparse

Why Sparse Development Data Require Care

At an older development age, only the most mature accident periods may have observations at both ages needed to calculate a link ratio. A factor based on one or two observations can be calculated exactly, yet still be highly sensitive to the particular claims in those periods. The key task is therefore not just to calculate a factor: it is to decide whether the available evidence is adequate to support it and to explain the selection.

A small number of observations does not automatically make a factor wrong, nor does it automatically justify replacing it with a factor from a younger age. The analyst should establish what data are actually available, calculate the available indications correctly, assess their limitations, and make any judgment explicit.

Definition: Let $C_{i,j}$ denote cumulative losses for accident period $i$ at development age $j$. An observed age-to-age link ratio is $r_i=C_{i,j+1}/C_{i,j}$, when both cumulative amounts are available and $C_{i,j}>0$. The volume-weighted factor for this transition is $f=\sum_i C_{i,j+1}/\sum_i C_{i,j}$, where the sums include only accident periods with both observations. The number of such periods is the number of usable observations supporting the factor.

What a Sparse Factor Can and Cannot Tell You

A factor supported by few periods has limited evidence about how the relationship between cumulative losses at the two ages behaves across accident periods. One unusually large or small ratio can materially change the result. In addition, the number of usable ratios is not the whole story: a few large-denominator periods can contribute much more to a volume-weighted factor than several small-denominator periods.

A useful first step is to make the data basis visible. Report the number of usable pairs, their individual ratios, their starting cumulative amounts, and the resulting selected indication. Then consider whether those periods are relevant to the current estimate and whether a suitable benchmark exists. A benchmark might come from a closely related segment or a broader group with a more stable history, provided the analyst can explain why the comparison is appropriate.

Do not treat a benchmark as automatically correct. Development patterns can differ by segment, coverage, claims practice, or data definition. Nor is a judgmental blend a universal statistical formula. If a blend is used, its inputs and weights should be stated as an assumption for the analysis, not presented as a standard rule.

Why Volume Weighting Gives a Different Indication

The volume-weighted factor describes the ratio of the total later-age cumulative losses to the total earlier-age cumulative losses across the included periods. It can also be understood as a weighted average of the individual link ratios. This explains why it need not equal the simple average when the starting cumulative amounts differ.

f=\frac{\sum_i C_{i,j+1}}{\sum_i C_{i,j}} =\frac{\sum_i C_{i,j}r_i}{\sum_i C_{i,j}} =\sum_i w_i r_i,\qquad w_i=\frac{C_{i,j}}{\sum_k C_{k,j}}.

To justify the result, use $C_{i,j+1}=C_{i,j}r_i$ for each usable period. Substituting into the numerator gives $\sum_i C_{i,j}r_i$. Dividing each term by the total starting amount shows that the factor is the weighted average of the ratios, with weights equal to each period’s share of the total starting cumulative losses. The weights are nonnegative and sum to one when the starting amounts are positive. A larger starting amount therefore has more influence on this particular indication.

That weighting is useful when the aim is to describe the aggregate change across the observed volume. It does not prove that the factor is reliable: with only a few periods, the observed experience may still be unrepresentative. The simple average answers a different question, giving each observed ratio equal weight regardless of its starting amount.

Worked Example: Three Observations at an Older Age

A hypothetical liability portfolio has three accident periods with observed cumulative losses at development ages 24 and 36 months. Calculate each ratio, the simple average, and the volume-weighted factor. Then assess whether the resulting factor is strongly supported.

Accident period	$C_{i,24}$	$C_{i,36}$
A	$1,200	$1,320
B	$800	$968
C	$400	$456

Step 1: Calculate the three ratios. Period A gives $1{,}320/1{,}200=1.100$. Period B gives $968/800=1.210$. Period C gives $456/400=1.140$.

Step 2: Calculate the simple average. The sum is $1.100+1.210+1.140=3.450$, so the simple average is $3.450/3=1.150$.

Step 3: Calculate the volume-weighted factor. The total at 24 months is $1{,}200+800+400=2{,}400$. The total at 36 months is $1{,}320+968+456=2{,}744$. Thus the factor is $2{,}744/2{,}400=1.143333\ldots$, or $1.143$ to three decimal places.

Assessment. The volume-weighted indication is $1.143$, while the simple average is $1.150$. The difference arises because the ratios are not identical and the periods have different starting amounts. Only three periods support the older-age indication, so $1.143$ is a calculated indication, not proof of a stable underlying pattern. If the three periods are relevant and there is no better-supported comparable evidence, selecting $1.143$ is explainable, provided its sparse basis is disclosed.

Check. Applying each period’s ratio to its starting amount returns the stated later amount: $1{,}200(1.100)=1{,}320$, $800(1.210)=968$, and $400(1.140)=456$. The later amounts also sum to $2,744.

Using a Benchmark Without Hiding the Judgment

If the few observed ratios vary materially, or if the available periods may not represent the current business, a relevant benchmark can help frame a selection. First calculate the indication from the sparse data. Then assess the benchmark’s comparability. If the selected factor lies between them, show exactly how the judgment was made.

For example, an analyst may state that a factor is selected as a weighted blend of a sparse-data indication and a benchmark. The weights are a transparent judgment for that analysis; they are not determined merely by counting observations. A weight should reflect the analyst’s assessment of relevance and reliability, and should not imply more precision than the evidence supports.

Worked Example: Two Ratios and a Comparable Benchmark

A hypothetical specialty insurer has only two usable periods for a 36-to-48-month transition. Their cumulative losses are $900 to $990 and $300 to $390. A broader, judged-comparable segment has a more stable indication of $1.120$. For this example, the analyst decides in advance to give 60% weight to the sparse-data indication and 40% to the benchmark. Calculate both indications, the selected factor, and a concise rationale.

Step 1: Calculate the individual ratios. The first is $990/900=1.100$. The second is $390/300=1.300$.

Step 2: Calculate the sparse-data indication. The simple average is $(1.100+1.300)/2=1.200$. The volume-weighted factor is $(990+390)/(900+300)=1{,}380/1{,}200=1.150$.

Step 3: Apply the stated judgmental weights. Using the volume-weighted sparse-data indication, the selection is $0.60(1.150)+0.40(1.120)=0.690+0.448=1.138$. The selected factor is $1.138$ to three decimal places.

Rationale and check. The selected value lies between $1.120$ and $1.150$, as a positive weighted blend should. The rationale is that the two observed periods are relevant but sparse, while the broader segment is judged comparable and more stable. The choice of 60% and 40% is an explicit judgment in this hypothetical analysis, not a standard credibility formula. The simple average of $1.200$ is not used in the blend because the stated method uses the volume-weighted indication; changing the indication would change the selection.

When Only One Period Is Available

One usable ratio deserves particular caution. It can be calculated, but there is no second observation at that age with which to assess the variation among periods. The ratio may also be especially sensitive when its starting cumulative amount is small. In such a situation, an appropriate comparable benchmark may be more informative than treating the single observed ratio as a dependable estimate.

If no relevant benchmark is available, the analyst should not manufacture certainty. State that the data provide only a single indication, explain the limitation, and make any selected factor conditional on the available evidence. The appropriate choice depends on the context; the arithmetic alone cannot decide whether the single ratio should be adopted.

Worked Example: A Single Small-Volume Observation

For a hypothetical book of equipment coverage, only one accident period has both 48- and 60-month cumulative losses: $200 and $260. A separate, broader group is considered comparable for this transition and has an indication of $1.120$. The analyst chooses an equal-weight blend for illustration. Calculate the single observed ratio and the selection, and explain why the single ratio is not adopted without qualification.

Step 1: Calculate the observed ratio. The ratio is $260/200=1.300$. It is based on one usable period and a starting amount of only $200.

Step 2: Calculate the stated blend. Equal weights give $0.50(1.300)+0.50(1.120)=0.650+0.560=1.210$. The selected factor is $1.210$.

Assessment and check. The selected value is exactly halfway between $1.300$ and $1.120$. The single-period ratio is high relative to the benchmark, but this comparison alone does not establish why they differ or which one is correct. The selection reflects the stated judgment that the one-period evidence and the comparable benchmark each receive half the weight. The result should be reported with that limitation, rather than described as a factor supported by a broad history.

Common Exam Trap: Confusing a Correct Calculation With a Strong Selection

A common error is to calculate a factor from the available rows and then describe it as reliable simply because the arithmetic is correct. Another is to use the simple average when the question asks for a volume-weighted factor. In the first worked example, the incorrect simple-average calculation would give $1.150$ as the volume-weighted factor. The correct calculation is $2{,}744/2{,}400=1.143333\ldots$, or $1.143$ to three decimal places.

Even after calculating $1.143$ correctly, it would be an overstatement to claim that three observations establish a stable older-age pattern. The correction is to distinguish the numerical indication from the selection judgment: report the method, the number of usable periods, the relevant limitation, and why the chosen factor is supportable in context.

Count usable pairs, not triangle rows. A period without both required cumulative amounts does not supply a ratio for that transition.
Keep the weighting method clear. A simple average gives equal weight to ratios; a volume-weighted factor weights by the earlier-age cumulative amounts.
Do not borrow a factor without justification. A factor at another age or in another segment is only a benchmark if its relevance is explained.
Label judgmental blends. State the indications and weights, and do not present chosen weights as a universal formula.
Do not overstate precision. Rounding a sparse indication to several decimal places does not make it more reliable.

Key takeaway: For a sparse development age, calculate the available indication correctly, show how many periods support it, assess the relevance of the observed experience, and explain any benchmark or judgment used in the selection. A defensible factor includes both a number and a clear account of its evidence.

Check Your Understanding

For each question, calculate the requested indication and consider what the sparse data support.

A hypothetical property portfolio has two usable 24-to-36-month pairs: $500 to $550 and $100 to $130. Calculate both individual ratios, the simple average, and the volume-weighted factor. (Answer: Ratios $1.100$ and $1.300$; simple average $1.200$; volume-weighted factor $680/600=1.1333\ldots$, or $1.133$ to three decimals.)
Three usable pairs at an older age have starting amounts of $400, $400, and $200, with ratios $1.05$, $1.10$, and $1.20$, respectively. Calculate the simple average and volume-weighted factor. (Answer: Simple average $1.1167$, or $1.117$; volume-weighted factor $(400(1.05)+400(1.10)+200(1.20))/1{,}000=1.100$.)
A sparse-data volume-weighted indication is $1.160$, and a relevant benchmark is $1.110$. An analyst specifies weights of 40% and 60%, respectively. Calculate the judgmental blend and state whether the result lies between the indications. (Answer: $0.40(1.160)+0.60(1.110)=1.130$; yes, it lies between $1.110$ and $1.160$.)
At a 48-to-60-month transition, only one period is usable: $250 to $300. A comparable benchmark is $1.140$, and the analyst specifies a 25% weight on the single-period indication and 75% on the benchmark. Calculate the selected factor and name one limitation to disclose. (Answer: The single ratio is $300/250=1.200$; the blend is $0.25(1.200)+0.75(1.140)=1.155$. Disclose that only one period supports the observed ratio.)
Two usable pairs are $1,000 to $1,080 and $200 to $260. A candidate reports the simple average as the volume-weighted factor. Calculate the correct volume-weighted factor and identify the error. (Answer: Correct factor $=(1{,}080+260)/(1{,}000+200)=1{,}340/1{,}200=1.1167$, or $1.117$; the candidate used an unweighted simple average instead of weighting by starting cumulative amounts.)

What You'll Learn