Extrapolate Unobserved Link Ratios to Estimate the Tail

Why Extrapolate Late-Age Link Ratios?

A development triangle often contains fewer observations at later ages: the newest accident years have not yet reached those ages. But a selected development pattern may still indicate that losses continue to emerge after the last age with reliable data. Extrapolating late-age link ratios provides a way to estimate that unobserved development and summarize it as a tail factor.

The central judgment is not just how to extend a numerical pattern. It is whether the observed pattern is sufficiently stable to support an extension, and where the extension should stop. A small error in several late-age factors can compound, while an implausible trend can produce a misleading tail.

Definition: Let $C_j$ denote cumulative loss at development age $j$. The link ratio from age $j$ to age $j+1$ is $f_j=C_{j+1}/C_j$, when $C_j\ne 0$. If age $m$ is the latest age used as the starting point, the tail factor through age $n$ is the product $T_{m:n}=\prod_{j=m}^{n-1}f_j$. It converts cumulative loss at age $m$ to the projected amount at age $n$, provided the selected factors apply to that loss measure and population.

A tail factor is therefore not one additional link ratio. It is the cumulative multiplier for all the selected, unobserved age-to-age development from the starting age to the chosen endpoint. If development beyond that endpoint is regarded as negligible, the endpoint serves as the practical ultimate age. The assumptions and endpoint should be stated, rather than implying that a finite projection proves no further development can occur.

Extending a Late-Age Pattern

A simple approach is to examine the changes in consecutive late-age link ratios. If those ratios have been decreasing by roughly a constant amount, continue that change for the unobserved ages. For example, ratios of 1.070, 1.055, and 1.040 show a decrease of 0.015 at each successive age. Continuing the pattern gives 1.025 at the next age. This is a linear extrapolation of the link ratios, not a guarantee that their true future behavior is linear.

When the observed ratios do not decline by exactly the same amount, a straight line can be fitted to the observed ratios or to their excess above 1. Let $x$ be a numerical index for development age and let $y_x=f_x-1$. A fitted line $y_x=a+bx$ forecasts the excess over 1 at unobserved ages. The forecast link ratio is then $1+\widehat y_x$. In ordinary least squares, $a$ and $b$ are chosen to minimize the sum of squared differences between observed excesses and fitted excesses. This method summarizes a trend; it does not remove the need to assess whether the trend is credible.

\widehat{f}_x=1+\widehat{y}_x=1+\widehat{a}+\widehat{b}x, \qquad \widehat{T}_{m:n}=\prod_{j=m}^{n-1}\widehat{f}_j

The expression $f_x-1$ is often useful because a link ratio of 1.000 represents no further cumulative development. It is not a universal requirement that all selected ratios exceed 1: for some loss measures, recoveries, adjustments, or data issues can produce a different pattern. For positive cumulative paid losses with expected further payments, however, an extrapolation that quickly gives ratios below 1 merits particular scrutiny.

Why the Extrapolated Factors Are Multiplied

The product follows directly from the definition of a link ratio. Each factor converts cumulative loss at one age to cumulative loss at the next. Applying successive conversions therefore requires multiplying the factors, just as each new balance is obtained by multiplying the previous balance by its growth factor.

C_{m+1}=C_m f_m,\quad C_{m+2}=C_{m+1}f_{m+1}=C_m f_m f_{m+1}, \quad \ldots,\quad C_n=C_m\prod_{j=m}^{n-1}f_j

This identity also explains why the tail factor measures cumulative development from the starting age: substituting each age-to-age relationship cancels the intermediate cumulative amounts. Adding the factors, or adding their excesses above 1, does not perform those successive conversions.

Worked Example: Continue a Constant Decline

A hypothetical portfolio has selected late-age link ratios of 1.072 from ages 7 to 8, 1.054 from ages 8 to 9, and 1.036 from ages 9 to 10. Assume the decline in the ratios continues at a constant amount through ages 10 to 12, and assume development beyond age 12 is negligible. Estimate the tail factor from age 10 to age 12.

Method. Continue the observed arithmetic change in the link ratios. The two observed changes are $1.054-1.072=-0.018$ and $1.036-1.054=-0.018$, so the assumed change per age is $-0.018$.

Project the unobserved factors. From age 10 to 11, the forecast is $1.036-0.018=1.018$. From age 11 to 12, it is $1.018-0.018=1.000$.

Multiply through the endpoint. There are two factors between ages 10 and 12, so $\widehat T_{10:12}=1.018\times1.000=1.018$. The selected tail factor is 1.018, or 1.018 to three decimal places.

Check. The projected factors continue the observed decrease, and their product exceeds 1 because the first projected factor allows for additional development. The assumed final factor of 1.000 represents no further change from age 11 to 12; it does not assert that all portfolios develop this way. The result depends on the stipulated linear pattern and endpoint.

Fitting a Line When the Pattern Is Uneven

Observed link ratios rarely trace a perfect line. A least-squares line can reduce the influence of one irregular step, but its extrapolation still relies on a substantive assumption: that the linear pattern remains a useful approximation over the forecast ages. The following example fits the line explicitly and keeps the unrounded forecasts through the product.

Worked Example: Fit a Straight Line to Late-Age Excesses

For a hypothetical book, the selected link ratios at three consecutive late ages are 1.100, 1.090, and 1.070. Label their ages $x=0,1,2$, respectively. Fit a least-squares line to $y_x=f_x-1$, and use it to project the next two link ratios. Assume the tail ends after those two projected factors.

Convert to excesses. The observed values are $y_0=0.100$, $y_1=0.090$, and $y_2=0.070$. Their means are $\bar{x}=1$ and $\bar{y}=(0.100+0.090+0.070)/3=0.0866667$.

Calculate the slope. The least-squares slope is $b=\sum (x-\bar{x})(y-\bar{y})/\sum (x-\bar{x})^2$. The numerator is $(-1)(0.100-0.0866667)+(0)(0.090-0.0866667)+(1)(0.070-0.0866667)=-0.0300000$. The denominator is $(-1)^2+0^2+1^2=2$. Thus $b=-0.015$.

Calculate the intercept. Since $a=\bar y-b\bar x$, $a=0.0866667-(-0.015)(1)=0.1016667$. The fitted line is $\widehat y_x=0.1016667-0.015x$.

Forecast the next factors. At $x=3$, $\widehat y_3=0.1016667-0.045=0.0566667$, so $\widehat f_3=1.0566667$. At $x=4$, $\widehat y_4=0.1016667-0.060=0.0416667$, so $\widehat f_4=1.0416667$.

Form the tail. Multiply the two forecast link ratios: $\widehat T=1.0566667\times1.0416667=1.1006944$, or 1.101 to three decimal places. The underlying product, before rounding, is approximately 1.1006944.

Check. Both forecasts continue the declining trend while remaining above 1. The line fit smooths the observed unequal decreases of 0.010 and 0.020; it does not establish that the future pattern will follow the fitted line. The specified endpoint is part of the estimate.

Applying a Tail Factor and Checking Its Scale

Once selected, a tail factor multiplies cumulative loss at its starting development age to estimate the amount at the chosen endpoint. This use is simply the age-to-age identity applied across several unobserved ages. It is important to keep the starting age consistent: a tail from age 10 cannot be applied to a cumulative amount already measured at age 12.

Worked Example: Apply Projected Factors to Cumulative Loss

At age 12, a hypothetical liability portfolio has cumulative paid loss of $4,800,000. Its selected late-age link ratios are 1.050, 1.040, and 1.030 across successive ages 9 to 10, 10 to 11, and 11 to 12. Assume this steady decline of 0.010 per age continues for ages 12 to 14, and that development beyond age 14 is negligible. Estimate the age-14 cumulative paid loss.

Project each factor. The factor from age 12 to 13 is $1.030-0.010=1.020$. The factor from age 13 to 14 is $1.020-0.010=1.010$.

Multiply the factors. The tail factor from age 12 to 14 is $1.020\times1.010=1.0302$.

Apply the factor. The age-14 estimate is $$4{,}800{,}000\times1.0302=$4{,}944{,}960$. The estimated increase is $$4{,}944{,}960-$4{,}800{,}000=$144{,}960$.

Check. The estimate is above the age-12 amount by 3.02%, consistent with a tail factor of 1.0302. This is a projection to the stated endpoint under the stated trend; it is not, by itself, evidence that development after age 14 is immaterial.

Common Mistakes and Limits of the Method

A projected tail depends on both the pattern and the range over which it is extended. Before using it, identify the observed ages, the projected ages, the numerical trend being continued, and the age at which the projection stops. If recent ratios reflect a change in claim handling, case reserving, settlement practice, or data quality, a simple extension may reproduce a temporary distortion rather than a persistent development pattern.

A frequent calculation error is to add the projected excesses above 1 instead of multiplying the factors. In the previous example, the incorrect calculation would be $1+(1.020-1)+(1.010-1)=1.030$. The correct calculation is $1.020\times1.010=1.0302$. The additive result misses the interaction between successive development amounts. Although the difference is small for these factors, it grows as the number or size of projected factors increases.

Do not extend a trend without checking its direction. A falling ratio may eventually imply a forecast at or below 1. Consider whether that is plausible for the loss measure, and whether the chosen projection horizon is appropriate.
Do not confuse a projected link ratio with the full tail. The tail is the product of every factor from the latest observed age to the selected endpoint.
Do not mistake a smooth extrapolation for certainty. Late-age data are often sparse, so the selected tail can be sensitive to the final observed factors, the extrapolation method, and the endpoint.
Do not round early. Retain adequate precision in forecast factors and their product; round the final reported factor or dollar amount as appropriate.

Key takeaway: Extrapolation extends an observed late-age pattern under an explicit assumption; multiplication converts the projected link ratios into a tail factor. State the starting age and endpoint, inspect the forecasts for plausibility, and treat the result as a selection rather than a mechanical certainty.

Check Your Understanding

For each hypothetical situation, identify the extrapolated factors or tail requested and check that the calculation uses the stated ages and method.

Selected ratios from ages 6 to 7, 7 to 8, and 8 to 9 are 1.060, 1.045, and 1.030. Continue the constant decline for ages 9 to 11 and calculate the tail from age 9 to 11. (Answer: projected factors 1.015 and 1.000; tail factor 1.015)
Three consecutive late-age ratios are 1.040, 1.030, and 1.020. Assume the same decrease continues for the next three ages. Calculate the tail across those three projected ages. (Answer: projected factors 1.010, 1.000, and 0.990; tail factor 0.9999)
A fitted linear model for excess above 1 is $\widehat y_x=0.080-0.012x$. Find the projected link ratios at $x=3$ and $x=4$, then calculate their product. (Answer: 1.044 and 1.032; tail factor 1.077408)
At the latest observed age, cumulative loss is $2,500,000. The two selected unobserved factors are 1.025 and 1.015, and no further development is assumed after the second factor. Calculate the projected amount at the endpoint. (Answer: tail factor 1.040375; projected amount $2,600,937.50)
A candidate projects two factors of 1.030 and 1.020 and reports the tail as 1.050 by adding their excesses to 1. What is the correct tail, and what is the error? (Answer: $1.030\times1.020=1.0506$; successive link ratios must be multiplied, not added.)

What You'll Learn