Geometric Averages of Link Ratios

Why Geometric Averaging Fits Multiplicative Development

A link ratio describes the multiplicative change from one development age to the next. When several positive link ratios are available for the same age-to-age interval, a geometric average summarizes their typical multiplicative level. It is not the same calculation as taking an arithmetic mean: the geometric mean is designed around products and proportional changes.

This distinction matters because development factors are used as multipliers. A selected factor of 1.10, for example, means that the projected cumulative value is 1.10 times its current value. A geometric average is useful when the intended average is multiplicative across observations. It does not, by itself, establish that all observations are equally suitable for selection; the observations still need to be relevant to the development age being considered.

Definition: Given $n$ strictly positive link ratios $r_1,r_2,\ldots,r_n$, their equal-weight geometric mean is $$ G=\left(\prod_{j=1}^{n}r_j\right)^{1/n}. $$ Equivalently, $G$ is the positive number whose $n$th power equals the product of the ratios. The ratios must be strictly positive for the logarithmic calculation and ordinary positive geometric mean used here.

A link ratio is a ratio of cumulative values at adjacent development ages. For example, if a hypothetical accident year has cumulative loss of $100,000 at one age and $110,000 at the next, its link ratio is $110{,}000/100{,}000=1.10$. The factor is dimensionless: it does not depend on whether the cumulative values are expressed in dollars or thousands of dollars.

Why the Geometric Mean Formula Holds

The formula follows from the multiplicative role of link ratios. If each of the $n$ observed ratios were replaced by one common factor $G$, preserving their product requires the product of the replacements, $G^n$, to equal the product of the original ratios. Solving that equation for the positive $G$ gives the geometric mean.

G^n=\prod_{j=1}^{n}r_j \qquad\Longrightarrow\qquad G=\left(\prod_{j=1}^{n}r_j\right)^{1/n}.

Taking logarithms gives another derivation and a practical calculation method. Since every $r_j$ is positive, its logarithm is defined. Taking the logarithm of the product equation converts multiplication into addition:

n\log G=\sum_{j=1}^{n}\log r_j \qquad\Longrightarrow\qquad \log G=\frac{1}{n}\sum_{j=1}^{n}\log r_j.

Thus, the logarithm of the geometric mean is the arithmetic mean of the logarithms. This can be convenient when there are many ratios or when multiplying them directly risks numerical error. The final factor is obtained by exponentiating the average log.

Confirm the observations.
Use strictly positive link ratios for the same development-age transition, with the intended observations included once each.

Multiply or take logarithms.
Calculate the product of the ratios, or calculate each logarithm and average those logarithms.

Take the $n$th root or exponentiate.
Use $n$, the number of ratios, as the root. Round the selected factor only at the final step.

Apply the factor consistently.
When projecting a cumulative value one development age forward, multiply that value by the selected link ratio.

Worked Calculations

Worked Example: Averaging Two Positive Link Ratios

A hypothetical liability portfolio has two positive observed link ratios for a particular age-to-age interval: 1.00 and 1.10. An accident year has cumulative loss of $240,000 at the earlier age. Calculate the equal-weight geometric average and use it to project the cumulative loss one age forward. Round the projected dollar amount to the nearest cent.

Method. Use the geometric mean because the requested average is multiplicative. There are $n=2$ ratios, and their product is

1.00\times1.10=1.10.

The geometric average is the square root of this product:

G=\sqrt{1.10}=1.048808848\ldots

Apply the unrounded factor to the cumulative loss:

240{,}000\times1.048808848\ldots =251{,}714.1235608\ldots

Rounding the final amount to the nearest cent gives $251,714.12. As a check, the selected factor is greater than 1, so the projected cumulative amount is greater than $240,000. It is also less than $1.10, so the projection is less than $264,000, which is $240{,}000\times1.10$.

Common mistake avoided. The arithmetic mean is $(1.00+1.10)/2=1.05$, which would project $252,000. That is not the requested geometric average. The geometric calculation is $\sqrt{1.10}$; the amount above uses the unrounded square root, not an intermediate rounded factor.

Worked Example: Geometric Mean From Cumulative Values

For a hypothetical line of coverage, three accident years have the following cumulative losses, in thousands of dollars, at two adjacent development ages: 100 and 90; 100 and 100; and 90 and 100. Calculate each link ratio, their geometric mean, and the projection from a separate accident year's current cumulative loss of $270,000.

Step 1: Calculate the ratios. For each accident year, divide the later cumulative value by the earlier value:

r_1=\frac{90}{100}=0.90,\qquad r_2=\frac{100}{100}=1.00,\qquad r_3=\frac{100}{90}=\frac{10}{9}=1.111111\ldots

Step 2: Multiply and take the cube root. The product is

0.90\times1.00\times\frac{10}{9}=1.00.

There are three ratios, so

G=(1.00)^{1/3}=1.00.

Step 3: Project the separate accident year. The current cumulative loss is $270,000, so the projected amount is

270{,}000\times1.00=\$270{,}000.

The calculation checks because the product of the three ratios is exactly 1, so their geometric mean must also be 1. The ratios are all positive, including the ratio below 1; a positive link ratio need not represent an increase.

Common mistake avoided. Replacing the geometric mean with the arithmetic mean gives $(0.90+1.00+1.111111\ldots)/3=1.0037037\ldots$, which would project approximately $271,000. That different result comes from averaging the ratios additively, rather than preserving their product.

Worked Example: Using the Logarithmic Calculation

A hypothetical property portfolio has three positive observations for the same development-age interval: 1.10, 1.21, and 1.331. Calculate the geometric average using powers of 1.10, then project a cumulative loss of $400,000. Round the selected factor to four decimal places and the projected loss to the nearest cent.

Step 1: Express the observations as powers.

1.10=1.10^1,\qquad 1.21=1.10^2,\qquad 1.331=1.10^3.

Step 2: Calculate their product and geometric mean.

1.10\times1.21\times1.331 =1.10^{1+2+3} =1.10^6 =1.771561, $$ $$ G=(1.771561)^{1/3}=1.10^2=1.21.

In logarithmic form, the same calculation averages the exponents: $(1+2+3)/3=2$, so the average log corresponds to $1.10^2$. The factor rounded to four decimal places is 1.2100.

Step 3: Project.

400{,}000\times1.21=\$484{,}000.00.

The check is direct: multiplying $400,000 by a factor of 1.21 produces a 21% increase, or $84,000, for a projected total of $484,000.00.

Common mistake avoided. Do not treat 1.21 as a 1.21% increase. It is a multiplier, equivalent to a 21% increase. Also, the third root is required because there are three observations, not because the largest exponent is 3.

Interpretation and Conditions

The geometric mean is a single factor that preserves the product of the observed positive ratios when each observation is replaced by that common factor. In a chain of development intervals, link ratios are themselves multiplied to obtain a cumulative development factor. This makes a multiplicative average conceptually consistent with using link ratios as projection multipliers.

For example, if ratios $r_1,\ldots,r_n$ are replaced by $G$, then the product is unchanged:

\underbrace{G\times G\times\cdots\times G}_{n\text{ factors}} =G^n =\prod_{j=1}^{n}r_j.

This product-preservation property is the central justification for geometric averaging. It does not mean the geometric mean is automatically the appropriate selection in every reserving analysis. The observations should be for a comparable development-age interval, and the chosen averaging approach should match the purpose of the selection. The geometric mean gives equal multiplicative weight to each listed observation; it does not adjust for differences in the amounts underlying those observations.

Important condition: The ordinary geometric mean of positive link ratios is defined only when every ratio is strictly greater than zero. If an observation is zero or negative, its logarithm is not defined as a real number, so the logarithmic formula cannot be used. Do not silently delete or alter such observations to force a geometric mean; investigate the data and use a method appropriate to the situation.

The geometric mean is also not a substitute for checking the data. A ratio can be positive and still be unsuitable because of a data issue, a structural change, or an unusual observation. Earlier material on loss development triangles covers data reconciliation and completeness; those checks remain important before calculating any summary. This tutorial’s formula assumes the stated positive ratios have already been identified as the observations to average.

Common Exam Traps

Using the wrong count in the root. For $n$ observations, take the $n$th root of their product. Do not use the number of development ages in the entire triangle if only a subset of ratios is being averaged.
Averaging factors as percentages. A link ratio of 1.08 is a multiplier of 1.08, or an 8% increase. Calculate with 1.08, not 8.
Rounding too early. Retain the full product or logarithmic result and the unrounded factor until the final answer. Small rounding differences can affect a projected dollar amount.
Ignoring positivity. The logarithmic geometric-mean calculation requires strictly positive observations. Zero and negative values require investigation rather than an unannounced adjustment.
Assuming geometric averaging settles factor selection. It defines one way to summarize positive ratios. It does not establish that the selected observations are homogeneous or that every observation should receive equal weight for a particular actuarial purpose.

Key takeaway: For strictly positive link ratios $r_1,\ldots,r_n$, calculate $G=(\prod r_j)^{1/n}$. This factor preserves the product of the observations when each is replaced by the same multiplier. Keep the ratios unrounded through the calculation, apply the factor to cumulative loss as a multiplier, and do not confuse the geometric mean with an arithmetic average.

Check Your Understanding

For each problem, identify the number of ratios, calculate the geometric mean, and apply it where requested. Round factors to four decimal places when needed and monetary answers to the nearest cent.

A selected development-age interval has two positive link ratios, 1.00 and 1.21. Find their geometric mean. (Answer: 1.1000)
Three hypothetical accident years have positive link ratios 0.80, 1.00, and 1.25. Find the equal-weight geometric mean. (Answer: 1.0000)
Four positive observations are 1.05, 1.05, 1.1025, and 1.00. Find the geometric mean. (Answer: 1.0500)
For a hypothetical coverage, three accident years have cumulative values of 200 and 220, 250 and 275, and 300 and 330 (in thousands of dollars) at adjacent ages. Calculate the three link ratios, their geometric mean, and the projection from a separate current cumulative loss of $750,000. (Answer: Link ratios 1.10, 1.10, and 1.10; geometric mean 1.1000; projected loss $825,000.00)
Three positive link ratios are 1.20, 0.90, and 1.00. Calculate their geometric mean to four decimal places. (Answer: $\sqrt[3]{1.08}=1.0260$)

What You'll Learn