Why Use a Geometric Average?
In chain ladder estimation, an age-to-age link ratio describes how cumulative losses at one development age compare with cumulative losses at the next. At a particular age-to-age interval, several accident years may provide observed ratios. A selected factor summarizes those observations for use in the development pattern.
The geometric average is a way to combine positive link ratios by multiplying them and taking a root. This matches the multiplicative character of development: successive factors compound, and a factor represents a proportional change. It is not the same as taking the arithmetic average, which summarizes ratios by adding them.
For a cumulative loss triangle, write \(C_{i,j}\) for the cumulative loss for accident year \(i\) at development age \(j\). The observed link ratio from age \(j\) to age \(j+1\) for accident year \(i\) is
A geometric-average selection is calculated separately for each development interval. Ratios from different intervals describe different parts of the development pattern and should not be mixed into one average. This tutorial concerns the equal-weight geometric average of the observed ratios, not a dollar-weighted selection.
Calculation and Interpretation
The product-and-root formula is convenient for a small set of ratios. The logarithmic formula is often more convenient for many ratios or when the product is awkward to calculate. Both give the same result. Keep full precision in intermediate calculations and round the selected factor only at the end.
Identify the positive observed link ratios that correspond to the same pair of development ages.
Calculate the product of the ratios, or add their natural logarithms.
For \(n\) ratios, take the \(n\)th root of the product. Equivalently, divide the sum of logarithms by \(n\) and exponentiate.
Confirm that the input ratios were positive, came from one age-to-age interval, and were not rounded prematurely.
The result has a direct multiplicative interpretation. If every observed ratio equaled \(G\), their product would equal the product of the actual observations:
Thus, the geometric average preserves the product across the set of observed ratios. It also treats reciprocal changes symmetrically in logarithmic terms: a ratio of \(0.80\) has log \(\ln(0.80)\), while its reciprocal, \(1.25\), has log \(-\ln(0.80)\). Their average log is zero, corresponding to a geometric mean of 1.
Why the Formula Is a Natural Multiplicative Summary
One justification follows by considering proportional, rather than absolute, deviations. For positive ratios, take logarithms so that multiplicative comparisons become additive. The value \(g>0\) that minimizes the sum of squared log deviations from the observed ratios minimizes
Set \(z=\ln g\). Then \(Q\), as a function of \(z\), is a sum of squared differences from a constant. Differentiating with respect to \(z\) and setting the derivative equal to zero gives
Since \(z=\ln g\), exponentiating yields \(g=\exp\bigl(\frac{1}{n}\sum_k\ln L_k\bigr)\), the geometric-average formula. This justification does not establish that every data set should use a geometric average; it explains what the calculation summarizes. The selection method still needs to be appropriate to the data and purpose.
Worked Example: Selecting a Factor From Three Accident Years
In a hypothetical paid-loss triangle, the cumulative losses (in thousands of dollars) for three accident years move from development age 12 months to 24 months as follows: \(1{,}000\) to \(1{,}100\), \(2{,}000\) to \(2{,}420\), and \(3{,}000\) to \(3{,}000\). Calculate the equal-weight geometric-average link ratio for this interval, and compare it with the arithmetic average. Round final factors to four decimal places.
Method: geometric average of the link ratios. First calculate each accident year's ratio:
All three ratios are positive, so the geometric-average method applies. Their product is
There are three observations, so take the cube root:
The geometric selection is 1.1000 to four decimal places. For comparison, the arithmetic average is \((1.10+1.21+1.00)/3=3.31/3=1.103333\ldots\), or 1.1033 to four decimal places. The geometric selection preserves the product because \(1.10^3=1.331\). This method avoids the mistake of using an arithmetic average when the requested selection is a geometric average.
Worked Example: Reciprocal Development Movements
A hypothetical incurred-loss triangle produces three positive link ratios for one development interval: \(0.80\), \(1.00\), and \(1.25\). The first ratio reflects a downward revision in incurred losses; the other two are unchanged and upward, respectively. Calculate both averages and explain the result. Round final factors to four decimal places.
Method: product-and-root calculation. The product is
Taking the cube root gives the geometric average \(G=(1.00)^{1/3}=1.0000\). The arithmetic average is
The geometric result is 1 because \(0.80\) and \(1.25\) are reciprocals: \(0.80\times1.25=1\). Their log contributions cancel, and the third ratio is also 1. The arithmetic average instead gives more influence to the larger absolute ratio. This example avoids assuming that an arithmetic average and a multiplicative average must agree. A ratio below 1 is not, by itself, an obstacle to the geometric method; the condition is that the ratio is positive.
Worked Example: A More Dispersed Set of Positive Ratios
For a hypothetical development interval, four accident years have cumulative paid losses (in thousands of dollars) that move from \(1{,}000\) to \(1{,}020\), from \(1{,}000\) to \(1{,}080\), from \(1{,}000\) to \(1{,}150\), and from \(1{,}000\) to \(1{,}250\). Calculate the equal-weight geometric-average factor and the arithmetic average. Round final factors to four decimal places.
Method: ratios first, then geometric averaging. Dividing each later cumulative amount by its earlier amount gives \(1.02\), \(1.08\), \(1.15\), and \(1.25\). Multiply the four positive ratios:
There are four ratios. Using natural logarithms, \(\ln(1.58355)\approx0.459669\); dividing by four gives \(0.114917\). Exponentiating gives
As a check, \(1.1218^4\) is approximately \(1.5836\), consistent with the product \(1.58355\) given the rounded factor. The arithmetic average is
The arithmetic selection is greater in this example. That difference is not an error: the two methods summarize the same observations differently. This calculation avoids prematurely rounding the product or logarithm, which could change the final digits of the selected factor.
Geometric Versus Arithmetic Averaging
Both methods can be calculated from the same observed ratios, but they answer different summary questions. An arithmetic average is the ordinary mean of the ratio values. A geometric average is the ratio whose repeated product matches the observed product. The choice should follow the intended selection method rather than whichever answer appears more convenient.
| Feature | Geometric average | Arithmetic average |
|---|---|---|
| Calculation | Product of ratios, then the \(n\)th root | Sum of ratios, then divide by \(n\) |
| Interpretation | Preserves the product of the positive ratios | Preserves their sum |
| Convenient scale | Logarithms; proportional differences become additive | Original ratio scale; absolute differences are averaged |
| Input condition | Every ratio must be positive for the log formula | Can be calculated for any real-valued observations |
The geometric average is no greater than the arithmetic average for positive observations, with equality when all observations are equal. This inequality is useful as a check: if a calculation from positive ratios gives a geometric average above the arithmetic average, recheck the arithmetic. The two values may be close when the link ratios are similar, but they need not be close when the observations vary.
Common Mistakes and Practical Checks
- Taking the wrong average. Adding the ratios and dividing by their count calculates the arithmetic average. For a geometric selection, multiply and take the appropriate root, or average the logarithms and exponentiate.
- Using the wrong root. The root is determined by the number of observed ratios in the set. Three ratios require a cube root; four ratios require a fourth root.
- Mixing development intervals. First group ratios by the same pair of development ages. Do not combine, for example, 12-to-24-month ratios with 24-to-36-month ratios in one selection.
- Ignoring nonpositive observations. The log form requires positive ratios. A zero ratio has no finite logarithm, and a negative ratio has no real logarithm. Investigate such data rather than inserting them into the formula as if they were positive.
- Rounding too early. Keep the product or logarithms at full precision, then round the selected factor at the end.
A useful final check is to raise the selected geometric factor to the number of observations. The result should reproduce the product of the observed ratios, apart from any small difference caused by rounding. Also compare the selection with the arithmetic average as a reasonableness check, not as a substitute calculation.
Check Your Understanding
For each problem, use the stated rounding rule and check that the inputs belong to the same age-to-age interval.
- A hypothetical triangle gives ratios of \(1.04\) and \(1.21\) for one interval. Calculate the geometric and arithmetic averages to four decimal places. (Answer: geometric \(1.1218\); arithmetic \(1.1250\).)
- Three accident years have cumulative losses of \(800\) to \(880\), \(1{,}200\) to \(1{,}260\), and \(1{,}500\) to \(1{,}800\) over the same interval. Calculate each link ratio, then find the geometric and arithmetic averages to four decimal places. (Answer: ratios \(1.10,\ 1.05,\ 1.20\); geometric \(1.1150\); arithmetic \(1.1167\).)
- For one interval, the positive ratios are \(1.02,\ 1.08,\ 1.15,\ 1.25\). Calculate their product, geometric average, and arithmetic average. Round the averages to four decimal places. (Answer: product \(1.58355\); geometric \(1.1218\); arithmetic \(1.1250\).)
- A hypothetical incurred triangle gives positive ratios \(0.90,\ 1.00,\) and \(1.111111\ldots\) for one interval. Calculate the geometric average and the arithmetic average to four decimal places. Explain why they differ. (Answer: geometric \(1.0000\); arithmetic \(1.0037\); the first and third ratios are reciprocals, so their log contributions cancel.)
- For an interval, an analyst has ratios \(1.10,\ 0,\ 1.05\). Can the equal-weight geometric average be calculated using the positive-ratio log formula? Identify the condition that fails. (Answer: No; the zero ratio is not positive, and its logarithm is undefined.)