Geometric mean:
the average that respects compounding

When the ordinary average gives the wrong answer

Suppose an investment gains 50% in one year and loses 50% the next. The simple average of +50% and −50% is zero, which suggests you broke even. In fact 100 becomes 150 and then 75, a loss of 25%. The error comes from averaging quantities that multiply, not add.

Whenever outcomes compound, such as returns, growth factors, ratios of prices or fold-changes in a lab assay, the typical value must be defined by the end result: the steady factor that, repeated each period, gives the same overall product. That steady factor is the geometric mean.

The formula

The nth root undoes multiplication in the way that dividing by n undoes addition. For 2 and 8 it is √16 = 4, and 4 × 4 = 16 = 2 × 8: replacing both values by 4 leaves the product unchanged, just as replacing values by their arithmetic mean leaves the sum unchanged.

Logs make large sets manageable. Take the natural log of each value, average those, and exponentiate: GM = exp(mean of ln x). This avoids overflow when you multiply many numbers and is how software computes it.

Worked example: three years of growth factors

A portfolio changes by +10%, −5% and +20% over three consecutive years. As growth factors those are 1.10, 0.95 and 1.20.

Check it: 1.0784³ is 1.254, so ₹100 grows to ₹125.40 in three years at 7.84% a year, which matches the actual path. Using the arithmetic 8.33% would predict 100 × 1.0833³ = 127.1, overstating the final value by about ₹1.70. The gap grows with the volatility of the returns.

Geometric versus arithmetic: which is bigger and why

For positive numbers that are not all equal, it is always smaller than the arithmetic mean. They coincide only when every value is the same. The more the values spread out, the wider the gap.

The last row shows how far apart they can get. For data that span orders of magnitude, such as bacterial counts or incomes, it is often the more representative centre, since a handful of huge values do not dominate it.

This ordering is known as the AM–GM inequality, and it has a pleasant intuition. For a fixed sum, a rectangle has the largest area when it is a square; for a fixed arithmetic mean, the product of the values is largest when they are all equal. Any spread at all reduces the product, and therefore the root, below the arithmetic average. Volatility drag on investment returns is the same effect in financial clothing.

The link with compound annual growth rate

Finance reports often state a compound annual growth rate, and it is just this average in disguise. If a value moves from a start amount to an end amount in a set number of years, the rate is the root of the end-to-start ratio, minus one.

The earlier portfolio went from 100 to 125.4 in three years, so (125.4 ÷ 100)^(1/3) − 1 = 0.0784, matching the root of the yearly factors. The only inputs needed are the two end points; the path in between does not matter, which is also its weakness, because a smooth 7.8% and a wild ride can share the same CAGR.

Treat such numbers as descriptions of the past. Reported returns depend on fees, taxes and the dates chosen, and they say nothing certain about what comes next.

Second example: counts spread over orders of magnitude

Four plates in a microbiology test give counts of 120, 450, 1,800 and 7,000 colonies per millilitre. The arithmetic mean is 2,342.5, pulled upward by the one large value. The alternative uses the fourth root of the product.

That value sits much closer to the middle of the data on a logarithmic scale, which is how counts of organisms are normally compared, and it is why laboratories report this kind of average for such data.

Where you will see it used

• Investment performance, where it gives the compound annual growth rate over several years.
• Index construction, for example inflation baskets that average price ratios.
• Biology and environmental science, where concentrations or counts are log-normally distributed.
• Photography and screen sizes, where successive steps are ratios of the same size.

A frequent mistake is to feed in zero or negative values. A single zero collapses the whole product to zero, and a negative value makes the root undefined or misleading, so the method is reserved for strictly positive data. If returns include a total loss, the compounded growth factor is simply zero, which is itself informative.

Checking the answer

Two quick tests protect against slips. First, the result must lie between the smallest and largest value. Second, raise it to the power n; you should recover the product you started with. A calculator that has an arithmetic mean function can supply the comparison figure, and a separate geometric-sequence tool deals with a constant ratio between consecutive terms, which is the special case where every period grows by the same factor.

In reports, state the number of periods and whether returns are before or after fees and taxes, which differ by product and country, because the same raw data can give different averages depending on what was netted off.