Geometric mean

The geometric mean is the nth root of the product of n values: GM = (x₁ × x₂ × ... × xₙ)^(1/n). Unlike the arithmetic mean, the geometric mean is the right tool for averaging ratios, percentages, and compounding growth rates.

Why it matters: imagine an investment that returns +50% in year 1 and −50% in year 2. The arithmetic mean of [1.5, 0.5] is 1.0 — suggesting the investment broke even. But $100 invested becomes $150 after year 1 and $75 after year 2 — a 25% loss. The geometric mean of [1.5, 0.5] is √(1.5 × 0.5) = √0.75 ≈ 0.866, correctly indicating a 13.4% annualised loss.

Standard use cases:

• Compound annual growth rate (CAGR). The geometric mean of the year-over-year ratios.
• Index numbers. Stock indices weighted by ratios use geometric means to avoid arithmetic-mean distortion.
• Ratio-of-ratios comparisons. When the inputs are multiplicative rather than additive (gains compounding, dosing).

The geometric mean is always ≤ the arithmetic mean. Equality holds only when all values are identical. For widely varying values, the gap can be large — which is precisely why using the right one matters.

Worked example

You invest $10,000 in a fund. Returns: year 1 +30%, year 2 −20%, year 3 +25%, year 4 −10%, year 5 +15%. Arithmetic mean of [30, −20, 25, −10, 15] = 8% — what the marketing brochure would quote. Now compound the actual money: 10,000 × 1.30 × 0.80 × 1.25 × 0.90 × 1.15 = 10,000 × 1.3455 = $13,455. Total gain 34.55% over five years. The geometric mean: (1.30 × 0.80 × 1.25 × 0.90 × 1.15)^(1/5) = 1.3455^0.2 ≈ 1.0612 — a 6.12% annualised return. That’s the “CAGR” the SEC requires US mutual funds to report instead of the arithmetic mean, because reporting 8% would imply 1.08⁵ × 10,000 = $14,693 — overstating actual performance by 9%. The mismatch grows with return volatility: two assets with the same arithmetic-mean return can have very different CAGR if one is more volatile.

One more illustrative case: an algorithmic-trading strategy backtest reports returns of [+40%, +30%, −50%, +20%, +10%] across five years. Arithmetic mean: 10% per year. Geometric mean: (1.40 × 1.30 × 0.50 × 1.20 × 1.10)^(1/5) ≈ 1.2012^0.2 ≈ 1.037 — only 3.7% annualised. The −50% year dominates everything; a single deep drawdown crushes the long-run return regardless of subsequent gains. This is why hedge-fund disclosures emphasise “maximum drawdown” alongside annualised returns.

When and why it matters

Geometric mean matters in any context where the numbers compound: investment returns, biological growth rates, viral spreading (R₀), and depreciation. The arithmetic-mean error is largest when individual values vary widely — symmetric ±50% returns produce a net loss because gains and losses compound asymmetrically (a 50% loss requires a 100% gain to recover). The same lesson applies to weight loss percentages, bacterial colony counts, and any “average rate of change” reported across multiple time periods. The opposite mistake — using geometric mean for additive quantities like daily temperatures or test scores — produces meaningless numbers. The rule: if the values multiply to produce the final result, use geometric mean; if they add, use arithmetic. For mixed scenarios (Sharpe ratios, weighted averages), specialised formulas exist. Reference: Investopedia — Compound Annual Growth Rate.

The numerical-stability trick — work in log space: computing the product of many values directly overflows or underflows IEEE 754 doubles quickly (10 values of 0.1 each produces 10⁻¹⁰, fine; 50 values produces 10⁻⁵⁰, fine; 500 values produces 0.0, definitely not fine). The standard fix is to compute exp(mean(log(x))) instead. Logarithms turn the product into a sum, the sum stays within representable range, and the final exp recovers the geometric mean. Every statistical library (NumPy, SciPy, R) implements geometric mean this way internally; rolling your own naïvely is one of the more common ways to silently produce zeros for large samples.

Why the geometric mean works for ratios — the intuition: ratios live on a logarithmic scale, not a linear one. “Doubled” (×2) and “halved” (×0.5) are symmetric multiplicative changes — they should average to “no change” (×1), not to “1.25×” as the arithmetic mean would suggest. The geometric mean respects this symmetry because it computes the average logarithm of the ratios. This is why finance, biology (cell-growth rates), and physics (radioactive decay) all default to geometric means for time-series ratios. Related: harmonic mean, arithmetic mean, compound interest. Reference: NIST/SEMATECH e-Handbook — Geometric Mean.