The Analyst's Path

Arithmetic vs geometric mean

The arithmetic mean is the simple average of a set of returns and, for independent periods, the best single estimate of next period's return. The geometric mean is the constant rate at which money actually grew over the whole series. They differ by roughly half the variance: a series alternating between a 40% gain and a 30% loss averages 5.00% arithmetically, yet the money shrinks at about 1.005% a period.

3 min readPart of Returns, Statistical Measures and the Discipline of InferenceOpen in course mode

Advertisement

Four averages appear in return work and they are not interchangeable. Each answers a different question, and picking the wrong one is the commonest quantitative error in published fund material.

How do the arithmetic and geometric means differ?

The arithmetic mean, Σr/n, answers: what is my best single estimate of next period's return, if periods are independent draws from one distribution? It is the right input to a forward-looking expected-return model and the wrong description of a realised path.

The geometric mean, (Π(1+rᵢ))^(1/n) − 1, answers: at what constant rate did the money actually grow? On the +40, −30, +40, −30 series the terminal wealth relative is 1.40 × 0.70 × 1.40 × 0.70 = 0.9604, so the geometric mean is 0.9604^0.25 − 1 = −1.005051%. Ten lakh rupees ends at ₹9,60,400. Report the arithmetic 5.00% as though it described the path and you claim ₹12,15,506, which is ₹2,55,106 of money that never existed.

The gap between the two is roughly half the variance. The population variance of the four returns is 0.1225, so 5.00% − 12.25%/2 = −1.125% against a true −1.005051%, good to 12 basis points on a series with a 35% standard deviation. Volatility is a tax on compounding, and this is the receipt.

Advertisement

When do you use a weighted or harmonic mean?

The weighted mean, Σwᵢrᵢ, answers: what did the portfolio return, given how much sat in each part? Sixty-two percent of a book in a sleeve returning +11.80% and 38% in a sleeve returning −2.40% gives 0.62(11.80) + 0.38(−2.40) = 6.404%. The unweighted average of the two sleeve returns is 4.70%, wrong by 1.704 points, and wrong in a direction that depends entirely on which sleeve was bigger.

The harmonic mean, n / Σ(1/xᵢ), answers: what did I pay on average, when I spent a fixed sum each time rather than buying a fixed quantity? Worked example 1 shows it reproducing a purchase plan's realised average cost to the paisa, while the arithmetic mean of the same prices overstates that cost by 6.3125%. The other place it is the honest average is a set of valuation multiples, and the relative-valuation module owns that case.

For a set of positive numbers the three sit in a fixed order: harmonic ≤ geometric ≤ arithmetic. Equality holds only when every number is identical. On the purchase plan above: ₹253.97, ₹261.53, ₹270.00. The spread between them is a measure of dispersion in disguise.

Advertisement

Part of the course

This page is part of Returns, Statistical Measures and the Discipline of Inference (QM1.04), in Phase 12: Finance Plus, AI and the quant-code track. The module's flashcards and its quiz are in the app.