Library / Econophysics

The Inverse Cubic Law of Return Tails

by , founder of Stochastly

Do extreme market returns follow the same tail pattern across assets and horizons?

Gopikrishnan and coauthors found an approximate tail exponent near three over specified equity-index samples and horizons. Estimate your own tail threshold and uncertainty before using that model; another asset or longer horizon may have a different tail shape.

The observed tail exponent near three is bounded to the authors? equity-index samples and horizons.

Return-tail survival is plotted on logarithmic axes with a fitted upper-tail slope.
Return-tail survival is plotted on logarithmic axes with a fitted upper-tail slope.

Evidence map

AspectFinding
What it isThe empirical finding that the tail of the distribution of returns, normalised by volatility, decays as a power law P(|r| > x) ~ x^(−α) with α close to 3.
Key result / formulaGopikrishnan, Plerou, Amaral, Meyer & Stanley (1999) measured it on the S&P 500 with three databases — about one million one-minute records for 1984-1996, daily records for 1962-1996 and monthly records for 1926-1996.
Why it matters for backtestingThe measured tail exponent near three is a warning against fitting a Gaussian risk model without checking its tails.

What it is

That is outside the Lévy-stable range α < 2, which the same measurement rules out: variance is finite, while the fourth moment sits at the edge of existence.

Key result / formula

The exponent α ≈ 3 holds for both tails from one minute up to horizons of a few days; at longer horizons the distribution slowly crosses over to a Gaussian. Gabaix, Gopikrishnan, Plerou & Stanley (2003) proposed a mechanism: large institutions whose sizes follow Zipf's law trade in markets where price impact is concave (square-root in volume), which turns the exponent 1 of fund sizes into an exponent 3 for returns and 3/2 for volumes — a derivation that predicts the two exponents jointly from one mechanism.

Why it matters for backtesting

The reported range belongs to the authors' equity-index samples and horizons; it is not a universal null for every market or period. At a specified confidence level, compare a Gaussian tail estimate with an empirical or suitably fitted heavy-tail estimate on the same data, then check exceedances out of sample. Whether the Gaussian estimate understates risk depends on the fitted distribution, horizon and quantile. A bootstrap or simulated null should preserve relevant dependence and tail behavior, but forcing an x^-3 law on a different asset could be just as misleading. Tail fitting also requires an explicit threshold and uncertainty interval; an eyeballed log-log slope does not establish an exponent. These constraints matter especially when the few most extreme observations dominate a risk estimate.

Source

Gopikrishnan, Plerou, Amaral, Meyer & Stanley, "Scaling of the distribution of fluctuations of financial market indices", Physical Review E 60(5), 1999, 5305-5316; Gabaix, Gopikrishnan, Plerou & Stanley, "A theory of power-law distributions in financial market fluctuations", Nature 423, 2003, 267-270. Primary source

Related in the library