Fat tails and the tail index

M3.1 established that the tails are enormous. This lesson asks the quantitative question — how heavy, measured how — and arrives at a result that undermines one of the numbers M3.1 leaned on.

Two ways for a tail to decay

The distinction is between exponential and polynomial decay:

   Gaussian      P(|r| > x)  ~  exp(−x²/2σ²)     dies faster than
                                                  anything
   exponential   P(|r| > x)  ~  exp(−λx)         "thin" tail

   POWER LAW     P(|r| > x)  ~  x^(−α)           "fat" tail
                                                  α is the TAIL INDEX

The difference is not one of degree. Under a Gaussian each extra sigma costs you exponentially in probability, which is why 6σ is effectively unreachable. Under a power law each doubling of the move costs a fixed factor 2^(−α), so extremes remain reachable no matter how far out you go.

Small α means fat. As α falls, the tail thickens.

Measuring it, two ways

By moments — the Student-t. The t-distribution with ν degrees of freedom has excess kurtosis 6/(ν−4) for ν > 4, so invert it:

   observed excess kurtosis  =  16.71

        ν  =  6/16.71 + 4  ≈  4.36 degrees of freedom

Which is already interesting: ν ≈ 4.4 is at the very edge of where kurtosis is defined at all.

By the tail directly — the Hill estimator. Rather than fitting the whole distribution, use only the k largest observations:

                    k
   α̂  =  ─────────────────────────────       x₍ᵢ₎ = the i-th largest loss
          Σᵢ₌₁ᵏ  ln( x₍ᵢ₎ / x₍ₖ₊₁₎ )

   S&P 500 losses, 2016–2026 (1,136 down days):

        top  25 losses   →   α ≈ 3.07
        top  50 losses   →   α ≈ 2.90
        top 100 losses   →   α ≈ 2.35
        top 150 losses   →   α ≈ 2.34

Note the estimate drifts as k grows. That is the Hill estimator's well-known bias-variance trade-off: too few points and the estimate is noisy, too many and you contaminate the tail with the body of the distribution. The honest reading of that column is α is somewhere around 2.5 to 3, not any single decimal.

The inverse cubic law

That range is not specific to this decade or this index. Across equities, indices, FX, and commodities, over horizons from minutes to days, measured on data going back a century, the tail index of returns clusters around

                          α  ≈  3

This is the inverse cubic law, and it is one of the closest things finance has to an empirical constant. Its universality is exactly why the econophysics literature got interested — this is the behaviour of a system with scale-invariant fluctuations, and it looks far more like critical phenomena in statistical physics than like anything in a finance textbook. The roadmap's claim that a physics background pays off most in M3 is really a claim about this fact.

The consequence that bites

Here is why the tail index is worth measuring rather than admiring. For a power-law tail with index α, moments of order ≥ α do not exist.

   α ≈ 3  ⟹    mean       (1st moment)   exists          ✓
               variance   (2nd moment)   exists          ✓
               skewness   (3rd moment)   marginal        ~
               kurtosis   (4th moment)   DOES NOT EXIST  ✗

Read that against M3.1, which reported a kurtosis of 19.71. If the fourth moment of the underlying distribution is infinite, then 19.71 is not an estimate of anything. It is a number computed from a finite sample of a population whose kurtosis does not converge — and it will keep growing, erratically, as you add data, jumping every time a new record day arrives.

This is not a technicality. It is the reason:

• Sample moments are unstable. M3.7 shows that removing ten days from 2,513 moves the Sharpe ratio from 0.70 to 0.37 or 1.17 depending which ten. With near-infinite higher moments, a handful of observations dominate.
• "σ" is a shakier object than it looks. Variance exists, but its estimate has a fat-tailed sampling distribution, so volatility estimates are themselves volatile.
• Any risk number built from a fitted normal is not conservative, it is wrong — and wrong in the direction that loses money (M3.4).

So what should you use?

Three honest options, in ascending order of correctness and effort:

Student-t instead of normal. Cheap, better, and captures much of the tail behaviour. ν ≈ 4 is a reasonable default for daily equity returns. Most risk systems that do anything at all do this.

Fit the tail separately — extreme value theory. The insight of EVT is that you should not fit a distribution to all the data and read off the tail; you should model only the tail. The Pickands–Balkema–de Haan theorem says exceedances over a high threshold converge to a generalised Pareto distribution, regardless of the parent distribution — a CLT for extremes. This is the right tool for "how bad is the 1-in-1000 day", and it does not require pretending the middle of the distribution is informative about the edges.

Accept irreducible uncertainty. With α ≈ 3 and a decade of data you have maybe a dozen genuinely extreme observations. You cannot estimate a 1-in-10,000-day event from 2,513 days no matter how good your method. The correct response is structural — position sizing that survives being wrong (M3.6) — not a better point estimate.

Where the tails come from

Two mechanisms, and they are complementary rather than competing.

Volatility clustering generates fat tails mechanically. A mixture of normal distributions with different variances is itself fat-tailed, even though every component is Gaussian. Since volatility clusters and varies enormously (M3.3), the unconditional distribution of returns is a mixture, and inherits heavy tails automatically. A good chunk of the observed kurtosis is this.

But not all of it. Fit a GARCH model, standardise the returns by its conditional volatility, and the residuals are thinner-tailed than the raw returns — but still not Gaussian. Something genuinely jumpy remains: news arrives discontinuously, liquidity withdraws (M1.2, as μ → 1), and crowded positions unwind at once. That residual is real, and it is what M0.7's leverage and M3.6's position sizing have to survive.

Source: Rama Cont (2001) again for the tail-index evidence. For the power-law literature, Gabaix, Gopikrishnan, Plerou & Stanley, “A theory of power-law distributions in financial market fluctuations” (Nature, 2003). For EVT done properly, McNeil, Frey & Embrechts, Quantitative Risk Management, ch.7 — the standard reference and worth owning if you end up anywhere near risk.