Fat Tails: The Black Swan Preview

"A tail model determines how quickly extreme-event probabilities shrink. Calling an event rare under a normal model does not make it rare under the real data-generating process."

What "Heavy-Tailed" Can Mean

Tail terminology varies. In rigorous probability, a heavy-tailed distribution is often one whose moment-generating function is infinite for every positive argument. In applied work, "fat-tailed" is also used more loosely for data with more extreme observations than a fitted normal predicts. State the definition and comparison model.

High kurtosis can signal tail weight or a sharp center, but kurtosis greater than 3 is not a complete definition. Some heavy-tailed distributions, such as the Cauchy, do not have finite mean or variance, so comparison in σ units is not even defined theoretically.

Normal Tails

After standardization, the normal density decays like exp(−x²/2). Its two-sided probability beyond |z|=5 is about 5.7×10⁻⁷, and beyond |z|=6 about 2.0×10⁻⁹. These are extremely small model probabilities, not impossibilities. Multiple testing, dependence, nonstationarity, and model error can all change how often such observations appear.

Student-t and Pareto-family models can place much more mass in the tails. But if variance is infinite, phrases such as "5σ event" lose their ordinary population meaning. If variance is finite, σ can still summarize spread while normal tail translation fails.

The 2008 Lesson Without a Single-Cause Myth

Some financial risk systems underestimated tail dependence, volatility changes, liquidity feedback, leverage, and model uncertainty. Normal or near-normal return assumptions contributed in some settings, but the crisis was not caused by one bell-curve formula. Incentives, underwriting, securitization, funding fragility, crowded positions, and policy failures interacted.

The practical lesson travels: validate tails and dependence, stress regimes outside the calibration sample, and treat model uncertainty as part of risk. Do not diagnose every failure as "fat tails" without checking the mechanism.

Operating Rule

Separate a spread statistic from a tail probability model. Ask how σ was estimated, whether the process is stable, what dependence exists, and how empirical extremes compare with candidate distributions. Quantiles, robust summaries, explicit tail models, and stress scenarios may complement or replace a normal approximation.

What Track 02 Has Done

This track introduced distributions, normalization as a bounded metaphor, normal and skewed shapes, candidate power laws, and tail behavior. Track 03 develops conditions for normal approximations; Track 07 examines failures without reducing them to a single cause.

Code

Normal vs Student-t tails: same sigma, different reality·python

import numpy as np
rng = np.random.default_rng(20)
N = 200_000

# Compare a normal and a Student's t with low degrees of freedom (fat-tailed).
normal = rng.normal(size=N)
student_t = rng.standard_t(df=3, size=N)

# Standardize both so their std=1 for fair tail comparison.
for name, x in [("normal", normal), ("t (df=3)", student_t)]:
    x = (x - x.mean()) / x.std()
    p4 = (np.abs(x) > 4).mean()
    p5 = (np.abs(x) > 5).mean()
    p6 = (np.abs(x) > 6).mean()
    print(f"{name:>10s}: P(|x|>4) = {p4:.5f}   P(|x|>5) = {p5:.6f}   P(|x|>6) = {p6:.6f}")

# In a normal, |x|>5 is about 6 in 10 million.
# In a t(df=3), |x|>5 happens roughly 1000x more often.
# Same axis, same standard deviation, different reality.
# This is what 'fat-tailed' looks like on the page.

Exercise

Choose a dataset that may have heavy tails. Plot empirical quantiles or the complementary CDF, compare at least two candidate models, and test sensitivity to the time window and threshold. Report whether σ is finite and stable enough to be useful.

Hint

A maximum ten or one hundred times the median is not proof of a heavy tail. Sample size, units, mixtures, and domain constraints can produce large ratios.

What "Heavy-Tailed" Can Mean

Normal Tails

The 2008 Lesson Without a Single-Cause Myth

Operating Rule

What Track 02 Has Done

Code

External links

Exercise

Progress

Comments 0