Simone De Bonis
Simone De Bonis
← Back to Blog
StatisticsJune 12, 2023· 3 min read

Outliers, Extremes and Black Swans

Distinguishing measurement anomalies from genuine extreme realizations and structural tail events in quantitative modeling.

#Statistics#Risk Modeling#Theory#Data Science

In applied data science and quantitative modeling, observations that fall far outside the bulk of a dataset are often grouped together under the generic label of “outliers”. However, treating every extreme observation as noise to be trimmed or winsorized can destroy the most important signal in the data.

Three Distinct Phenomena

1. Outliers (Anomalies / Contamination)

An outlier is an observation generated by a different data-generating process than the phenomenon under study—such as a sensor malfunction in telemetry data, a corrupted server log entry, or a data-entry error.

When an observation is a true measurement anomaly, robust preprocessing or anomaly detection is essential so that parameter estimates are not distorted.

2. Extremes (Tail Realizations)

An extreme value belongs to the same data-generating process, but sits in the far tail of the underlying probability distribution. Under Extreme Value Theory (EVT), the distribution of block maxima Mn=max⁡(X1,…,Xn)M_n = \max(X_1, \dots, X_n) converges to the Generalized Extreme Value (GEV) distribution:

G(x;ξ)=exp⁡(−(1+ξx)−1/ξ)G(x; \xi) = \exp\left( -\left(1 + \xi x\right)^{-1/\xi} \right)

where the shape parameter ξ\xi governs the tail behavior:

  • ξ=0\xi = 0: Gumbel domain (light-tailed, e.g., Gaussian or Exponential)
  • ξ>0\xi > 0: Fréchet domain (heavy/fat-tailed, e.g., Pareto, Student-tt)
  • ξ<0\xi < 0: Weibull domain (bounded upper tail)

3. Black Swans

Coined and formalized by Nassim Nicholas Taleb, a Black Swan is a high-impact, rare tail event in the “Mediocristan vs. Extremistan” framework where single realizations dominate cumulative aggregates and retrospective probability models suffer from severe epistemic uncertainty.

Practical Rule for Modeling

Before removing an extreme point during data preprocessing, always ask:

  • Is this point physically impossible or caused by instrumentation noise (Outlier)?
  • Or is the system genuinely governed by a heavy-tailed distribution where rare realizations carry the majority of the risk (Extreme / Black Swan)?
← Back to Blog
Share article: