Statistical Consequences of Fat Tails: Real World Preasymptotics, Epistemology, and Applications

COAUTHORS 1

• Contributors listed include Yaneer Bar-Yam, Pasquale Cirillo, Raphael Douady, Andrea Fontanari, Hélyette and Donald Geman, Espen Haug, and collaborating teams.The section also credits the Uncertainty Reading Group and the Universa Investments team.Source: COAUTHORS 1 · p. 3

SKEPTICAL EMPIRICAL TRADITION (Philosophy)

• The section marks no overlap between heuristics and biases and between economics.It also lists philosophical traditions and figures spanning ancient skepticism, induction, Bayesian thought, and negative empiricism.Source: SKEPTICAL EMPIRICAL TRADITION… · p. 4

ASYMMETRY /FRAGILITY/

• The section describes almost no overlap between large-deviation theory and fat tails, alongside a genealogy of related research traditions.The listed researchers include Pareto, Lévy, Mandelbrot, Feller, Zolotarev, Taqqu, Samorodnisky, Gnedenko, and Resnick.Source: ASYMMETRY /FRAGILITY/ · pp. 4–5

Loss Models /INSURANCE (I)

• The section lists Cramer, Lundberg, Denbo, Zeitouni, and Varadhan among researchers associated with loss models and insurance..Source: Loss Models /INSURANCE (I) · p. 5

CONVERGENCE / STOCHASTIC

• The section lists convergence researchers and notes a stated lack of overlap between stochastic theory and fat-tail work.The chapter markers distinguish nontechnical, discussion, and adapted-from-published-paper chapters.Source: CONVERGENCE · p. 5, STOCHASTIC · pp. 5, 7

NOTES FOR THE THIRD EDITION

• The third edition adds chapters on forecasting, Lindy, and maximum-entropy distributions, including Tsallis entropy and associated q-distributions.Chapter 12 addresses single-point forecasting under fat-tailed threats; Chapter 13 discusses the forecasting industry; Chapter 23 derives Lindy from distance away from absorption.Source: NOTES FOR THE THIRD EDITION · p. 15
• Standard statistical tools designed for thin tails require preasymptotic adaptation to fat tails or may need to be abandoned.The book draws on published papers and uncensored commentary, and connects mathematics, philosophy, social science, contract theory, decision theory, and practical concerns.Source: §1 · pp. 17–18

2.1 general notations and frequently used symbols; 2.2.2 Law of Large Numbers (Weak); 2.2.3 The Central Limit Theorem (CLT)

• The glossary sets conventions for probability, expectation, variance, and distribution functions, and states that the weak law of large numbers gives convergence of the sample average to the expectation while the standard finite-variance CLT gives a Gaussian limit.The weak law requires an integrable i.i.d. sequence but not finite variance; the standard CLT passage assumes i.i.d. variables with finite variance.Source: §2.1 · p. 21, §2.2.2 · p. 24, §2.2.3 · p. 24

2.2.1 Power Law Class P; 2.2.4 Expected Excess Deviation Over a Threshold

• The power-law class is defined through a regularly varying survival tail, and beyond the Karamata point its tail can be calibrated using techniques such as the Hill estimator.The section also describes expected excess over a threshold as a diagnostic: thin-tailed variables have a characteristic scale, while fat-tailed regular-variation variables do not.Source: §2.2.1 · p. 23, §2.2.4 · p. 25

2.2.6 Kappa Metric

• Kappa measures preasymptotic behavior, ranging from 0 for the Gaussian benchmark to 1 for a Cauchy or a random variable without a mean.The metric is defined using the expected mean absolute deviation of partial sums centered around their mean.Source: §2.2.6 · p. 25

2.2.7 Elliptical Distribution; 2.2.8 Statistical independence

• An elliptical distribution requires a single covariance matrix for the joint distribution; regime switching or stochastic covariances violate ellipticality, and zero correlation implies independence for the bivariate Gaussian but not multivariate Student T or Cauchy.The text says that violating ellipticality can allow linear combinations of thin-tailed variables to produce explosive thick-tailed properties.Source: §2.2.7 · p. 26, §2.2.8 · p. 26

2.2.9 Stable (Lévy stable) Distribution; 2.2.10 Multivariate Stable Distribution; 2.2.13 Student T as Proxy

• A multivariate stable vector is defined by every linear combination of its components being stable; the Student T is used as a two-tailed power-law proxy, becoming Cauchy at α = 1 and Gaussian as α → ∞.The stable-distribution discussion presents stable laws as a generalization of the central limit theorem.Source: §2.2.9 · p. 26, §2.2.10 · p. 27

2.2.12 Subexponentiality

• For subexponential distributions, a sufficiently large sum is dominated by its largest observation, so the sum has the same magnitude as the sample maximum.The probability that two independent copies sum to more than x is asymptotically twice the probability that either copy alone exceeds x.Source: §2.2.12 · pp. 27–28

2.2.14 Citation Ring; 2.2.15 Rent seeking in academia; 2.2.16 Pseudo-empiricism or Pinker Problem

• The glossary describes pseudo-empiricism as using uninformative evidence or metrics, including means and correlations for fat-tailed variables, and identifies academic citation incentives as a related concern.It attributes pseudo-empiricism to thin-tail-focused education, weak probabilistic understanding, and neglect of dimensionality; it describes citation-driven prominence and career incentives as fostering research corners.Source: §2.2.14 · p. 28, §2.2.16 · p. 29, §2.2.15 · p. 29

2.2.5 Law of Medium Numbers or Preasymptotics; 2.2.17 Preasymptotics

• Preasymptotics concern behavior for large but finite n, where distributions that are Gaussian asymptotically may not yet behave that way.The book calls this intermediate regime the law of medium numbers and notes that convergence is especially rapid for Gaussian variables, unlike many others.Source: §2.2.5 · p. 25, §2.2.17 · p. 29

2.2.18 Stochasticizing

• Stochasticizing replaces a deterministic parameter with a two-state or full distribution, and metaprobability compares distributions or checks the robustness of resulting prices and risk metrics.The glossary also defines VaR and CVaR/Expected Shortfall at a threshold; when standard deviation is stochasticized, its square can represent stochastic volatility with variability termed Vvol.Source: §2.2.18 · pp. 29–30, §2.2.19 · p. 30, §2.2.31 · p. 34

2.2.21 MS Plot

• The MS plot tracks the maximum observation’s contribution to the sum as sample size grows, while the sample maximum converges to the distribution’s right endpoint in its maximum domain of attraction.The MS plot uses this behavior to examine moment existence through the law of large numbers.Source: §2.2.21 · p. 30, §2.2.22 · p. 31

2.2.23 Substitution of Integral in the psychology literature

• The psychology literature’s substitution of threshold impact times exceedance probability for expected payoff is valid only for constant impact, or approximately under thin tails.For a variable impact function with positive first derivative, the two quantities need not be close under fat-tailed distributions.Source: §2.2.23 · p. 31

2.2.24 Inseparability of Probability (another common error)

• Probability is a kernel within a payoff equation, and the glossary links its application to contingent-claim pricing, skin in the game, six-sigma evidence for a power law, and tail dependence.Tail dependence is defined through upper- and lower-tail indices; skin in the game filters for exposure to harm after failure.Source: §2.2.20 · p. 30, §2.2.24 · p. 32, §2.2.25 · p. 32

2.2.26 Black Swans

• Black Swans reflect observer-dependent knowledge gaps and consequential events that should be met with convexity, not prediction; fat tails magnify their impact.Their statistical attributes may remain elusive even when fragility to classes of events can be detected or measured.Source: §2.2.26 · pp. 32–33

2.2.27 The Empirical Distribution is Not Empirical

• The empirical distribution’s convergence concerns probabilities, not moments, and sample censoring at observed extremes can have large consequences for fat-tail analysis.The text argues that fat tails should be analyzed in payoff space rather than probability space.Source: §2.2.27 · p. 33

2.2.28 The Hidden Tail

• The hidden tail above the sample maximum has distribution-independent behavior for probabilities, but higher moments depend on the underlying distribution.Plug-in estimation uses maximum-likelihood parameters, such as tail exponent α, to derive shadow means or higher moments rather than relying on the directly observed sample mean, which is biased under fat tails.Source: §2.2.28 · p. 34, §2.2.29 · p. 34

2.2.32 Dynamic Hedging

• Dynamic hedging does not achieve replication in fat-tailed environments because of their special preasymptotic properties.The proposed replication is intended to make a European call payoff deterministic as hedge intervals approach zero.Source: §2.2.32 · p. 35

3.1 on the difference between thin and thick tails

• Insurance works in Mediocristan, where losses are more likely to arise from many events, but catastrophe risk makes uncapped contracts unsafe.For subexponential distributions, ruin is more likely to come from a single extreme event than a series of bad episodes; the text calls this the catastrophe principle and cites its risk-theory roots in Lundberg and Cramer.Source: §3.1 · p. 40
• For a non-fat-tailed distribution, two 3-sigma events (1.8 × 10^-6) are more probable than one 6-sigma event (9.86 × 10^-10); farther into the tail, large deviations arise through intermediate deviations.Source: §3.1 · p. 40

3.2 tail wagging dogs: an intuition

• As tails thicken, information shifts from the distribution’s body to rare tail deviations, slowing the law of large numbers and making a single disconfirming observation more informative than many confirmations.The chapter illustrates this with the example that a million observations of white swans do not establish that black swans do not exist.Source: §3.2 · pp. 43–44

3.3 a (more advanced) categorization and its consequences

• The chapter ranks tails from subexponential distributions to power laws: lognormals can shift from thin- to very thick-tailed behavior, while sufficiently heavy power laws lack variance or even a mean.It notes that power laws with α ≤ 2 have no variance, and the Fuhgetaboudit category has no mean; the text warns that conventional methods stop working as distributions leave the statistical comfort zone.Source: §3.3 · pp. 45–46

3.4 the main consequences and how they link to the book

• Because real samples may not reach asymptotic limits, statistical methods designed around the central limit theorem and law of large numbers can misrepresent thick-tailed data.The chapter stresses that convergence is distribution-dependent and that even when a Gaussian center emerges, consequential remote tails may persist.Source: §3.4 · pp. 46–47
• Finite-variance distributions with infinite higher moments can look Gaussian centrally only within ±n log n under summation, while remote tails remain consequential; life happens in the preasymptotics.The cited discussion attributes this convergence result to Bouchaud and Potters and Sornette.Source: §3.4 · p. 47

Consequence 1

• In very thick-tailed, one-tailed distributions, rare events prevent direct estimation of the population mean from the sample mean; some 80/20 power laws place 92% of observations below the true mean.The text says orders of magnitude more data may be needed for an informative sample average, and variance is likely to be underestimated.Source: Consequence 2 · p. 49
• Of 40,000 examined securities, none appeared thin-tailed, and standard deviation, variance, Beta, and Sharpe ratios are unreliable or uninformative.The Sharpe ratio is reported to fail out of sample, sometimes acting in the reverse of its intended predictive role; variance estimates can remain imprecise even when variance exists.Source: Consequence 4 · pp. 49–50, Consequence 3 · p. 49

Consequence 5

• Robust statistics can suppress informative tail observations, while winsorization distorts expectations and the empirical distribution misrepresents tail payoffs.The warning applies to finance and risk management; the text also cautions that nonparametric methods can increase distribution dependence problems.Source: Consequence 5 · pp. 50–51
• Under thick tails, least-squares regression and R2 can mislead because finite samples distort squared-deviation fits and variance may be infinite.If the second moment exists, very large samples may still be needed; if it does not, least squares is unavailable, and minimizing absolute deviations may not yield a unique slope. Finite samples can make R2 appear positive even when it should be zero.Source: Consequence 6 · p. 51
• Maximum-likelihood estimates of some distribution parameters can improve mean estimation, but when no reliable estimator exists, the author advises avoiding unbounded exposure.For a Pareto distribution, estimating the tail exponent and scale can yield a plug-in mean estimate more reliably than directly sampling the mean; the approach does not work in every situation.Source: Consequence 7 · pp. 51–52

Consequence 8

• Thick tails can produce spurious PCA or factor-analysis results, invalidate moment matching when higher moments are unstable or nonexistent, and make large deviations non-convergent.The method-of-moments problem also applies to GMM; conditional on a large move, its magnitude may not converge, especially for power-law tails.Source: Consequence 9 · p. 53, Consequence 10 · p. 53, Consequence 11 · p. 53

Consequence 12

• Conventional Gini and related concentration measures can become super-additive under thick tails, making aggregate inequality appear higher than the weighted average of its parts.The warning also applies to measures such as the share of wealth held by the top 1%.Source: Consequence 12 · p. 54

Consequence 13

• Large deviation theory, as described here, does not apply to thick-tailed distributions.The passage distinguishes this limitation from thin-tailed settings, where the cited methods are useful.Source: Consequence 13 · p. 54

Consequence 14

• The author argues that dynamic hedging never mitigates financial option risk because its required exponential decline of tail probabilities is incompatible with thick tails.The described hedging basis requires distributions outside the sub-exponential class.Source: Consequence 14 · p. 54

3.5 epistemology and inferential asymmetry

• In Extremistan, risk management should reduce the impact of an event rather than focus only on its probability or frequency.The banking losses cited for 1982, 1991, and 2008–9 illustrate how a single event can erase cumulative gains; tallying events also requires the Cramer condition of non-subexponentiality.Source: §3.5 · p. 58

3.4.2 The Law of Large Numbers

• 1011 observations in the Pareto case versus 30 in the Gaussian case are needed to stabilize the mean to the same level, assuming the mean exists.The section cautions that sample means alone cannot support claims about their stability under thick tails.Source: §3.4.2 · p. 57, Table 3.1 · p. 59

3.5 epistemology and inferential asymmetry

• The Russian school of probability emphasized inequalities and bounds rather than precise estimates, accommodating one-sided knowledge and skepticism about probability claims.The passage connects this approach to risk bearing, where one may know a minimum acceptable insurance price without knowing an upper bound, or vice versa.Source: §3.5 · p. 61
• A large deviation can rule out thin-tailedness, but an absence of large deviations cannot establish it; inference can instead rank distributions by elimination.For example, the passage says a 20 sigma event rules out a thin-tailed distribution, while no large deviation does not rule out thick tails without understanding the process.Source: §3.5 · p. 63
• Future parameter uncertainty should make tails fatter, not thinner; fitting a stable distribution is more likely to overestimate the tail index, where lower values mean fatter tails.The passage gives Argentina’s market before and after the large drop of August 12, 2019, as an illustration and attributes the tail-index finding to Rafal Weron.Source: §3.5 · p. 63

3.6 naive empiricism

• Comparing observed Ebola deaths with diabetes or bedsheet deaths is naïve: extreme-scale outcomes from multiplicative fat-tailed risks cannot be compared directly with thin-tailed risks.The text argues that this error leads people to worry too much about diabetes and too little about Ebola and other ailments with multiplicative effects.Source: BEFORE · p. 64, §3.6 · p. 65, Figure 3.22 · p. 65

3.6.1 How some multiplicative risks scale

• Individual risk comparisons can miss collective effects: contagion raises others’ infection risk, making evidence-based approaches inadequate for second-order risks and risk management.The text contrasts this with swimming-pool drowning, which does not increase neighbors’ drowning risk, and frames contagious-disease harms as extending beyond one’s own outcome.Source: §3.6.1 · p. 67

3.7 primer on power laws (almost without mathematics)

• Across 55 years of S&P 500 data, the largest observation contributed about 80% of kurtosis, versus a Gaussian maximum contribution of around 0.008 ± 0.0028 over the same span.The passage says Pareto distributions have no higher moments: they either do not exist or become increasingly unstable. For uncapped data or open variables, Pareto distributions offer a less informative but more rigorous approach.Source: §3.7 · pp. 68–69
• Principal components can show spurious structure for uncorrelated variables under thick tails, requiring much more data than under Gaussian tails for that structure to flatten.The cited simulation compares 30 uncorrelated variables at n = 100 and n = 1000; Gaussian components flatten as n increases, while the thick-tailed case does not in reasonable time.Source: §3.7 · p. 69, Figure 3.26 · p. 70

3.8 where are the hidden properties?

• The sample mean can misrepresent fat-tailed distributions; fitting the distribution and deriving a “shadow” mean is preferable to treating the observed mean as the population mean.The text says that for a Pareto distribution 98% of observations fall below the mean, and that short-volatility series can overestimate performance while long-tail volatility series underestimate it.Source: §3.8 · p. 71

WITTGENSTEIN’S RULER: WAS IT REALLY A "10 SIGMA EVENT"?

• At 10 standard deviations, the Gaussian survival probability is 1 in 1.31 × 10^23, versus 1 in 203 for a same-scale Student T distribution with tail exponent 2.The text recommends rejecting Gaussianity when an event exceeds 4 or 5 standard deviations, even if the probability that the data are non-Gaussian is below 10^-10.Source: WITTGENSTEIN’S RULER: WAS IT REALLY A… · p. 72

3.9 bayesian schmayesian

• Bayesian methods cannot manufacture missing information under thick tails and depend on a reliable prior; observers with the same information need not converge in real time or at all.The text allows Bayesian parameter estimation when the parameter range is clear and the parameter has a tractable, low-variance distribution, as with a Pareto tail exponent.Source: §3.9 · p. 74

3.10 x vs f(x): exposures to x confused with knowledge about x

• The subsection introduces F(X) as a variable’s exposure, payoff, or effect on an observer, distinguishing that outcome from X itself.Source: §3.10 · p. 74
• Nonlinear F(X) can have statistical properties and risks markedly different from X, so the section urges focusing on the exposure one can alter rather than trying to predict X.For convex-concave transformations of wealth, F(X) is described as more stable and thinner-tailed than X; nonlinearities also make the mean of F(X) differ from F(mean of X).Source: §3.10 · pp. 75–77

3.11 ruin and path dependence

• Ensemble probability across gamblers does not transfer to an individual exposed to ruin, and time probability differs from ensemble probability.Ruin is an absorbing barrier: once an individual is ruined, there is no next day; the passage says the correspondence requires a risk allocation policy compatible with the Kelly criterion, using logs.Source: §3.11 · p. 77
• Equal mean variation (0.71) produced P/L 2 under thin tails versus −18 under thick tails, showing that tail lumping can change outcomes despite correct forecasts.The example uses a variance-swap payoff; the thick-tailed succession ends in ruin.Source: §3.11 · p. 78

Remark 3.2: Psychology of decision making

• Psychological analyses of single-episode risks can label people paranoid for overestimating small risks, while overlooking collective tail risks and risks that accumulate.The passage connects precaution to units expected to survive and says the ecosystem's destruction is the largest risk.Source: Remark 3.2: Psychology of decision… · p. 80

3.12 what to do?

• Valid analysis requires distinguishing Mediocristan from Extremistan, which the section describes as domains that almost never overlap.The passage says failing to distinguish them leaves no valid analysis.Source: §3.12 · p. 80

3.13 kahneman’s parting gift

• The section recommends heuristics and policies to detect exposure harm instead of fabricating statistical properties; detecting convexity and concavity is described as simpler than probability.Fragility entails a concave, option-like exposure whose harm accelerates in the tails, a property used to build stress-testing heuristics.Source: §3.13 · p. 81

Remark 4.1: Fat Tails and Jensen’s inequality

• Because Gaussian tail probabilities are convex in scale, randomizing variance or standard deviation while conserving the mean fattens tails; the author’s two-volatility option-price averaging heuristic captured much of the effect.Source: §4 · p. 83, Remark 4.1: Fat Tails and Jensen’s… · pp. 84–85

4.1.1 A Variance-preserving heuristic

• The simple variance-preserving heuristic preserves the second moment and odd moments are nil, but raises kurtosis only to twice the Gaussian level, making it pedagogical.Source: §4.1.1 · p. 86

Remark 4.2: Peaks

• Fat tails manifest as higher peaks, concentrating more observations around the distribution’s center.Source: Remark 4.2: Peaks · p. 86

4.1.2 Fattening of Tails With Skewed Variance

• With p = 1/1000 and maximum a = 999, switching between variance-preserving Gaussian states can produce kurtosis as high as 3000.Source: §4.1.2 · p. 86
• A gamma-distributed Gaussian variance provides an analytically tractable shortcut to a full variance distribution and extends beyond narrow heuristic methods.Source: §4.1.2 · p. 87

4.2 does stochastic volatility generate power laws?

• Whether stochastic volatility yields a power law depends on whether σ, σ², or their reciprocals are randomized; lognormal σ² is explicitly not power-law.Source: §4.2 · p. 88

4.3.1 The Crossovers and Tunnel Effect.

• Scale perturbation increases density in the high-peak inner tunnel and outer tails but decreases it in the intermediate shoulders; Gaussian crossovers are −2.13σ, −.66σ, .66σ, and 2.13σ.Source: §4.3.1 · pp. 90–91

In Summary, Where Does the Tail Start?

• The tail for symmetric power-law distributions begins around 2–3 standard deviations; biased-low historical standard deviation estimates can push this beyond 2–3 standard deviations.Source: In Summary, Where Does the Tail Start? · p. 92
• Under some conditions, one-tailed bell-shaped distributions share four tail crossover points, with the Lognormal as a special case.The section also notes that asynchronous changes in individual distribution scales can make a distribution cease to be elliptical.Source: In Summary, Where Does the Tail Start? · pp. 92–93

4.4.1 The Common Errors

• Professionals often substitute mean absolute deviation (MAD) for standard deviation (STD), underestimating STD by 20–40% on average and up to 90% in some markets.For daily absolute moves averaging 1%, respondents commonly answered 1% rather than about 1.25% under a Gaussian model; the yearly answer was often about 16%, roughly 80% of the true answer.Source: §4.4.1 · p. 94

4.4.2 Some Analytics

• For a Gaussian, the ratio of standard deviation to mean deviation is about 1.25, and it rises with fat-tailedness.The section presents this ratio as a heuristic for finite moments and notes that the ratio-of-moments heuristic works only in sample when moments are undefined.Source: §4.4.1 · p. 94, §4.4.2 · p. 94, §4.4.4 · p. 100

Finalmente, the Asymptotic Relative Efficiency For a Gaussian

• Conditional on Gaussian data, STD is 12.5% more efficient than mean deviation, but even slight contamination can overturn that advantage.The section calls this efficiency result fragile outside the Gaussian condition.Source: Finalmente, the Asymptotic Relative… · p. 96

4.4.3 Effect of Fatter Tails on the "efficiency" of STD vs MD

• With p = .01 and n = 1000, small outliers of 5 standard deviations make MAD five times more efficient than STD.For the finite-variance power-law example, the mean deviation of standard deviation is five times higher; the text attributes this to STD having infinite variance.Source: §4.4.3 · p. 97

4.4.4 Moments and The Power Mean Inequality

• For a fixed norm, greater dispersion raises its value, with higher norms becoming rapidly explosive as spread increases.Finite samples always produce computable moments even when population moments are undefined, but different samples can yield very different values and apparent convergence can be interrupted by jumps.Source: §4.4.4 · pp. 98–100

4.4.5 Comment: Why we should retire standard deviation, now!

• For the changes (-23, 7, -3, 20, -1), root mean square deviation is 15.7 while mean absolute deviation is 10.8.The section argues that mean deviation better reflects real-life variation and that people often act on STD figures as if they represented expected mean deviation.Source: §4.4.5 · p. 101
• Some people mistake MAD for STD because MAD comes to mind more readily, and economists dismissed infinite-variance models by conflating infinite variance with infinite mean deviation.The section cautions that whenever the mean exists, MAD exists, so infinite variance need not imply infinite mean deviation.Source: §4.4.5 · pp. 101–102
• MAD is more accurate in sample measurements and less volatile than STD because STD gives large observations greater weight, overweighing tail events.Whenever the mean exists, MAD exists; the reverse case— infinite MAD with finite STD—never occurs, although economists have conflated infinite variance with infinite mean deviation.Source: §4.4.5 · pp. 101–102

4.5 visualizing the effect of rising p on iso-norms (

• As p rises, iso-norms occupy a larger share of space, while the p = 1 share relative to p = ∞ falls from 1/2 in two dimensions to 1/6 in three.The text presents this growing ratio of higher to lower moments with dimensionality as a hint of the curse of dimensionality, relevant to model error.Source: Figure 4.13 · p. 102, §4.5 · pp. 103–104, Figure 4.14 · p. 103

5.0.1 Revisiting the Rankings

• The subexponential class marks the boundary between thin and fat tails, while power-law moment existence depends on the tail exponent.Power laws have a finite mean only when the exponent exceeds 1 and finite variance only when it exceeds 2; the section also distinguishes thin-tailed D1, fat-tailed regular-variation D2, and a borderline exponential class.Source: §5.0.1 · p. 105, §5.0.2 · p. 108
• For subexponential distributions, a large sum is dominated by its largest observation, which captures why tail events matter most.For two independent copies, the exceedance probability of their sum is asymptotically twice the individual exceedance probability; the text extends maximum dominance to sums of n variables.Source: §5.0.1 · pp. 106–107

5.0.3 Let Us Invent a Distribution

• The constructed borderline distribution has mean zero, zero odd moments, kurtosis 21/5, and tails slightly fatter than a Gaussian.The section proposes constructing a distribution at the boundary between thin and fat tails.Source: §5.0.2 · p. 107, §5.0.3 · p. 109

5.1.1 Scalable and Nonscalable, A Deeper View of Fat Tails

• Only scalable distributions have “true” fat tails; their tails are regularly varying, while nonscalable distributions turn Gaussian under summation.The tail exponent is asymptotic and may not be reached in practice; real processes can also switch exponents across deviation sizes.Source: §5.1.1 · pp. 109–111

5.1.2 Grey Swans

• Tail-exponent calibration is difficult because tail data are insufficient, and observed data will generally show thinner tails than they should.Student t is used to simulate symmetric power laws for convenience, not as a claim about the generating process.Source: §5.1.2 · pp. 111–112

5.2.1 Sums of variables

• Adding even one power-law summand can make a sum power-law tailed, regardless of thin-tailed components, and can leave moments undefined when that summand has undefined moments.The sum's asymptotic tail is dominated by the fatter tail; the same implication applies to mean, variance, or higher moments.Source: §5.2.1 · p. 113

5.2.2 Product of variables

• Products of power-law variables likewise show asymptotic dominance by the fatter tail, and mixing power-law with thin-tailed variables produces power laws.The passages state that this holds regardless of the composition.Source: §5.2.2 · p. 114

Property 5.2

• A finite-variance random variable with tail exponent below 4 can have infinite variance after the discussed power transformation.The section presents this as a consequence relevant to stochastic-volatility models and as an example of how convex transformations thicken tails.Source: Property 5.2 · p. 114

5.3 bell shaped vs non bell shaped power laws

• A double Pareto and a Student T with the same power-law exponent can have contrasting bell-shaped and angular double-convex forms, although their moments eventually coincide.The shape differences appear for small deviations and reflect the slowly varying part of the tails.Source: §5.3 · p. 115, Figure 5.6 · p. 115

5.4 interpolative powers of power laws: an example

• The COVID-19 jobless-claims jump barely changed the tail exponent, which remained robust to the out-of-sample observation.The Zipf plot indicates the data were already power-law distributed without the pandemic jump; the scale could vary, but the exponent was robust.Source: §5.4 · p. 116, Figure 5.8 · p. 116

5.5 super-fat tails: the log-pareto distribution

• The log-Pareto is the Paretian analog of the lognormal, and its log-log survival slope tends to 0, so no moment exists for any tail parameter α.This contrasts with the asymptotically linear log-log slope of a regular power law.Source: §5.5 · pp. 116–117, Figure 5.9 · p. 117

5.6 pseudo-stochastic volatility: an investigation

• A power law can be mistaken for heteroskedasticity because observed data do not reveal the generating process, allowing volatility changes to be recast as changing sigma deviations.The section proposes an inverse-problem perspective: a constant-scale power law can masquerade as a heteroskedastic process.Source: §5.6 · pp. 117–118
• Constant-scale Student T returns can produce realized-volatility patterns that resemble stochastic volatility, while squaring returns yields a power-law tail with exponent α/2.The apparent volatility pattern arises even though the return distribution's scale is constant.Source: §5.6 · pp. 118–119, Figure 5.10 · p. 118

6.1 thick tails in higher dimension, finite moments

• Thick-tailed multivariate distributions concentrate more density near the center than thin-tailed distributions with the same variance.The text attributes this pattern to making variances stochastic while keeping correlations constant to preserve positive definiteness.Source: §6.1 · p. 122

Diversification

• Under ellipticality, portfolios with identical location and scale have identical return distributions, but nonelliptical returns make Markowitz-style diversification fail to reduce risk, and leverage makes long-run blowups certain.Elliptical distributions are closed under linear transformation; the text says financial returns are not elliptical by any conceivable measure.Source: §6.2 · pp. 125–126, Diversification · p. 126

6.3.1 Ellipticality and Independence under Thick Tails

• The multivariate Student T offers a convenient model, including the Cauchy at α = 1, but in these elliptical distributions, zero correlation does not imply independence.The bivariate Gaussian with zero correlation is both independent and uncorrelated; this does not hold for Student T or Cauchy distributions.Source: §6.3 · p. 127, §6.3.1 · p. 127

6.4 fat tails and mutual information

• Mutual information can detect nonlinearities and is more potent than correlation for measuring dependence in fat-tailed settings.The section notes that mutual information need not be zero without correlation in the distributions considered.Source: §6.4 · pp. 128–129

6.5 fat tails and random matrices, a rapid interlude

• For random matrices, the fourth moment reaching the Gaussian benchmark of 3 is equivalent to eigenvalues reaching Wigner’s semicircle distribution.The semicircle is the limiting eigenvalue distribution for symmetric matrices with finite moments as matrix size tends to infinity.Source: §6.5 · pp. 129–130

6.6 correlation and undefined variance

• Correlation can remain bounded between −1 and 1 even when moments and covariance are undefined, but its sampling noise and slow convergence can make it uninformative.Finite correlation does not mean low variance; the section also likens the problem to PCA’s sampling-error issue.Source: §6.6 · pp. 130–131

6.7 fat tailed residuals in linear regression models

• Fat-tailed regression residuals can make R2 misleading: clipping income tails creates an illusion of high R2, and an infinite-variance example has high in-sample R2 but ultimately a value of 0.The section frames regression of power-law income or wealth on Gaussian IQ scores as improper; its model assumes Gaussian X and power-law residuals.Source: §6.7 · p. 131, Figure 6.12 · p. 132, Figure 6.11
• .985: a Cauchy regression with expected R2 = 0 can show a high value in a small sample.The sample statistic is finite and highly sample-dependent, stabilizing only at large, potentially astronomically large n.Source: Figure 6.13 · p. 132, §6.7 · p. 133
• When a fat-tailed variable is regressed against a thin-tailed one, R2 is biased upward and needs a much larger sample to converge, if it does.Because R is bounded in [0, 1], it reaches its true value very slowly; the text says this invalidates many IQ–wealth and IQ–income studies.Source: §6.7 · p. 133
• Nonlinear transformations can establish a linear relationship only when the required transformation is exact; discrepancies can cause errors, and correlation is delicate, non-additive, and often uninformative.Source: §6.7 · p. 133

A SPECIAL CASES OF THICK TAILS

• Thick tails are described as the norm for unimodal distributions, while multimodal distributions can produce surprises, including kurtosis below the Gaussian value.The War and peace model is shown with kurtosis = 1.7, below the Gaussian value; another figure depicts negative relative kurtosis with bimodality.Source: A SPECIAL CASES OF THICK TAILS · p. 135, Figure A.2 · p. 135, Figure A.3 · p. 136

a.1 multimodality and thick tails, or the war and peace model

• Stochasticizing a deterministic variable's variance can increase thick-tailedness while preserving its mean, whereas stochasticizing the mean and separating outcomes can lower kurtosis.Associating different variances with different means yields regimes, each with its own probability.Source: a.1 multimodality and thick tails, or… · p. 136
• Under multimodality, “thick tails” may lose significance or take on a new meaning as the region around the expectation ceases to matter.Source: a.1 multimodality and thick tails, or… · p. 136
• In the war-and-peace variety, calm periods have positive means and very low volatility, while turmoil has negative means and extremely high volatility.The model's central condition is σ1 > σ2, with p1 < (1 − p1) and m1 < m2.Source: a.1 multimodality and thick tails, or… · p. 137
• In the bond model, realizations at or near the expectation can be rarer than outcomes at either regime mean, making the mean-region realization a tail event.The model describes discrete outcomes: a high-probability payoff of B(1 + r) and a small possibility of default, with receiving exactly B very unlikely.Source: a.1 multimodality and thick tails, or… · p. 137
• Bimodal option payoffs raise at-the-money option values and lower out-of-the-money values, the opposite of the described volatility smile.Source: a.1 multimodality and thick tails, or… · p. 137

3. Professional choices and outcomes

• Conflicts include interpersonal, general, or martial situations with no intermediary between harmony and hostility.Source: §3 · p. 138

a.2 transition probabilities: what can break will break

• In the absorbing-state case, regime 1 eventually dominates with probability 1 when what can break is irreversible.The section frames this as a long-term exercise under an unchanged regime structure.Source: a.2 transition probabilities: what can… · pp. 138–139

b.2 tsallis entropy

• Tsallis entropy generalizes additive Shannon entropy for nonextensive systems and is presented as a way fat-tailed distributions emerge naturally.For independent systems, its entropy of the sum includes an interaction term proportional to (1 − q) Sq(A)Sq(B); when q < 1, p(x)^q increases small probabilities.Source: B FAT TAILS FROM MAXIMUM ENTROPY · p. 141, b.2 tsallis entropy · p. 144

b.1 shannon/boltzman-gibbs entropy

• For continuous densities, the section maximizes differential entropy under normalization and constraints using calculus of variations.The maximum follows from the functional's concavity.Source: b.1 shannon/boltzman-gibbs entropy · p. 142

B.1.1 Constraining the Mean and Variance leads to the Gaussian

• Under a mean and central-variance constraint, maximizing Shannon entropy yields the Gaussian distribution.The text connects the Gaussian's emergence to systems constrained by energy (x^2).Source: §B.1.1 · pp. 142–143

B.1.2 Power Law Constraints

• A finite multiplicative-outcome constraint, expressed through E(log(x)) on [L, ∞], yields a power law.The derivation sets λ1 = −1 − α to obtain the familiar form.Source: §B.1.2 · p. 143

B.1.3 A Good Epistemic Intuition

• Maximizing entropy for a distribution on [0, 1] with normalization alone gives the uniform distribution.The text presents this as the maximum-ignorance choice when the probability distribution is unknown.Source: §B.1.3 · p. 143

b.2 tsallis entropy

• Maximizing Tsallis entropy under an escort-mean constraint gives a q-exponential distribution: q > 1 has heavy tails, q < 1 compact support, and q → 1 recovers the exponential.Source: b.2 tsallis entropy · pp. 144, 146

7.1 refresher: the weak and strong lln

• For i.i.d. variables with finite mean, the sample average converges to the expectation as n approaches infinity; finite variance is not required.The refresher distinguishes weak convergence in probability from strong, almost-sure convergence and notes that finite higher moments accelerate convergence.Source: §7.1 · pp. 151–152

7.2.1 The Stable Distribution

• The generalized CLT describes normalized sums of i.i.d. variables converging in distribution to a stable law, a class stable under summation that includes Gaussian, Cauchy, and Lévy distributions.The chapter distinguishes finite-n convergence from the limiting stable distribution and its basin.Source: §7.2 · p. 153, §7.2.1 · pp. 153–154, §7.2.2 · p. 155

7.3.1 Fast Convergence: the Uniform Dist.

• With three independent uniform [0, 1] summands, the sum is already almost bell-shaped, illustrating fast convergence toward Gaussianity.The passage describes the two-summand density as triangular and the three-summand density as almost bell-shaped; Figure 7.1 illustrates sums through four summands.Source: Figure 7.1 · p. 153, §7.3.1 · p. 155

7.3.2 Semi-slow convergence: the exponential

• Exponential sums approach the Gaussian more slowly than uniform sums, largely because their skewness takes time to disappear.The chapter says that getting to the Gaussian requires symmetry.Source: Figure 7.4 · p. 156, §7.3.2 · p. 157

7.3.3 The slow Pareto

• At n = 100 and n = 1,000, Pareto sums show little improvement toward Gaussianity, although the α = 2 case eventually converges.For αp > 2, the passage says the limit is Gaussian; computing Pareto-sum densities can be exponentially slow, so the authors use Monte Carlo simulations.Source: Figure 7.6 · p. 158, §7.3.3 · p. 159

7.3.4 The half-cubic Pareto and its basin of convergence

• The half-cubic Pareto never becomes symmetric in real life and is, in a sense, more fat-tailed than a stable distribution.The passage cautions against conflating its properties with those of a stable distribution.Source: §7.3.4 · p. 159, Figure 7.7 · p. 159

7.4 cumulants and convergence

• Economic variables show no convergence of kurtosis under summation, while thin-tailed distributions have finite cumulants and a clear basin of converging versus diverging moments.The chapter notes that the expected N^-2 kurtosis drop for stochastic volatility does not occur; it reports an absence of convergence across economic variables.Source: §7.4 · p. 160, Figure 7.8 · p. 160

7.5 technical refresher: traditional versions of clt

• The chapter’s CLT refresher states that Lyapunov’s condition implies Lindeberg’s, but not conversely, while Lindeberg’s condition suffices for Gaussian convergence.The standard Lindeberg–Lévy version assumes i.i.d. variables with finite variance; Lyapunov allows independent, non-identically distributed summands with moments of order 2 + δ.Source: §7.5 · pp. 161–162
• Lindeberg’s condition holds if and only if the central limit theorem result holds.Source: §7.5 · p. 162

7.6.1 Higher Moments

• For moments that do not exist, cumulative averages can jump as observations are added; when moments exist, further additions eventually prevent jumps.The chapter also proposes examining the maximum observation’s share of the total through the maximum-to-sum (MS) plot.Source: §7.6.1 · pp. 162–163, Figure 7.9 · p. 164

8.2 the metric

• The κ metric uses expected mean absolute deviation from the mean for sums of finite-mean variables, rather than dispersion measured in the L2 norm.The passage motivates this choice for finite-mean, infinite-variance settings, where L2 cannot measure compression around the mean.Source: §7.7 · p. 165, §8.2 · pp. 170, 172

7.7 mean deviation for a stable distributions

• κ can guide sample-size and Monte Carlo planning, compare convergence across distributions, and assess diversification needs and lognormal behavior.The stated applications also include measuring convergence speed toward the Gaussian or stable basin.Source: §7.7 · p. 167

8.1 introduction and definitions

• κ compares sums of finite-mean variables before the asymptote and measures how sample-size growth changes mean absolute deviation.It is intended for comparing distributions or sample sizes and assessing the speed of the law of large numbers.Source: §7.7 · p. 168, §8.1 · pp. 169–170, §8.2 · p. 170

8.3 stable basin of convergence as benchmark

• κ also measures distance from the limiting Lévy α-stable basin, with the Gaussian as a special case.Source: §8.1 · p. 170

8.3.2 Practical significance for sample sufficiency

• κ above .15 heuristically signals unreliable normal approximation and disproportionally insufficient confidence intervals.The passage presents this as a practical heuristic for judging sample sufficiency.Source: §8.3.2 · p. 173
• A Pareto distribution with α ≈ 1.14 requires >109 more observations than a Gaussian, while at α = 3 a one-tailed Pareto needs 543 observations to match a Gaussian sample of 30.A Student T with 3 degrees of freedom needs 120 observations to match the variance reduction of a Gaussian sample of 30; the contrast shows tail exponent and finite variance alone do not determine sample sufficiency.Source: §8.3.2 · p. 174

8.4.1 Some Oddities With Asymmetric Distributions

• A skewed stable distribution retains the same κ index as its symmetric counterpart, but a one-tailed Pareto is fatter-tailed than an equivalent symmetric distribution.The text emphasizes that Pareto and stable distributions are not interchangeable, even for large numbers of summands.Source: §8.4.1 · p. 175

8.4.2 Rate of Convergence of a Student T Distribution to the Gaussian Basin

• For the Student T with α = 3, κ converges to 0 at rate 1/log(n), and even 10^6 summands may not yield Gaussian behavior.The convergence rate differs between the center and tails; the appendix also supplies an explicit density for the sum. The exponential distribution likewise has slow asymptotic convergence.Source: §8.4.2 · p. 175, Proposition 8.1 · pp. 179–180, §8.6.3 · p. 182

8.4.3 The Lognormal is Neither Thin Nor Fat Tailed

• At low σ, a lognormal behaves like a Gaussian, while at high σ it appears Cauchy-like; sums behave Gaussian-like at low σ and lognormal-like at high σ.For high σ, κ approaches 1 as the law of large numbers slows; the text also notes that some apparent Paretian cases are high-variance lognormal situations with smaller-than-imagined practical statistical consequences.Source: §8.4.3 · pp. 175–176, §8.6.2 · pp. 180–181

8.4.4 Can Kappa Be Negative?

• κ can be negative in mixed-Gaussian distributions with switching means and fixed variance when the modes are far apart, but such cases do not appear in unimodal distributions.The appendix considers a Gaussian mixture with switching means and variances.Source: §8.4.4 · p. 176, §8.6.4 · p. 183

8.5.1 Portfolio Pseudo-Stabilization

• The method extends sample-sufficiency reasoning to portfolio diversification, where adding securities has a stabilizing role, and modern portfolio criteria can underestimate risk by at least 8 orders of magnitude for a near-80/20 Pareto.The text reports κ around .2 for the S&P 500 and between .3 and .7 for single stocks, and argues broader portfolios are needed.Source: §8.5.1 · pp. 176–177, §8.5 · p. 176
• The noisy 1/n diversification heuristic is described as rational rather than biased under fat tails, and equally weighted portfolios are said to outperform the S&P 500 across a broad range of metrics.The text warns that policy nudges against this diversification can increase portfolio risk.Source: §8.5.1 · p. 177

8.5.2 Other Aspects of Statistical Inference

• In higher dimensions, sampling extreme fat-tailed variables is not captured by Marchenko-Pastur or Wishart distributions, and adding variables does not easily remove random-matrix noise.The section concludes that statistics is not standard, though heuristics can indicate where and how much it departs from standard behavior.Source: §8.5.2 · pp. 177–178, §8.5.3 · p. 178

8.6.4 Negative Kappa, Negative Kurtosis

• For equal-variance bimodal distributions, kappa is negative across wide mean separations, with thinner-than-Gaussian tails becoming acute when means differ by several standard deviations.Source: §8.6.4 · p. 183
• The section presents simple power-law diagnostics applied to the SP500.Source: §8.6.4 · p. 183
• The diagnostics contrast naive methods with ML estimators that allow extrapolation into the tails.Source: §8.6.4 · p. 183

9 EXTREME VALUES AND HIDDEN TAILS ∗,†

• Past maxima are poor predictors of future extrema because thick-tailed distributions have unobserved tail outcomes.Treating the highest level personally observed as the highest possible is described as the Lucretius fallacy.Source: §9 · p. 185

9.1 preliminary introduction to evt

• Exact finite-sample maximum distributions can be used to study preasymptotic behavior, when the sample maximum has not reached its limiting regime.The passage notes that this approach may require patience, computing power, and investigation.Source: §9.1 · p. 186

9.1.1 How Any Power Law Tail Leads to Fréchet

• For power-law variables, the maximum retains the variables’ tail exponent and converges asymptotically to a Fréchet distribution.The chapter also reports an exact asymptotic fitting for power-law extrema.Source: Property 9.1 · p. 188, §9.1.1 · p. 188, Property 9.2 · p. 188

9.1.2 Gaussian Case

• Gaussian maxima are harder to parametrize than power-law maxima, and the exact Gaussian maximum distribution is preferred for tail-risk work.The Gaussian limiting distribution is approached slowly, and the cited figure caption says the tail approximation is poor.Source: §9.1.2 · p. 189, Figure 9.4 · p. 189, Property 9.3 · p. 190

9.1.3 The Picklands-Balkema-de Haan Theorem

• Above a threshold, the limiting excess distribution is Pareto for ξ > 0, exponential for ξ = 0, and uniform for ξ = −1.The theorem supports inference from exceedances.Source: §9.1.3 · p. 191

9.2 the invisible tail for a power law

• For power-law samples, the hidden moment above the observed maximum has a derived distribution when the threshold exceeds K* and p > α.The passage defines the visible and hidden parts of the moment and states the conditions for the hidden-moment result.Source: §9.2 · pp. 191–192
• The empirical sample survival function’s exceedance probability depends only on sample size n, not on tail fatness.Separately, although empirical CDF errors have Gaussian asymptotics, tail error effectively increases when scaled by the payoff associated with the probability.Source: §9.2 · p. 193, §9.3 · pp. 194–195

9.3 appendix: the empirical distribution is not empirical

• For U.S. stock indices, empirical methods underestimate tail payoffs by 5 to 70 times.The comparison is with a smoothed empirical distribution and concerns a tail CVaR-style measure.Source: §9.3 · p. 195, Figure 9.11

C GROWTH RATE & OUTCOME: NOT IN THE SAME DISTRIBUTION CLASS

• Cirillo and Taleb (2020) show that future pandemic fatalities have a tail exponent α < 1, with survival function P(X > x) = L(x)x−α.They simplify by assuming L(x) is approximately constant at its minimum value.Source: C GROWTH RATE & OUTCOME: NOT IN THE… · p. 197

c.1 the puzzle

• For 106 realizations, exponentiating r changes sample kurtosis from 9 to 106, theoretically infinite for X=e^r, with values dominated by one large deviation.Source: Figure C.1 · p. 198
• Small but nonzero errors in r can become explosive in XT through exponentiation; conversely, exponential r implies Pareto-distributed XT under Eq. 15.13.Source: c.1 the puzzle · p. 199
• The change-of-variables sketch maps r to u=g(r), with transformed support [g(a), g(b)]; the text considers exponential and Laplace rate distributions.Source: c.1 the puzzle · pp. 198–199

c.2 pandemics are really fat tailed

• Random subsamples of the 60 largest pandemics retain Pareto properties and parameterization; EVT gives a higher tail exponent, still below one.The data are normalized by today’s population.Source: c.2 pandemics are really fat tailed · p. 199, Figure C.2 · p. 199
• The authors describe the preliminary pandemic analysis as having an exceptionally low tail exponent and recommend EVT or precautionary principles for policy making.They distinguish policy making from using epidemiology studies for research.Source: c.2 pandemics are really fat tailed · p. 200

• Large-deviation reasoning helps explain thin-tailed behavior and why fat-tailed outcomes may occur when the Cramér condition fails; exponential tail decay is necessary for effective diversification under the catastrophe principle.Source: D THE LARGE DEVIATION PRINCIPLE, IN… · p. 201
• For large N, the tail probability of the sample mean is expressed using the rate function I, which can be obtained by Legendre transformation when the distribution is known.The rate function is connected to the cumulant generating function.Source: D THE LARGE DEVIATION PRINCIPLE, IN… · p. 201
• The function θ(x) indicates a single event’s contribution to overall payoff and connects to the Cramér condition’s requirement of exponential moments.Source: D THE LARGE DEVIATION PRINCIPLE, IN… · p. 201
• For n = 1000, the cited 50%-above-mean coin-flip event occurs 1 in 1.24 × 10^18.Chernoff bounds give tight bounds for Bernoulli variables.Source: D THE LARGE DEVIATION PRINCIPLE, IN… · pp. 201–202
• Hoeffding extended Chernoff bounds to bounded independent random variables that need not be Bernoulli.Source: D THE LARGE DEVIATION PRINCIPLE, IN… · p. 202

E CALIBRATING UNDER PARETIANITY

• MLE tracks Pareto sample means more reliably than sample-mean methods, which underestimate under skew; the tail-exponent estimate converges rapidly toward Gaussian uncertainty, and the plug-in mean may work under one-tailed Paretianity.For Pareto α = 1.2, the passage says 107 more data are needed for Methods 1 and 2 to attain the same error rate.Source: E CALIBRATING UNDER PARETIANITY · p. 204, Figure E.2 · p. 204, e.1 distribution of the sample tail… · p. 205

10 "IT IS W H AT I T I S": D IA GNO SING T H E S P 5 00 †

• S&P 500 returns are power-law distributed with upside-downside asymmetry, and the diagnostics deem L2 methods including portfolio theory, factor analysis, GARCH, conditional variance, and stochastic volatility invalid.The diagnostic battery includes aggregation kurtosis, conditional expectations, maximum-to-sum plots, drawdowns, and records/extremes.Source: §10 · p. 207

10.1 paretianity and moments

• A single 1987 crash point makes the log-log survival plot consistent with Paretianity, while removing it could allow a lognormal or other fat-tailed mixed distribution; finite variance does not ensure informative sample moments.With an infinite third moment, the observed second moment almost never converges to its value, even though the second moment exists.Source: E CALIBRATING UNDER PARETIANITY · p. 204, §10.1 · pp. 208–209

10.2.1 Test 1: Kurtosis under Aggregation

• One-month kurtosis is not lower than daily kurtosis and shows no decline as data accumulate, contrary to the expected ∼n^-1 drop; the passage says this excludes simple stochastic-volatility formulations such as gamma variance.For α ≈ 3, the central limit mechanism may require n of the order of 10^6 to become acceptable.Source: §10.2.1 · pp. 209–210, Figure 10.2 · p. 209

10.2.2 Maximum Drawdowns

• S&P 500 drawdowns are Paretian, and empirical bootstrapping maps κ to a power law with α < 3 for negative returns.Source: §10.2.2 · p. 212, §10.2.3 · p. 213

10.2.4 Test 2: Excess Conditional Expectation

• Conditional expectation remains proportional to threshold K and does not drop for large K, which the passage says is incompatible with non-Paretian distributions; one observation accounts for >80% of kurtosis in 50 years.The cited passage treats this sample dependence of kurtosis as evidence that the fourth moment lacks stability.Source: §10.2.4 · p. 214, §10.2.5 · p. 214

10.2.6 Test 4: MS Plot

• Across 16,000 observations, the S&P 500 maximum-to-sum plot does not drop enough to indicate that the law of large numbers is functioning for the examined moment.Source: §10.2.6 · p. 215

10.2.7 Records and Extrema

• The records test indicates independence for negative-return extremes but dependence for positive ones; reshuffling makes independence of positive extrema implausible.The test's scope is limited to dependence at extremes.Source: Figure 10.15 · p. 216, §10.2.7 · p. 216
• Negative extrema produced 9 records versus 10.3 expected, close enough to fail to reject independence; subrecords likewise cannot reject independence, and EVT can be applied to negative observations.The subrecord analysis used 170 subperiods.Source: §10.2.7 · p. 217, Figure 10.16 · p. 217

10.2.8 Asymmetry right-left tail

• The left tail is considerably thicker than the right, a potential modeling challenge but not necessarily a problem for people interested in risk or option trading.Source: §10.2.8 · p. 219

10.3 conclusion: it is what it is

• SP500 returns are power-law distributed, making Modern Portfolio Theory and results depending on variance or correlation “totally bogus,” according to the chapter.The chapter also rejects “base rate crash” claims that people overestimate the probabilities of tail events.Source: §10.3 · p. 219

F THE PROBLEM WITH ECONOMETRICS

• Standard econometric results are not expected to replicate, and the methods are not dependable enough for risky decisions.The text says alternative techniques also cannot reliably measure rare events, though they can indicate whether a remote event is underpriced.Source: F THE PROBLEM WITH ECONOMETRICS · p. 221, f.1 performance of standard parametric… · p. 222
• One daily observation can dominate kurtosis: the 1987 U.S. stock-market crash accounted for 80% of kurtosis from 1952 to 2008, and estimates vary across periods.The text also reports that one April 2020 observation raised the kurtosis of U.S. weekly jobless-claim log changes from 8 to > 550.Source: F THE PROBLEM WITH ECONOMETRICS · p. 221, f.1 performance of standard parametric… · p. 223
• Methods based on squaring variables, including variance, correlation, and regression, are described as scientifically invalid except in rare cases where variables are bounded.The text also says p values have no meaning for economic and financial variables.Source: F THE PROBLEM WITH ECONOMETRICS · p. 222

f.2 performance of standard nonparametric risk estimators

• In higher dimensions, covariance-matrix sensitivity makes observed and conditional moments extremely unstable, while vector tail events become harder to calibrate as dimensions increase.The text separately warns that unstable variances make correlations problematic and undermine covariance-based portfolio methods.Source: f.2 performance of standard… · p. 225, Figure F.8 · p. 226
• The author proposes retaining extremes and building risk management around them, rather than treating outliers as analytically negligible.The passage compares this approach to safety practices that focus on catastrophe avoidance.Source: f.2 performance of standard… · pp. 225–226
• The author attributes continued use of unreliable risk measures partly to researchers and risk managers lacking personal exposure to the harms those measures can cause.The passage says those hurt by incorrect measures are other people.Source: f.2 performance of standard… · p. 226

G MACHINE LEARNING CONSIDERATIONS

• Weighted call options are presented as building blocks for nonlinear functions, while sigmoid sums cannot reproduce an unbounded payoff without infinitely many summands.The text describes sigmoid functions as bounded and says the final observation may match their flat portion.Source: G MACHINE LEARNING CONSIDERATIONS · pp. 227–228
• ReLU or ramp functions act like call options and are described as handling out-of-sample tail events better than sigmoid-based approaches through their extrapolation properties.The text links this property to there being no unexpected tail events in sample.Source: G MACHINE LEARNING CONSIDERATIONS · pp. 227–228

G.0.1 Calibration via Angles

• A weighted sum of call options can represent nonlinear univariate functions, with option weights calibrated implicitly from the function’s angles relative to Cartesian coordinates.The text contrasts this with sums of sigmoids, which it says map to tails less well.Source: §G.0.1 · p. 229, Summary · p. 230

11 PR O B A B IL I TY C A LI BR ATI ON UNDER FAT TA IL S ‡

• The chapter maps binary beliefs to continuous payoffs and argues that conflating them can mischaracterize decision-science findings.It says some psychological claims about overestimating tail probabilities are spurious, while binary forecasting ability does not necessarily correspond to performance, especially under nonlinearities.Source: §11 · p. 233
• The text says M4-M5 competition scores appear closer to real-world variables than the Brier score.It also says some nonlinear payoff functions are captured by machine learning or expressed in option contracts.Source: §11 · p. 233

11.1 continuous vs. discrete payoffs: definitions & comments

• A binary belief that the market is more likely to rise can coexist with a short position when a small chance of a large fall gives that exposure positive expected return.The example distinguishes forecasting direction from the payoff of a position.Source: §11.1 · p. 234

Commentary 11.1

• Under thick tails, real-world decisions require all moments, while probabilistic calibration estimates only the zeroth; higher moments can be explosive or infinite.Source: Commentary 11.1 · p. 234

11.1.1 Away from the Verbalistic

• Prediction markets may help gamblers but cannot hedge economic exposures: binary bets and open-ended continuous payoffs cannot properly replicate one another, especially under fat tails and parameter uncertainty.Probability of success can also mislead for thick-tailed sales: an out-of-the-money option may be attractive despite less than 1-in-1,000 odds.Source: §11.1.1 · pp. 237–238

11.1.2 There is no defined "collapse", "disaster", or "success" under fat tails

• Under fat tails, verbal labels such as “collapse” or “disaster” cannot define a binary event because its magnitude has no typical scale; continuous-payoff insurance or options can hedge it.Replicating an open-ended payoff with binary bets requires an infinite series.Source: §11.1.2 · pp. 238–239

11.2 spurious overestimation of tail probability in psychology

• For variable impacts with positive derivative, expected impact above a threshold and impact-at-threshold times exceedance probability are close only under thin tails, not fat tails.They are identical only when impact is constant above the threshold.Source: §11.2 · p. 239

11.2.1 Thin tails

• For Gaussian thin tails, the threshold-based quantity I2 corresponds to the inverse Mills ratio used in insurance.Source: §11.2.1 · p. 240

11.2.2 Fat tails

• The fat-tail analysis considers regular-variation distributions through their tail survival behavior for large thresholds.Source: §11.2.2 · p. 240

11.2.3 Conflations

• Psychological findings that agents overestimate small probabilities in single-payoff experiments may hold in those conditions, but generalizing them to real risks is dubious; probability overestimation can coexist with payoff underestimation.The section calls conflating expected payoff with probability-times-impact the ludic fallacy, following The Black Swan.Source: §11.2.3 · p. 241
• Default VaR can differ substantially from actual expected loss because it uses default probability times expected recovery, while collateral value conditional on default may be much lower.The passage says collateral values fell to about one-third of expected value after 2007 foreclosures and argues faulty CVaR approximations may have contributed to the 2008 crisis.Source: §11.2.3 · p. 241
• For the Pareto 80/20 distribution (α = 1.16), the adjusted probability is more than 7 times the naive probability.The passage describes the adjustment as de-binarizing the event probability.Source: §11.2.3 · p. 242
• As lognormal uncertainty rises, the probability of success tends to 0 while the expected payoff tends to +∞.The example likens this divergence to a binary option losing value while a same-strike call rises, and to riskier ventures being less likely to succeed but more rewarding when they do.Source: §11.2.3 · p. 244

Remark 11.1: Distributional uncertainty

• Distributional uncertainty increases the gap between I1 and I2, with fatter Paretian tails making uncertainty about expected payoff worse.The discrepancy increases with uncertainty in volatility or tail index; for α > 1, the ratio is positive and rises markedly as α falls.Source: Remark 11.1: Distributional uncertainty · p. 244

11.3 calibration and miscalibration

• Calibration methods are highly flawed beyond narrow cases of purely binary payoffs, and generalizing them to other outcomes can mislead.The passage notes that calibration metrics such as the Brier score are thin-tailed even when the measured variable is fat-tailed.Source: §11.3 · pp. 244–245

11.4 scoring metrics

• Binary forecasts converge rapidly to a thin-tailed Gaussian, whereas tracking economic P/L for a Pareto exposure requires a considerably larger sample.The passage attributes the sample-size difference to how the law of large numbers operates across distributions; real-world P/L is measured under a survival condition.Source: §11.4 · pp. 245–246, Commentary 11.5 · p. 246

11.4.1 Deriving Distributions

• The tally of binary forecasts is asymptotically normal regardless of the distribution of the underlying variable.For finite n, the Brier score has moments of all orders regardless of that distribution, even without assuming independence of its summands.Source: §11.4.1 · p. 248, Theorem 3 · p. 248

Theorem 3

• For nonbinary outcomes under fat tails, absolute-deviation scores are vastly more efficient than the Brier score, even when the second moment exists.The passage states that an absolute deviation has the same probability class as the variable itself, while the Brier score is based on the second moment.Source: Theorem 3 · p. 249

11.5 non-verbalistic payoff functions/machine learning

• A nonlinear payoff function makes the probability-payoff conflation worse; event probability is an expected payoff only when the payoff is a Heaviside function.The section frames probability as an integration kernel within quantitative payoffs, rather than as an outcome in itself.Source: §11.5 · p. 249, Remark 11.5: Inseparability of… · p. 250

11.6 conclusion:

• Risk management should change the payoff function rather than seek “good forecasts,” and net performance—not calibration mistakes that do not harm or help survival—is what counts.The conclusion states that risk management should never operate in probability space.Source: Commentary 11.7 · p. 251, §11.6 · p. 252

11.7.2 Distribution of the Brier Score

• The probability integral transform yields a uniform variable regardless of the distribution of X, while the Brier score has thinner tails than a Gaussian because its kurtosis is below 3.The appendix also derives law-of-large-numbers and central-limit convergence for a Bernoulli tally.Source: §11.7.1 · p. 252, §11.7.2 · p. 253

12.1 main statements

• Science should explain properties rather than forecast single outcomes, and sound risk management should focus on extremes and tails rather than distributional averages or bulk.The claims are framed alongside warnings that single-variable forecasts in fat-tailed domains violate common sense and probability theory, and that absence of evidence can matter greatly.Source: §12.1 · p. 255

12.2 commentary

• Pandemics have extremely fat tails (α < 1), so much of the distribution’s body is noise even though its extreme-value properties are remarkably stable.Random variables with unstable sample moments may still have stable tail properties useful for inference and risk taking.Source: §12.2 · p. 256, Remark 12.1 · p. 256

Remark 12.1

• Under systemic survival risks, uncertainty calls for precaution and “evidence of no harm,” because evidence may follow rare impactful events rather than precede them.The passages argue that waiting for evidence of harm can leave action too late.Source: Remark 12.1 · p. 257

12.3 more technical commentary

• For power-law variables with α ≤ 1, the law of large numbers fails; heuristically, variables up to α <= 5 2 are not forecastable, and a Pareto 80/20 average needs 10^14 observations versus 30 for a Gaussian.The source gives the Pareto 80/20 tail exponent as α ≈1.13 and says this sample size is needed for the average to emulate the significance of the Gaussian estimate.Source: §12.3 · pp. 257–258
• Tail survival functions can be forecastable even when higher moments are intractable, but binary metrics such as the Brier score do not directly serve risk management for the underlying variables.The source says the Brier score is suited to binary survival functions, while finance and insurance use expected shortfalls for risk management.Source: §12.3 · p. 258

12.4 further commentary specific to pandemics

• Pandemic forecasts must account for decisions and their consequences, since interventions can alter outcomes and make forecast accuracy an inadequate measure.Forecasts can become targets that participants game, as described through Goodhart’s law.Source: §12.4 · p. 259, §12.3 · p. 259
• COVID point predictions that omit action-conditioned consequences are unreliable; Spain’s response roughly two weeks later than Greece’s coincided with two orders of magnitude more cases.The commentary says early action distinguished countries that limited the outbreak from those that continued to suffer, and urges analysis grounded in affected people’s experiences.Source: §12.4 · pp. 259–261

Remark 12.4: Scaling of probability

• Systemic contagion risk cannot be inferred from individual risk: COVID-19 spreads multiplicatively, unlike additive car-accident risk, making the comparison in and misleading [This volume].The passage says the distinction can lead to macroscopic errors.Source: §12.4 · pp. 261–262

Remark 12.5: Ruin Problems

• Pandemics are systemic risks that should be contained early, while existential-risk and ruin settings defeat naive cost-benefit analysis [318]; disease costs include service saturation and long-term morbidity.The text also rejects treating the economy and disease as independent trade-off categories and raises ethical objections to sacrificing older generations.Source: Remark 12.4: Scaling of probability · p. 262, Remark 12.5: Ruin Problems · p. 263

Remark 12.7: Domain Dependence

• Pandemic unemployment cannot simply be attributed to authorities’ mitigation: prevention is described as orders of magnitude cheaper than cure, while observed fatalities omit the unseen counterfactual.The commentary notes that local or governmental mitigation occurred in practically every affected location.Source: Remark 12.7: Domain Dependence · p. 264

Remark 12.8: Economic Fragility

• Extreme-value analysis is not selection bias: pandemic fatality tail estimates remain below 1 under resampling, and mitigation can target early growth through superspreader reduction, quarantine, and low-cost measures.The section says the tail properties remain invariant as knowledge grows; it also describes historical quarantine practices and mitigation’s role in buying time for better treatments.Source: Remark 12.8: Economic Fragility · pp. 264–266

13 THE PROBABILITY CONFLATION

• Expert judgment can miss tail risk, forecasting tournaments do not align with extreme-value methods, and tail expectation diverges further from its corresponding probability.The section presents this as a new result while responding to Tetlock et al. (2022).Source: §13 · p. 267

13.2 the fat tails problem

• The so-called masquerade problem conflates the partial tail impact expectation G_K with exceedance probability P_K, often treating probability as a constant impact.The setup defines K as a threshold and P_K as the probability that X exceeds K.Source: §13.2 · p. 268
• The additional term in the expectation formula can dominate the first term, particularly for heavy-tailed distributions.Source: §13.2 · p. 268

Remark 13.2: Probability Classes

• PK is always thin-tailed, whereas GK reaches higher moments and extrema, so PK forecasts cannot serve as practical proxies when GK is fat-tailed.As a bounded random variable, PK has an unconditional Uniform distribution and is often modeled conditionally with a Beta distribution; sums of such bets become rapidly Gaussian.Source: Remark 13.1: Moments · pp. 268, 270, Remark 13.2: Probability Classes · p. 270
• Risk decisions should consider full distributional properties, especially tail shape and the dynamics that generate it, rather than rely on point forecasts alone.The passage contrasts multiplicative effects, which cause thick tails, with additive effects, which tend toward benign stochastic outcomes.Source: Remark 13.2: Probability Classes · p. 270

13.3 a new result: value at risk fails under error propagation

• A small, controlled error in tail probability can produce an explosive or infinite error in the corresponding threshold, increasingly so for fatter-tailed distributions.The passage relates this error propagation to Value at Risk (VaR) and expected shortfall (CVaR).Source: §13.3 · p. 271, Remark 13.5: Error Propagation · p. 272

13.4 a rather unscientific challenge

• The authors reject a single track record as evidence for general claims about tail-risk hedging, noting that trading records can conceal tail risks.They compare relying on one record to anecdotal claims with n = 1.Source: §13.4 · p. 272

13.5 conclusion: some more evidence required

• The authors call for more rigorous evidence linking PK estimates to GK and question the practical benefit of refining a South China Sea conflict estimate from 15% to 17%.They conclude that they see neither theoretical nor practical benefits to the superforecasting enterprise as it stands.Source: §13.5 · p. 273

ELECTION PREDICTIONS AS MARTINGALES: AN ARBITRAGE AP P R OA C H ‡

• The election framework treats a forecast probability as a tradable binary-option price and requires revised estimates to satisfy martingale pricing to avoid intertemporal arbitrage.It constructs a bounded shadow process for the vote variable from an unbounded Brownian motion.Source: ELECTION PREDICTIONS AS MARTINGALES:… · p. 277
• Under high uncertainty, an arbitrage-valued election probability moves toward 50% and should resist large changes even when polls or other inputs vary significantly.The model also caps the forecast probability at the estimated vote proportion when that proportion lies below or above 0.5.Source: Figure 14.1 · p. 276, ELECTION PREDICTIONS AS MARTINGALES:… · p. 277, Notation · p. 277

14.1 the bachelier-style valuation

• The Bachelier-style valuation prices a binary threshold event using the survival function of the Normal distribution under the process for X.The main equation is presented as the option price, with no-arbitrage motivating the continuously made forecast’s martingale condition.Source: §14.1 · p. 281

14.2 bounded dual martingale process

• Sigmoidal transformations preserve option prices across dual processes, while a transformed process is a martingale only at its specified central starting value.For the stated transformations, X martingale implies Y martingale only when Y0 = 1/2, and Y martingale implies X martingale only when X0 = 0; the transformed volatility declines toward the boundaries and is approximated by a scaled y(1 − y).Source: §14.2 · pp. 282–283

14.0.2 Organization / 14.0.3 A Discussion on Risk Neutrality

• The model has E(Yt) = Y0; higher-moment expansions add little precision, and risk neutrality is adopted because risk premiums would be difficult to justify and would degrade the Brier score.The variance lacks a closed-form solution from the distribution of Y and can instead be estimated from X using the delta method; higher-order terms are immaterial given the low vote volatility and bounded sigmoid transformations.Source: §14.0.2 · pp. 279–280, §14.0.3 · p. 280, §14.2 · p. 282

14.3 relation to de finetti’s probability assessor

• Continuous-time arbitrage consistency links a forecaster’s evolving binary bets to the Brier score, requiring the bet price to remain a martingale.De Finetti formulated probabilistic assessment with a nonlinear loss function; the section also notes that election forecasters’ revisions are not generally subjected to martingale valuation.Source: §14.3 · pp. 283, 285

14.4 conclusion and comments

• Binary options reveal more about uncertainty than about the true estimate, and the process is heuristically extended to multiple candidates by making the final candidate the residual share.For more than two candidates, the section constructs successive bounded candidate processes and assigns the remaining probability mass to the nth candidate.Source: §14.4 · p. 285

Incorrect claims

• The cited paper does not assess FiveThirtyEight’s forecast record or mention it; its only 2016-election reference is illustrative.The passage says Clayton’s attribution confuses the scientific paper with Twitter debates.Source: Incorrect claims · p. 286

Incorrect arbitrage valuation

• Under the pricing framework, a binary option converges to 1/2 as future volatility tends to infinity, regardless of the underlying’s distance from the threshold.The argument distinguishes an option price, which is an expectation, from a probability distribution; it concerns future unrealized volatility given the underlying information.Source: Incorrect arbitrage valuation · p. 286

Arbitrage matters

• A bounded binary-option price process lets traders arbitrage excessively volatile tradable forecasts by selling when prices are high and buying when low.The section uses the price bound [0, 1] to derive limits on forecast volatility; the strategy applies when the forecaster is willing to trade on the forecast.Source: Arbitrage matters · p. 287
• Numerical probabilistic forecasts should be treated as choice prices with skin in the game; binary forecasting then follows arbitrage and derivative-pricing rules while allowing Bayesian updating.The argument says arbitrage pressures constrain forecasts toward 1/2, not that probabilities should equal 1/2; the final outcome is 1 or 0.Source: Arbitrage matters · p. 287

15 GINI ESTIMATION UNDER INFINITE VARIANCE ‡

• The chapter proposes correcting the nonparametric estimator’s small-sample bias using the distance between the mode and mean of its asymptotic distribution.Source: §15 · p. 291

15.1 introduction

• The analysis focuses on regularly varying fat-tailed data with finite mean and infinite variance, corresponding to tail index α ∈(1, 2).The authors warn that conventional nonparametric estimation can be unreliable for Gini-based assessments of economic inequality.Source: §15 · p. 291, §15.1 · p. 293
• For Pareto data with α = 1.1, the nonparametric estimator converges extremely slowly and monotonically toward the true Gini value, g = 0.8333.The chapter compares the nonparametric and maximum-likelihood estimators across sample sizes using Monte Carlo samples.Source: §15.1 · pp. 293–294

15.2 asymptotics under infinite variance

• A lemma provides a way to handle order-statistic sequences generated by fat-tailed random variables that are integrable only in L1.This addresses the lack of the L2 integrability typically required by standard approximations.Source: §15.2 · p. 295

Theorem 1

• With α ∈(1, 2), the nonparametric Gini estimator remains consistent but approaches its true value more slowly from below, with a right-skewed α-stable limit.The limit is skewed and fat-tailed rather than the usual symmetric Gaussian limit, explaining the downward bias under fat tails.Source: §15.1 · p. 294, Theorem 1 · pp. 297–298

15.3 the maximum likelihood estimator

• For a specified exponential-family model, plugging the maximum-likelihood estimate into the Gini index produces an asymptotically normal and efficient estimator.The chapter argues that this parametric approach can reach efficiency with fewer observations than the nonparametric method, while noting the risk of model misspecification.Source: §15.1 · p. 292, §15.3 · p. 298

15.4 a paretian illustration

• For Pareto observations with α ∈(1, 2), the chapter applies its asymptotic results to derive corollaries for the nonparametric Gini estimator.Source: §15.4 · pp. 298–299, Corollary 15.1 · p. 299

Corollary 15.1

• As α approaches 1, the nonparametric estimator’s distribution mode moves farther from its mean, producing small-sample bias absent from the MLE’s normal limit.Figure 15.3 compares the asymptotic estimator distributions across tail indices.Source: Corollary 15.1 · p. 300, Figure 15.3 · p. 300
• At α = 1.2, the nonparametric estimator needs at least 80 × 10^6 observations to match the MLE’s probability of exceeding the ±0.02 threshold with n = 100.The comparison uses tail probabilities of the centered estimators.Source: Corollary 15.1 · p. 301, Table 15.2 · p. 301

15.5 small sample correction

• The proposed correction shifts the nonparametric Gini estimate by the mode–mean distance of its α-stable limiting distribution, without requiring the true Gini value.The correction depends on the tail index and sample-size scale parameter; it decreases with sample size and tends to zero, so the corrected and original estimators become asymptotically equivalent.Source: §15.5 · pp. 301–303
• For α ∈ (1, 2), the corrected estimator has smaller absolute deviation in the simulations, especially for n ≤ 1000; the estimators coincide at α = 2.The simulated samples are Paretian, and the improvement decreases as sample size grows.Source: §15.5 · pp. 303–304, Figure 15.4 · p. 304

15.6 conclusions

• The chapter recommends parametric methods for suspected fat-tailed data; when those cannot be used, it proposes the correction but cautions that estimating its terms adds uncertainty.The correction’s quality also depends on how closely the finite-sample distribution matches its α-stable limit.Source: §15.5 · p. 303, §15.6 · pp. 304–305

Proof of Lemma 15.1

• The Lemma 15.1 proof bounds the empirical-distribution terms to show that the relevant error tends to zero, including after multiplication by n^(α−1)L0(n) for α ∈ (1, 2).Source: Proof of Lemma 15.1 · pp. 305–306

Proof of Theorem 1

• Theorem 1’s proof rewrites the order-statistic expression using i.i.d. terms Zi = (2F(Xi) − 1)Xi and establishes that Zi remains in the fat-tails domain of attraction when X is in it.The argument assumes α ∈ (1, 2) and then applies a central-limit argument for fat-tailed sums.Source: Proof of Theorem 1 · pp. 306–307
• The proof establishes Z ∈ DA(Sα) and concludes that the ordered sequence of Z(i) has the same weak limits.The argument uses the generalized central limit theorem, the result β = 1, and Slutsky’s theorem.Source: Proof of Theorem 1 · pp. 308–309

Proof of Theorem 2

• The proof shows that Ẑ shares Z’s α and slowly varying function L, retains β = 1, and yields a Gini-index weak limit characterized by an i.i.d. sequence.The limit is characterized by the i.i.d. sequence rather than the ordered sequence, and the α-stable law is closed under scaling by a constant.Source: Proof of Theorem 2 · p. 310
• The observed quantile-contribution estimator is downward biased, with bias increasing with tail fatness and remaining significant for very thick tails even in very large samples.The discussion identifies in-sample quantile contribution as a biased estimator for one-tailed fat-tailed variables bounded on the left.Source: §16.1 · p. 312

16.2.1 Bias and Convergence

• κ0.01 ≈ 0.657933, and for an α-Pareto distribution with α = 1.1, even 100 million observations leave the estimator severely biased.The authors report simulations across sample sizes and present the biases in Table 16.1.Source: §16.2.1 · p. 313, Table 16.1 · p. 314
• Simulations suggest b(q) = 1 across α and q, but the estimator’s bias and standard deviation converge slowly toward zero.For α farther from 1, bias falls more rapidly and becomes weak near 2 for constant α; the discussion focuses on the top 1% and low α values.Source: §16.2.1 · pp. 313–314
• Concavity makes the finite-sample estimator clip large deviations, creating an upper bound on measured concentration and contributing to aggregation problems.The estimator’s concavity above the threshold dominates the opposing effect below it; the effect strengthens with a fatter right tail.Source: §16.2.1 · pp. 314–315

16.3 an inequality about aggregating inequality

• Aggregating measurements across sub-samples increases estimation bias: broad-sample concentration appears higher than the average of local concentration measures.The passage frames this as an “inequality about inequality,” using European data aggregated across countries as an example.Source: §16.3 · p. 316
• Weighted averaging of subsample concentration measures by their total sums underestimates the full-sample concentration measure.The same concavity and bias apply to lognormal distributions and worsen with higher variance.Source: §16.3 · p. 316

16.4 mixed distributions for the tail exponent

• A standard error of 0.3 in α raises κq(α) by 0.25, because convexity makes uncertainty about the tail exponent increase estimated concentration.For uncertain exponents, the minimum across a range can be more appropriate than the probability-weighted average; an estimate near 3/2 can yield results like exponents closer to 1.Source: §16.4 · pp. 320–321, §16.6 · p. 322

16.6 conclusion and proper estimation of concentration

• In thick-tailed settings, larger wealth sums accompany higher measured inequality, while smaller units can show lower concentration and create an illusion of historical increase.Year-on-year changes in concentration require adjustment for estimation stochasticity and inconsistent sample sizes.Source: §16.4 · p. 321, §16.6 · p. 322

16.6.1 Robust methods and use of exhaustive data

• Direct centile estimates are tied to a specific population and do not aggregate, so even exhaustive data cannot support out-of-sample claims about true properties.For insurance, one year's accounting profits do not reflect stochastic economic status or future tail exposure.Source: §16.6.1 · pp. 322–323

16.6.2 How Should We Measure Concentration?

• Risk managers can reduce bias by using extrapolative, nonconcave measures such as CVaR, lower-bound tail exponents, or integrating over possible exponents.These adjustments avoid aggregation problems and account for parameter uncertainty, analogous to stochastic-volatility adjustments in option pricing.Source: §16.6.2 · p. 323

17.2 the dual distribution

• A smooth transformation maps the bounded variable to a dual variable with the apparent tail, allowing extreme value theory to estimate its tail and recover the original variable's moments.The transformation is smooth, unlike simple truncation, and the method can compute excess moments above a threshold.Source: §17 · p. 327, §17.2 · p. 328

17 SHADOW MOMENTS OF AP PA R E NT LY I N F I NI TE -ME AN PHENOMENA ‡

• A bounded heavy-tailed variable can appear to have infinite mean when observations miss its remote upper bound, although bounded support ensures all moments are finite.The proposed approach avoids strong parametric assumptions for this apparent-infinite-mean problem.Source: §17 · p. 327, §17.1 · p. 328

17.2 the dual distribution

• A parameter-free dual transform makes bounded Y resemble unbounded Z below the cap, enabling Fréchet-tail modeling of Z with generalized Pareto threshold methods.A bounded Y belongs to the Weibull, not Fréchet, class; the GPD shape parameter ξ governs moment existence.Source: §17.2 · pp. 328–330

17.3 back to y: the shadow mean (or population mean)

• Even when dual Z has infinite mean, finite upper bound H permits estimating Y’s shadow mean by transforming fitted GPD tail moments back to Y.The conditional mean above L* is obtained by plugging in tail estimates; when L*=L, it is the population mean.Source: §17.3 · pp. 331–332

Value-at-Risk and Expected Shortfall

• The inverse tail quantile yields conditional VaR above L*, while expected shortfall follows from the corresponding mean-excess function.Source: Value-at-Risk and Expected Shortfall · p. 332

17.4 comparison to other methods

• The proposed method is contrasted with hard truncation and an absorbing cap; parameter-dependent EVT truncation can create a density cliff and overstate expectation, especially for α<1.Source: §17.4 · p. 333

17.5 applications

• Applications include operational losses, capped reinsurance, violence, credit risk, environmental harm, and company size, each with a finite upper bound.Source: §17.5 · pp. 334–335

18 ON THE TAIL RISK OF VIOLENT CO NFL IC T ( W I T H P. C I RI LLO) ‡

• The violent-conflict study applies extreme-value methods to log-transformed data, retransforms under the casualty cap, and checks robustness with high/low sampling and jackknifing.Its stated focus is statistical methods for thick-tailed, unreliable data and bounded variables with local power-law behavior.Source: §18 · p. 337, §18.1 · p. 337

18.1 introduction/summary

• Estimated future mean casualties are ≈3× the past mean, while conflict risk shows no decline and large-conflict arrivals are memoryless.Across raw and rescaled data, casualties are power-law distributed; log-rescaled α=.4–.7 indicates an undefined mean, and conventional sample means severely understate the true mean.Source: §18.1 · pp. 338, 340

18.2.1 Results

• Peak-over-threshold analysis finds strong Pareto-tail behavior in raw and rescaled casualties; the rescaled-data GPD fit gives ξ≈1.88 ±.14.Source: §18.2.1 · p. 340
• Rescaled data fit a Lomax-style distribution with the same tail as the peaks-over-threshold fit and strong goodness of fit.The analysis fits Pareto II (Lomax) distributions at different casualty thresholds, with tail parameters derived from the GPD.Source: §18.2.1 · p. 340
• The maximum-likelihood or “true” mean is 3–4 times the observed mean, so sample averages understate conflict violence and later observations alone cannot establish a rise.The discussion also reports no statistically detectable temporal structure or autocorrelation in event intervals.Source: §18.2.1 · pp. 340–341, §18.2.2 · p. 341

18.3.1 Rescaling Method

• Population rescaling uses a smooth transformation to remove the theoretical upper bound, enabling power-law analysis and rescaling back while leaving very large values essentially unchanged.The supplied passages also describe conditioning as a less rigorous alternative, but provide only a formula fragment.Source: §18.3.1 · p. 342, §18.3.2 · p. 343

18.3.3 Reliability of Data and Effect on Tail Estimates

• Despite anecdotal and uncertain historical casualty estimates, Monte Carlo treatment of bounded values found the tail exponent stable; bundling or splitting Mongolian wars also left conclusions invariant.The event definition is difficult because historical naming is inconsistent and direct war deaths can be hard to separate from famine or other indirect deaths.Source: §18.3.3 · p. 344, §18.3.4 · p. 345

18.3.5 Missing Events

• Raising the minimum casualty threshold to 500,000 and randomly removing events, including 30%, did not materially change tail-focused findings, though missing Black Swan events could still fatten the tail.Survivorship bias was not modeled and was assumed negligible before 1960, although the risk became substantially higher afterward with weapons of mass destruction.Source: §18.3.5 · p. 346, §18.4 · p. 346, §18.3.6 · p. 346

18.4 data analysis

• Raw and rescaled casualty plots suggest opposite trends, but both readings are naive; the data show fat tails, motivating Peaks-over-Threshold analysis.The upward mean-excess pattern for actual casualties begins around 5,000 victims, with 50,000 suggested as a higher goodness-of-fit threshold.Source: §18.4 · p. 346, §18.4.1 · p. 347

18.4.2 Gaps in Series and Autocorrelation

• 101.58 years is the average interval between actual-estimate events with more than 10 million victims, making trend extrapolation unreliable for such extremes.The reported mean absolute deviation is 144.47 years; a WW2-scale event occurred only once in 2014 years of data.Source: Table 18.2 · p. 341, §18.4.2 · p. 349

18.4.3 Tail Analysis

• The study models exceedances above a 50,000-victim threshold with a generalized Pareto distribution, using maximum-likelihood estimates for actual and rescaled casualties.The threshold is described as the best choice, but the supplied passage does not provide parameter values.Source: Table 18.4 · p. 342, §18.4.3 · p. 349
• The GPD tail estimate ξ = 1.5886 implies no defined moments; its fit captures most observations above 50,000 victims but misses the largest events.Events with 5–10 million victims also appear more frequent than the fit predicts; removing the two largest events means the GPD hypothesis cannot be rejected at the 5% level.Source: §18.4.3 · p. 350, Figure 18.9 · p. 350, Figure 18.10 · p. 351

18.4.4 An Alternative View on Maxima

• With 21 annual-block maxima, the GEV estimate ξ > 2 places casualties in the Fréchet domain of attraction and indicates a very heavy tail with no moments.The last block covers 2001–2014.Source: §18.4.4 · p. 351

18.4.5 Full Data Analysis

• 3.1 × 10^7 is the estimated true mean for L = 10K, versus a sample average of 9.12 × 10^6; the true-to-observed mean ratio stays between 3 and 4.The truncated mean is about 1.8835 × 10^7, and 96% of conflicts above 10,000 victims fall below the mean, making the sample average a poor guide.Source: §18.4.5 · p. 352

18.5.1 Bootstrap for the GPD

• Across 100K bootstrap samples of 90% of observations, ξ remained positive, and the Table 18.4 estimates approximated the GPD shape parameters despite data imperfections.The passage reports this stability for raw and rescaled data.Source: §18.5.1 · p. 352, Figure 18.11 · p. 353, Figure 18.12 · p. 353

18.5.2 Perturbation Across Bounds of Estimates

• Across 100K estimate combinations, near-tail p-values retained similar significance and goodness of fit, while sample and alpha-derived MLE means were examined across permutations.The checks assess sensitivity to parameter estimates rather than establish a new p-value.Source: §18.5.2 · p. 353, Figure 18.14 · p. 354, Figure 18.15 · p. 354

18.6 conclusion: is the world more unsafe than it seems?

• A threefold rise in average violence would remain consistent with the fitted parameters, so neither such a rise nor a large event alone would establish a change.The conclusion also notes that ignoring survivorship bias understates risk; homicide was excluded by the 10,000-victim threshold and is described as having different dynamics.Source: §18.6 · pp. 354–355

Inter-arrival times

• For conflicts causing at least 10 million casualties, the mean waiting time is 136 years, with a 267-year mean absolute deviation, so 70 years of “Long Peace” cannot establish a trend or rule out near-term war.The authors argue that fat-tailed data require more evidence to claim a decline than a rise, and that their robustness analysis addresses substantial historical-data error.Source: H WHAT ARE THE CHANCES OF A THI R D WO… · p. 357, Foundational Principles · p. 358, Inter-arrival times · p. 359

19 TAIL RISK OF CONTAGIOUS DI S EA S ES

• Extreme-value analysis of pandemic fatalities over about 2,500 years finds an extremely fat-tailed distribution, indicating potentially existential risk and a need to account for tail risk in risk management.The authors argue that such tail risk can override decisions based on naive interpolations, expected averages, or compartmental epidemiological models.Source: §19.1 · p. 361, §19 · p. 361

19.1 introduction and policy implications

• The fatality maximum-to-sum ratio does not converge to 0 for p = 1, 2, 3, or 4, suggesting no finite moments and making naive sample averages and standard deviations meaningless for inference.The authors caution that Figure 19.1 does not establish infinite pandemic risk; their method can estimate large but finite risk.Source: §19.1 · p. 362

The tail wags the dog effect

• Policy should not compare fat-tailed pandemic fatalities with thin-tailed accident deaths or use thin-tailed models that underestimate tail risk.The authors say epidemiological models such as SIR can aid scientific discussion of the bulk, but precautionary risk management should focus on maxima and tail exposures; EVT is their candidate approach.Source: The tail wags the dog effect · p. 363

The dual distribution

• The real fatality distribution is bounded by world population, so an apparently unbounded tail in finite data can mislead; dual data enable EVT modeling while retaining that bound.Values observed far below the upper bound can look like an unbounded tail, even though the bounded real distribution cannot have infinite moments.Source: The dual distribution · p. 368

19.3 the dual tail via evt and the shadow mean

• The estimated overall pandemic death mean is 7M, versus the naive sample mean of 4.9M, after replacing the tail above 200K with a 20.1M shadow mean.The shadow mean exceeds the 13.9M sample tail mean; the fitted threshold has 34.7% of observations above it, with ξ = 1.62 and β = 1′174.7K.Source: §19.3 · p. 371

19.4 data reliability issues

• Across 10K dual-data perturbations of ±20%, ξ values were always above 1 and averaged 1.62 (SD 0.10); jackknife samples removing 1–7 events preserved the fat-tail risk conclusion.The authors report that tail estimates are robust to imprecise observations and that the jackknife results leave their main message unchanged.Source: §19.4 · pp. 371–372

20.1.2 Higher Order Integrals in the Standard Gaussian Case

• For any constant a > 0, recursive uncertainty makes moments explode and the second moment unbounded as N →∞, invalidating L2 distributions unless recursion is cut off a priori.The process multiplies variance by (1 + a^2) at each recursion; the text says even a .001% error rate can produce this outcome.Source: Discretization using nested series of… · p. 387
• Adding uncertainty layers makes the density approach a power-law shape while the first moment remains invariant; as a approaches 1, the tail exponent approaches 1 but remains >1.The first absolute moment is also reported to remain independent of a and N.Source: Discretization using nested series of… · p. 386, Convergence to Properties Similar to… · p. 387

20.1.3 Effect on Small Probabilities

• Increasing N produces power-law-style flattening in the exceedance-probability plot, and the analysis compares exceedance probabilities with those of a standard Gaussian.Source: §20.1.3 · p. 387, Figure 20.4 · p. 388

20.2.1 Regime 2-a;“Bleed” of Higher Order Error

• 1.28σ2 is the limiting second moment and 9.88σ4 the limiting fourth moment for λ=.9 and a(1)=.2; the fourth moment remains finite versus 3σ4 for a Gaussian.When higher-order error parameters decline, moments tend to be capped, with inherited tails coming from the lognormality of σ.Source: §20.2 · p. 388, §20.2.1 · p. 389

20.2.2 Regime 2-b; Second Method, a Non Multiplicative Error Rate

• The second, nonmultiplicative error-rate method has a very mild effect on moments.Source: §20.2.2 · p. 390

20.3 limit distribution

• Lognormal approximations work well when errors on errors are in constant proportion, and the limit distribution is lognormal under the right conditions [291].Source: §20.3 · p. 391

21 STOCHASTIC TAIL EXPONENT FOR ASYM ME T R I C P O W ER L AW S †

• Stochastic tail exponents bias means and higher moments under data uncertainty; for right-tailed variables finite moments increase with exponent variance, while those of left-asymmetric variables decrease.The bias persists under summation even when enough summands warrant convergence to a stable distribution; uncertainty in historical data also increases the true mean for capped power laws.Source: §21 · p. 393

21.1 background

• Just as options are convex to distribution scale, expectations can be convex to the power-law tail exponent; options far from the money increase in value with variance uncertainty.Source: §21.1 · p. 394

Proposition 21.2

• Under the stated conditions, E(X^p) is convex in α, so mean-preserving stochastic variation in α increases the expected pth moment.The text also states that tail conditional expectation E(X|X>K) converges to a constant as K increases; for class P, convexity extends when the expectation approximation has the specified form and ν is approximately linear in α.Source: Proposition 21.2 · pp. 395–396, §21.2.2 · p. 395, §21.2.3 · p. 396

21.3 sums of power laws

• For 1 < α < 2, mean convexity is invariant under summation of power-law variables, and the limiting stable distribution’s mean depends on α despite conventional parameterizations.For Pareto summands, the asymptotic mean is the scale times α/(α−1), independent of n.Source: §21.3 · p. 397

Remark 21.1: Preservation of Asymmetry

• A normalized sum from the one-tailed class P1 converges to an asymmetric stable distribution with β ≠ 0 when its expectation depends on α as specified.The section also notes that flipping distributions in subclasses P1 and P2 to negative values degrades the stochastic-α effect.Source: Remark 21.1: Preservation of Asymmetry · p. 398, §21.4 · p. 398

21.5 pareto distribution with lognormally distributed α

• For stochastic-α Pareto models, the section considers a shifted lognormal exponent with lower bound b ≥ 1 and a shifted gamma exponent α−1 with values above 1.For b = 1, it approximates the density by expanding α around its lower bound and integrating the terms.Source: §21.5 · p. 399, Assuming finite expectation for X’ and… · p. 399, Approximation of the Density · p. 399

21.7 the bounded power law in cirillo and taleb (2016)

• Around α = 1/2, expectation is extremely convex, so uncertainty about α tends to raise the true mean estimated for violence data.The bounded-power-law construction gives finite expectation for every positive α and was used in studies of violence and operational risk,.Source: §21.7 · pp. 400–401

21.8 additional comments

• Exponent-uncertainty bias can inform analyses with insufficient, unreliable, or forged data, and the convexity result extends to compound Poisson and Lévy processes,.For a Lévy process, the text proposes analysis through successive slice distributions or discretization.Source: §21.8 · p. 401

22 META -D IST R I B U T I ON OF P-VALUES AN D P- H A C K I N G ‡

• The chapter derives exact meta-distributions for p-values across identical trials and their minimum, with small-sample and limiting forms parameterized by median pM.The derivation includes the ensemble distribution for the sample-derived one-tailed paired t-test with unknown variance.Source: §22 · pp. 403–404, §22.1 · p. 405
• 75% of realizations for a true p-value of .05 fall below .05, while 60% for a true p-value of .12 fall below .05.For an observed p-value of .02, the true value is likely above .1 and possibly near .2; the distribution is highly skewed, and setting power offers little remedy without a markedly larger sample or a p-value at least one order lower.Source: §22 · pp. 403–404

Proposition 22.2

• For large n, φ(p; pM) has the stated limiting distribution, applicable to paired tests with known or assumed sample variance as the test converges to Gaussian.The limiting expression is given in Equation 22.2.Source: Proposition 22.2 · pp. 406–407

Proposition 22.3

• The limiting distribution yields the distribution of the minimum among m statistically identical p-values, providing a way to assess p-hacking across multiple trials.The text also describes numerical integration outside the limiting case to calculate expectations for different m.Source: Proposition 22.2 · p. 408, Proposition 22.3 · p. 409

22.3 application and conclusion

• A test treated as 5% confidence may require a p-value at least one order of magnitude smaller; close-margin significant results may fail replication, and power has the same problem.The passage says replication attempts, including the open science project, should account for procedural error and bias toward favorable results; it recommends lowering p-values or setting tests at higher levels, such as .99.Source: §22.3 · p. 410

i.1 case study: how the myopic loss aversion is misspecified

• Using ensemble probabilities and the law of large numbers to explain the equity premium misses investor non-ergodicity; positive market returns alone do not ensure positive investor expectations.The passage says investors need a path-scaling strategy such as a Kelly-style strategy or path-dependent dynamic hedging to obtain a positive expectation.Source: i.1 case study: how the myopic loss… · p. 413

Myopic loss aversion

• Second-order effects from stochastic variance can cancel the myopic-loss-aversion explanation of the equity premium, while more frequent performance sampling lowers prospect-theory utility.The variance illustration shifts probability into the tails without changing total variance; the text says this does not rule out effects of myopia, but it cannot explain the equity premium from the inside of the value function.Source: I SOME CONFUSIONS IN BEHAVIORAL ECO NO… · p. 413, Myopic loss aversion · p. 415

Time Preference Under Model Error

• Making the discount parameter stochastic can flatten future preferences and produce hyperbolic discounting, so the passage treats the pattern as a consequence of model uncertainty rather than an empirical pathology.The setup assumes geometric discounting; the gamma-distribution example says preferences become flat in the future regardless of how steep they are in the present.Source: Time Preference Under Model Error · pp. 416–417

23 LINDY AS DISTANCE FROM AN ABSORBING BARRIER

• In the absorbing-barrier model, zero drift produces a Lindy survival function, whereas any negative drift moves the process out of that class.The section characterizes Lindy behavior by conditional expected additional life increasing with elapsed time, corresponding to a power-law survival function.Source: §23 · p. 419

23.1 force of mortality

• For a power-law time variable, force of mortality declines; it is constant for borderline exponential distributions and should increase for thin-tailed classes.The power-law result is stated for a time variable with a power-law probability density; the thin-tail behavior is presented as expected.Source: §23.1 · p. 421

23.2 the process

• Cagliostro’s account suggests that experience can reduce danger and the hazard rate over time, making the survival of people who do not age power-law distributed.Source: §23.2 · p. 422

23.2.1 Process without a drift

• For driftless Brownian motion, the first-passage distribution is Lévy and has no moments, while its force of mortality is not Lindy in the pre-asymptotic zone.The reflection-principle argument does not extend to arithmetic Brownian motion with drift or geometric Brownian motion; the passage next uses a change of measure.Source: §23.2.1 · pp. 424–425

Remark 23.2

• Any negative drift makes the stopping-time distribution leave the power-law class and lose its Lindy attribute; positive drift can make the hit never occur.The passage says all moments are finite for negative drift and calls the positive-drift possibility “immortality.”Source: §23.2 · p. 421, Remark 23.2 · p. 426

23.3 some considerations on life expectancy and aging

• Supercentenarians’ mortality is consistent with an exponential distribution: their annual death probability is around .47 and remains flat.The passage also reports no survival differences at extreme ages across sex, lifestyle, genetic background, time period, or ages 110 and 115; lifestyle factors important earlier cease to play a role after 110.Source: §23.3 · pp. 426–427

24.1 bachelier not black-scholes

• Bachelier’s model prices options using the actuarial expectation of final payoffs rather than dynamic hedging, allowing any distribution.Source: §24.1 · p. 431

24 FIN A N C I A L T H E O RY’S FAI LU RES WI T H O PT I O N P RI CI NG †

• The passage frames Black–Scholes–Merton’s dynamic-hedging assumptions as highly idealized, and says power-law moments can prevent its portfolio equation from working.Higher-order terms can matter when the distribution has significant cubic or quartic contributions, particularly for strikes away from its center.Source: §24.1.1 · p. 433

24.1.1 Distortion from Idealization

• Black–Scholes–Merton’s portfolio equation requires finite moments, but power-law tails make every moment above the tail exponent α infinite.The approach also assumes frictionless trading, no price impact, known Gaussian returns with fixed parameters, and no scalable jumps.Source: §24.1.1 · p. 433

24.1.3 Failure: How Hedging Errors Can Be Prohibitive.

• Dynamic hedging in a power-law world removes no risk, and hedge errors can be especially consequential for strikes away from the money.Under fat tails, increasing hedge frequency does not guarantee lower risk; a missed hedge or gap can be disastrous.Source: §24.1.3 · p. 434, §25.4 · p. 441
• Put-call parity supports distribution-free option valuation by expected terminal payoff without dynamic hedging or complete markets, for distributions with finite first moments, including infinite-variance cases.The framework requires a tradable forward and uses put-call parity across strikes; the valuation measure’s mean matches the forward, which need not be risk-neutral.Source: §24.1.3 · p. 435, §25.1 · pp. 436–437

25.2.2 Derivations

• For a given maturity, a unique measure prices European puts and calls by expected terminal payoff; it may or may not be risk-neutral, and remains unique when the forward is not risk-neutral.In the latter case, the framework replaces the measure with another measure while option prices remain unique.Source: §25.2.2 · p. 439, §25.3 · p. 441

Lemma 25.1

• The lemmas establish that put and call valuations determine unique measures, that both options share the same measure, and that static arbitrage aligns it with the tradable forward’s risk-neutral measure.The uniqueness argument uses strike-spread valuations; equality on semi-closed intervals extends to Borel sets.Source: Lemma 25.1 · pp. 439–440, Lemma 25.2 · p. 440, Lemma 25.3 · p. 440

25.4 comment

• Adding strike prices constrains the recovered distribution between strikes, making gaps less consequential than in dynamic hedging and turning replication into an interpolation problem.The authors compare the approach to kernel smoothing across strikes and note that additional over-the-counter strikes can be added as needed.Source: §25.4 · pp. 441–442

26 OPTION TRADERS NEVER USE T H E B L ACK-S C H O LES FO RMULA ∗,‡

• The chapter argues that traders have long used adaptable, bottom-up heuristics rather than the Black-Scholes-Merton dynamic-hedging argument, drawing on historical trading methods and earlier formula work.It credits Edward Thorp with writing down and using the formula later called Black-Scholes-Merton, while describing traders’ methods as more robust to rare events.Source: §26 · p. 443, §26.1 · p. 444

26.2.1 Black-Scholes was an argument

• Practitioners, the section argues, use a formula called Black-Scholes-Merton without realizing it differs from their methods; among historical formulas, it is uniquely fragile to jumps and tail events.The passage identifies the formula with Black and Scholes, 1973, Merton, 1973.Source: §26 · p. 443, §26.2.1 · p. 444
• BSM’s dynamic-hedging argument depends on strong assumptions, including thin-tailed randomness and continuous trading, and can expose practitioners to blowups.The authors characterize the argument as a fragile theoretical construction and say their account will instead examine how traders actually priced and hedged options.Source: §26.2.1 · pp. 445–446

26.3 myth 1: traders did not price options before bsm

• Options were actively traded and handled by practitioners well before BSM, including sophisticated markets in the 1870s and early 1900s.De La Vega describes Dutch option trading in the 1600s, and Nelson’s account shows early-1900s traders pricing and hedging options rather than feeling helpless.Source: §26.3 · pp. 447–448, §26.4 · p. 450
• Put-call parity and option arbitrage let earlier traders remove the need for a risk premium in option valuation without BSM-style dynamic hedging.Bronzin’s 1908 formula used arbitrage principles, while the authors say Higgins, Nelson, and Reinach described parity and robust option-hedging practices before BSM.Source: §26.3 · p. 447, §26.4 · p. 450

26.4.1 Option formulas and Delta Hedging

• Boness (1964) derived a call-option formula identical to BSM’s 1973 formula, making BSM’s claimed novelty its derivation rather than the equation itself.Earlier formulas by Sprenkle (1961) also anticipated the equation; BSM’s method used continuous dynamic delta hedging or CAPM to remove dependence on expected return.Source: §26.4.1 · p. 451
• Static market-neutral delta hedging was described before BSM, including methods for options beyond at-the-money contracts.Thorp and Kassouf (1967) and a 1970 Arnold Bernhard & Co. publication described balanced hedges across strike or asset-price levels.Source: §26.4.1 · pp. 451–452

26.5 myth 2: traders today use black-scholes

• Traders produce prices compatible with other instruments and arbitrage restrictions, not valuations based on beliefs about future probabilities; the proposed theoretical option formula is not robust in practice.The constraints include put-call parity and nonnegative spreads; option holding time is stochastic.Source: §26.5 · pp. 452–453, §26.4.1 · p. 452

26.5.1 When do we value?

• Valuation would be necessary in a hypothetical market without offsetting trades or a forward, but the authors say meaningful option trading requires an offsetting book.They argue that without offsetting trades, positions would remain negligible because dynamic hedging is not possible.Source: §26.5.1 · p. 453

26.6 on the mathematical impossibility of dynamic hedging

• Scale-free tails make dynamic hedging require finite moments that may not exist; higher-order terms can dominate, and replication risk remains stochastic rather than shrinking as Black–Scholes predicts.The authors argue that preasymptotic tail behavior and limited long-horizon data can obscure this problem, while dynamic hedging depends on high-frequency return properties.Source: §26.6 · pp. 453–455

26.6.1 The (confusing) Robustness of the Gaussian

• The volatility smile can fit option prices but conflicts with Black–Scholes–Merton; the authors describe the volatility surface as a map, not a territory, rather than evidence for that theory.They distinguish fitting a surface consistent with arbitrage-free prices from a robust hedging method.Source: §26.6.1 · p. 455

26.6.2 Order Flow and Options

• When traders hedge options with other options, pricing is largely demand- and supply-based; out-of-the-money prices do not estimate rare-event odds, and implied distributions express Dutch-book compatibility.This contrasts with the Black–Scholes–Merton idealization in which dynamic delta hedging makes demand and supply irrelevant to option prices.Source: §26.6.2 · p. 456

26.6.3 Bachelier-Thorp

• The authors call the approach Bachelier–Thorp: earlier formulas accommodate scale-free distributions, while dynamic hedging is not central to option pricing.They note that Black–Scholes–Merton was rarely used by traders before the 1980s, and that traders priced options from butterflies and conversions.Source: §26.6.3 · pp. 456–457

27 OPTION PRICING UNDER POWER LA WS : A R O B U ST H EUR I STI C ∗,‡

• The power-law heuristic extends a tail option price across strikes beyond the Karamata point using the tail index α; an anchor price and α suffice, without requiring finite variance.The first approach says other distribution parameters are eliminated; the stated moment condition is α > 1, corresponding to a finite first moment.Source: §27 · p. 459, §27.2.1 · p. 461, Remark 27.3 · p. 462

Theorem 7: Log returns

• A price formulation based on S in the regular-variation class does not imply that its log return is in that class; the geometric-return approach says nearly all distribution information is embedded in l.The tail index and Karamata constant depend on whether the underlying price or its geometric return is modeled.Source: §27.2.1 · p. 461, Theorem 7: Log returns · p. 462, §27.2.2 · p. 462

27.5 comments

• Market put prices fit a power law with thinner tails than the market-matched model; claims of tail overpricing from thin-tailed models are therefore misspecified.The comparison is relative to a closer-to-the-money strike, and the put-pricing derivation eliminates l and scaling λ.Source: Figure 27.3 · p. 463, §27.3 · p. 464, §27.5 · p. 466

27.4 arbitrage boundaries

• Adjacent-strike call prices can permit arbitrage if the call-spread boundary is violated, and that boundary imposes a lower bound on tail index α.The text says there is no arbitrage for strikes above the baseline strike K1.Source: §27.4 · p. 465

28.1 conflation of second and fourth moments

• The VIX primarily reflects the distribution’s center and second moment, not tail events or the fourth moment; tail bets instead use out-of-the-money options.The text describes buying tail options while selling around-the-money options to reduce second-moment exposure.Source: §28.1 · pp. 467–468

28.2 missing jensen’s inequality in analyzing option returns

• A 5-standard-deviation option can gain 16 times in value when implied volatility rises by 4, showing why VIX-based return estimates miss option convexity.The passage also reports 144 times for a 10-standard-deviation option and around 210,000 times for a 20-standard-deviation option, and says the example excludes in-the-money payoff.Source: §28.2 · pp. 468–469

28.3 the inseparability of insurance and insured

• Insurance costs must be assessed alongside the returns from the portfolio it enables; the market rose more than 100% over the period discussed.The author says an investor might otherwise have stayed out of the market or not added funds.Source: §28.3 · p. 469

28.4 the necessity of a numéraire in finance

• Without a numéraire—a unit to which other units relate—there is no probability measure for quantitative finance, the authors argue.They present the numéraire as the baseline for relating economic variables and applying the expectation operator.Source: §28.4 · p. 470

28.5 appendix (betting on tails of distribution)

• Fourth-moment strategies can use out-of-the-money options or calendars; the described trades benefit from fat tails and high peaks while having close to flat modified vega.One method buys 60-day options and sells 20-day options on 80% of the amount.Source: §28.5 · pp. 470–471

29 PO RT FO L I O S S H OU LD NE VER RELY ON CORRELATION

• MPT fails to control tail risk: volatility does not map to risk, and correlation is unreliable for fat-tailed or nonlinear stock-bond returns.The authors also argue that future returns and correlations are not observable, undermining ex ante optimal allocation.Source: §29 · p. 473, Remark 29.2: Second Fallacy of MPT · p. 474, correlation and risk parity · p. 474

Remark 29.3: Correlation and association

• Correlation is an unreliable measure of association for non-Gaussian variables or nonlinear dependence, and its observed swings violate portfolio theory’s assumptions.The passage says the random correlation structure makes the joint distribution non-elliptical, leaving correlation non-informative at best.Source: Remark 29.3: Correlation and association · pp. 474, 476
• Assuming correlation −0.5 rather than +0.5 produced about 1.5 times greater accumulated wealth, despite worse risk-adjusted return and volatility above the 10% target.The −0.5 assumption led to a 73% stock allocation versus 50% at +0.5, although the passage says there were no empirical grounds for assuming negative correlation.Source: Remark 29.3: Correlation and association · p. 476, Table 29.1 · p. 477
• The 3-year lookback dynamic portfolio performed no better, and much worse than portfolios based on assumptions that produced higher equity allocations.The comparison covers static portfolios based on +0.3 and −0.3 correlation assumptions.Source: Figure 5

What is the failure of Ellipticality?

• Market returns are described as non-elliptical, so stochastic correlations violate ellipticality’s single-covariance condition and leave MPT without mathematical justification for fat-tailed returns.Ellipticality is presented as necessary to portfolio theory’s results.Source: What is the failure of Ellipticality? · p. 480

Distinguishing between sample error and random correlation

• For 36-month rolling correlations, the constant-correlation sampling-error null yielded p-values of order 10^-17, supporting random correlation rather than sampling error.The Monte Carlo comparison accounts for dependence between overlapping windows.Source: Distinguishing between sample error… · p. 482, Figure 29.6 · p. 482

30.1 left tail risk as the central portfolio constraint

• The proposed framework hardwires robust left-tail loss constraints and maximizes entropy elsewhere; left-tail constraints can override conventional considerations, and a barbell portfolio emerges.The barbell combines maximal certainty or low risk in some holdings with maximal uncertainty in others.Source: §30.1 · p. 485

30.1.1 The Barbell as seen by E.T. Jaynes

• The barbell approach echoes Jaynes’s view that systems may move toward increasing entropy subject to constraints imposed by nature and government.The passage presents Jaynes’s insight as a parallel to constraining what can be constrained and maximizing entropy elsewhere.Source: §30.1.1 · p. 486

30.2 revisiting the mean variance setting; 30.2.1 Analyzing the Constraints

• The proposed framework replaces variance with two left-tail VaR constraints and adds a constraint on portfolio performance or growth outside the danger zone; the VaR-feasible density set is convex.Source: §30.2 · p. 487, §30.2.1 · p. 487

30.3 revisiting the gaussian case

• Under Gaussian returns, the VaR constraints determine mean and variance; a positive mean requires a variance lower bound that rises with ϵ and sufficiently large risk tradeoffs.Source: Figure 30.1 · p. 488, §30.3 · pp. 488–489

30.3.1 A Mixture of Two Normals

• Two-normal mixtures can represent fat tails, and with λ=ϵ, µ1=ν−, and small component variances they approximate two spikes; a dynamic stop loss induces positive skew.Source: §30.3.1 · p. 489

30.4 maximum entropy

• The maximum-entropy extension uses VaR constraints plus additional constraints to model unknown returns, since VaR alone leaves the right tail unrestricted and admits no maximum-entropy distribution.Source: §30.4 · p. 490

30.4.1 Case A: Constraining the Global Mean; 30.4.2 Case B: Constraining the Absolute Mean

• The global-mean and absolute-mean cases provide maximum-entropy extensions; in Case A, the left tail decays faster as ν− approaches K and converges to a point mass at K.Source: §30.4.1 · p. 491, §30.4.2 · p. 492

30.4.3 Case C: Power Laws for the Right Tail

• For a power-law right tail, the maximum-entropy model adds a constraint on E(log(1 + |X|)|X > K), with its value determining the tail decay rate.Source: §30.4.3 · p. 493

30.4.4 Extension to a Multi-Period Setting: A Comment

• Across periods, Case A approaches Gaussian behavior, while Case C converges to Gaussian only for α≥2 and does so slowly.Source: §30.4.4 · p. 494, Figure 30.7 · p. 495

30.5 comments and conclusion

• The authors say stop-loss behavior matters more than portfolio composition and propose entropy maximization as uncertainty recognition, yielding a barbell portfolio under VaR and Expected Shortfall.Source: §30.5 · pp. 494–495

INDE X

• The index includes entries on heteroskedasticity, hidden tails, independence, the kappa metric, and large-deviation theory.It also lists the generalized Pareto distribution, higher-dimensional thick-tailedness, and the law of large numbers.Source: INDE X · p. 522
• Other entries cover Lindy effects, fat-tailed linear regression, location-scale families, and the Lucretius fallacy.The index also lists life expectancy, log-Pareto distributions, and the law of large numbers for higher moments.Source: INDE X · pp. 522–523
• The remaining entries include the Sharpe ratio, universal approximation, var der Wijk’s law, violence, von Neumann, and the Wigner semicircle distribution.Numéraire, shadow moment, and Shannon/Boltzmann-Gibbs entropy also appear in the index.Source: INDE X · p. 523