Claims

Only 14% of men in the United States are 6 feet or taller, meaning a woman who insists on a 6-foot minimum is eliminating 86% of the male population from consideration, which becomes irrational filtering if that criterion hasn't worked out after years of dating.

Bitcoin's young age (12-13 years) means 25% of its history is being evaluated when someone analyzes three-year periods, making such analysis statistically weak and insufficient for trend characterization.

Epidemiological studies reporting correlations between protein and disease outcomes are unreliable as evidence of causation because people reporting precise protein amounts ('I ate 6 ounces of steak') are inherently different from people reporting vague carbohydrate amounts ('a bowl of pasta'), introducing measurement error bias that masks carbohydrate effects

Beef consumption in the U.S. has declined 40% from 1975 to present, yet heart disease, cancer, obesity, and diabetes have increased; therefore epidemiological claims blaming beef for modern disease are contradicted by actual consumption trends

A confidence interval in medical research (e.g., '10 points plus or minus 3 points with 95% confidence') means that if a study is repeated 100 times, approximately 95 of those runs will yield results within the stated range, not that there is 95% certainty about the point estimate.

When 10,000 people make market forecasts over a 3-6 month period, some will necessarily be correct due to pure randomness, but this success is attributed to forecasting skill rather than statistical inevitability.

In domains like price or contagion where there are no physical limits to the size of the variable and the process is multiplicative, you cannot rule out a thick-tail distribution and must assume fat tails, unlike bounded systems like height.

Every statistical problem needs a new class of estimators adapted to the problem; you cannot make general claims using Gaussian-derived tools, and this principle should guide statistical methodology rather than trying to force all problems into standardized frameworks.

The central disagreement between Taleb and the superforecasting project is binary versus continuous payoffs: binary options lose value when tails fatten (because probabilities of being above/below a threshold drop), whereas continuous payoffs benefit from tail fattening.

If you are using a thin-tail probability distribution, you can always be surprised by an outlier with respect to your distribution, but if you are using a fat-tail or extremist model, the reverse cannot be true—nothing can surprise you because a quiet period is entirely within statistical properties and so is a large deviation.

In Gaussian distributions, the expected deviation above 3 sigma is slightly more than 3 sigma, and as you move to higher sigmas the expected additional deviation shrinks, but in extremist time the scale stays the same—the expected life expectancy of a Gaussian-distributed entity at 100 years old is about 2 additional years, but for a Pareto distribution it scales the same way.

When you exponentiate a Gaussian distribution, you get a log-normal; when you exponentiate a Gamma or exponential, you get a Pareto; when you exponentiate a Pareto, you get a log-Pareto—each transformation moves the distribution to fatter tails.

The shadow mean is the true mean of the data-generating process, which is often substantially higher than the empirical sample mean for fat-tail distributions; for warfare, the shadow mean is approximately three times higher than what historical records show.

Correlation is not an objective measure but depends on the sample; the real issue in decision-making is the relationship between changes in two variables under your observation, and understanding this requires accounting for the payoff function—the Kelly Criterion or entropy-maximizing frameworks show that low correlations are mostly noise and 0.5 is nowhere near halfway between 0 and 1 in terms of utility.

Mutual information (an entropy measure) is far more informative than correlation for decision-making; a correlation of 0 can coexist with infinite mutual information, showing the fundamental inadequacy of Pearson correlation for understanding dependence.

The tail exponent (alpha) converges to its true value much faster and with lower standard error than the mean in fat-tail distributions; empirically, alpha estimates stabilize within a few observations while mean estimates scatter widely.

Standard deviation is often confused with mean absolute deviation; standard deviation is the square root of the average of squared deviations and lacks intuitive physical meaning, whereas mean absolute deviation is the average deviation and is more interpretable.

Hypothesis testing and standardized statistics are fundamentally different from probability theory, with hypothesis testing being a separate discipline—you cannot do non-standard statistics without understanding probability, which explains why most practitioners fail.

In social science, the metric correlation is widely misunderstood; a correlation of 0.5 is not halfway between zero and one because the effect size non-linearly scales with payoff; further, correlation is sub-additive in absolute terms, meaning you cannot sum correlations across subsamples to get the population correlation.

For extremist distributions, the tail exponent (alpha) can be estimated reliably with relatively small samples using inverse gamma distribution theory because the exponent itself follows a predictable distribution, whereas estimating the mean is nearly impossible due to tail bias.

Multiplicative processes (like wealth accumulation through compound returns or contagion dynamics) generate fat-tail distributions because exponentiating a thin-tail base distribution creates fatter tails; specifically, exponentiating a Gaussian yields a log-normal distribution, and exponentiating a gamma or exponential distribution yields power-law distributions.

For Gaussian distributions, the expected deviation above any threshold k-sigma shrinks as k increases (e.g., above 3-sigma is ~0.8 sigma, above 5-sigma is ~5.1 sigma), whereas for extremist distributions, the expected deviation stays proportionally constant and explosively scales in absolute terms, meaning old-age expectations become nonsensical in power-law domains.

The shadow mean for a fat-tailed domain is the true population mean that would be observed if you had infinite data; it differs from the sample mean because finite samples cannot capture tail events that have disproportionate impact on the expectation, creating a gap between observed mean and true mean.

The ratio of standard deviation to mean absolute deviation is the best indicator of distribution fatness; for a Gaussian it is approximately 1.25, but for power-law distributions with alpha below 2 it becomes infinite, providing a simple diagnostic for detecting tail behavior.

Standard deviation became the traditional measure of dispersion because Fisher proved it was more statistically efficient than mean absolute deviation under Gaussian assumptions, but this efficiency advantage does not hold under fat-tail conditions, yet the field persisted with standard deviation without accounting for domain-specific distributional assumptions.

Every statistical problem requires adapted estimators suited to the specific distributional context; practitioners who only know Gaussian-based regression and hypothesis testing cannot move to non-standard problems without developing entirely new estimation theory.

Probability is not an independent object but a kernel—a function that must integrate to one and cannot be meaningfully discussed in isolation from the payoff function; practitioners who treat probability and payoff separately fundamentally misunderstand probability theory.

The reverse law of large numbers states that when you average individual outcomes in thin-tailed domains, aggregated effects differ dramatically from individual effects; a medication might lower hypertension by 4 points on average, but the distribution of individual effects may be wide, making population-level claims inapplicable to individuals.

Mean absolute deviation is more intuitive than standard deviation because it represents the average deviation from the mean; standard deviation is the square root of the average of squared deviations and lacks physical intuition; for a process with all observations at zero and one observation at a million, standard deviation would be 500,000 while MAD is 50,000.

The critical distinction in designing decision frameworks is whether a domain exhibits thin-tail or fat-tail distributions, because thin-tail distributions can be surprised by outliers that violate their assumptions, whereas fat-tail distributions incorporate large deviations within their statistical properties, making the latter epistemic class mandatory unless there are compelling physical reasons to rule it out.

The specific tail exponent (alpha) that appears in financial markets, wars, pandemics, and other domains is often an emergent property of the underlying generating process rather than a free parameter; for example, the square-root market impact model plus power-law firm size distribution theoretically generates a cubic tail exponent, but Taleb disputes both the exponent value and the robustness of such theoretical predictions because empirical data is too noisy and domain-specific.

Applying mediocristan (Gaussian-based) statistical tools to extremistan phenomena like pandemics is mathematically invalid: mean and variance are uninformative for fat tails; prediction error is 'monstrous'; sampling error is very high; the law of large numbers works very slowly or not at all; 95-99.9% of observations fall below the mean, making average-based predictions useless for extreme events.

Most empirical data on human heights, IQ, and apple sizes clusters around an average value following a normal distribution, but many phenomena in nature follow power laws instead, where extreme values are far more common than normal distribution would predict.

The probability calculation reveals that the three classical lunar conjunctions appearing with maximum frequency in three separate batches has an improbability that would only occur once in 2,500,000 cases, making the result so remarkable it looks like deliberate arrangement, yet it is statistically within limits that allow it to be explained as accidental.

The central limit theorem and law of large numbers only hold when variance is bounded and finite, but in fat-tailed distributions like Pareto distributions these conditions fail, requiring sample sizes of 10^13 instead of 30 to achieve the same statistical stability.

Fat tails emerge from a Gaussian distribution when volatility itself is variable; by constructing a mixture of Gaussian distributions with different variance parameters, the resulting composite distribution exhibits heavier tails and higher shoulders than a single Gaussian.

Correlation is a meaningless metric outside of linear relationships; it is not transitive and exhibits paradoxes that make it unsuitable for causal reasoning, yet it is routinely misused in psychology and social science to claim causal relationships.

P-values are random variables themselves; observing a p-value of 0.12 means the true parameter could produce p-values ranging widely, so approximately 53% of repeated experiments could yield p < 0.05 even though the true p-value is 0.12, making thresholds like p < 0.05 arbitrary.

The central limit theorem ensures that the sum of many independent random variables converges to a Gaussian distribution, even if the underlying distribution is uniform or other shapes, and this convergence happens very quickly (by 3 samples for uniform distribution you see non-flat distribution, by 10 samples you are approaching Gaussian).

Pareto distribution (80/20 distribution) is characterized by having the property that 80% of values lie in 20% of the range, which implies that if you recurse it, 1% of people own 50% of wealth, and this is used to illustrate fat tails because the distribution has most values concentrated but with extreme tail events.

The average human has one breast and one testicle; using group averages to make claims about individuals or to derive prescriptions for individuals is a category error—it conflates collective properties with individual properties.

You can rule out probability distributions and rank them by explanatory power even without knowing the true distribution; if you observe a 10-sigma event and the Gaussian distribution would assign it probability 10^-100, then the true distribution is ~10^100 times more likely to be non-Gaussian than Gaussian; this approach (Popperian falsification applied to distributions) is more powerful than trying to estimate the distribution directly.

Human height is bounded (cannot be fat-tailed) for scientific reasons: humans need wombs for gestation, and if heights were fat-tailed, occasionally someone would be 1,000 kilometers tall, which is physically impossible; therefore, you can rule out fat-tailed distributions for human height using a priori scientific reasoning rather than just fitting distributions to data.

An odds ratio is the probability of something happening in one group divided by the probability of it happening in another group, and an odds ratio of 1.78 means there is a 78% increased chance of the outcome in the affected group (calculated as: odds ratio minus one equals percent increase).

Group differences in IQ are not evidence of individual differences: knowing that group A has lower mean IQ than group B tells you nothing about whether a specific individual from group A is smarter or dumber than a specific individual from group B, making group data irrelevant to individual assessment.

Birth year effects and test year effects in IQ data are mathematically confounded (birth year = test year minus age), so it is impossible to distinguish between birth cohort effects and period effects from raw trend data, though ancillary analyses of other variables can suggest which interpretation fits better.

The Black-White gap on standardized tests has substantial sample size limitations: WISC standardizations have 100-300 Black samples (in some cases 135 for children ages 2-6, less than 200 for Stanford-Binet), making strong inferences about group means unreliable because the samples are designed to represent national distributions, not to measure precise group differences.

The process of extracting meaningful conclusions from large datasets with many variables is not mechanistic: researchers can identify endless spurious correlations that are technically true but practically useless, and must use judgment to extract valid conclusions, which is where pragmatism intersects with statistics.

The blurriness of racial categories (fuzzy boundaries between groups) acts as noise in detecting genetic IQ differences, not as artificial signal; if genetic differences exist, they should survive this noisy categorization, whereas environmental-only explanations might be confounded by racial classification errors.

A 20% Chance Is Not Nothing

Trusting the most reputable polls that gave Trump roughly a 20% chance was not a failure of judgment, because a 20% chance is not nothing—spend a few minutes with dice and you'll see it comes up frequently, sometimes on the first roll.