← Back

Risk & Uncertainty

·Bryan Lai
TechnicalRiskUncertaintyAntifragile

The world isn't the tidy, predictable place many models assume.

1. Fat Tails – Where Extremes Rule

  • Many crucial variables (market returns, pandemic deaths, wealth, war casualties) don't follow the Gaussian bell curve. Their distributions have "fat tails" (often Power Laws), meaning extreme events are far less rare and vastly more impactful than in thin-tailed "Mediocristan."
  • In fat-tailed "Extremistan," a single extreme observation can dominate the sum or average, rendering traditional statistical properties unstable. The mean you calculate from a sample might tell you almost nothing about the true, long-run expected value. The Law of Large Numbers and Central Limit Theorem converge impractically slowly or fail entirely – you need astronomical amounts of data for averages to stabilize. [1]

2. Non-Linearity – Response is Not Proportional

  • How systems respond to events (F(X)) is often non-linear and more critical than the event itself (X). F is like the payoff or consequence function. [2]
    • Fragile (Concave Response): Suffers increasingly from volatility/stress. Think of a coffee cup: small taps do nothing, a moderate tap shatters it (accelerated harm). Fragile systems hate variance. [3]
      • Example: Driving Speed & Trip Success
        • Imagine driving from New York to Boston.
        • X is the input: driving speed in km/h (the stressor).
        • F(X) is the consequence: the trip success score (the outcome of driving at speed X).
        • Scores might look like: 10 km/h -> low score; 100 km/h -> high score; 300 km/h -> score crashes (accident).
        • Excessive speed leads to disproportionately bad outcomes, making the function F(X) concave for high speeds.
        • If speed X fluctuates (e.g., 10 to 300 km/h), the average outcome E[F(X)] from varying speeds will be worse than the outcome from driving steadily at the average speed F(E[X]). The lesson: drive fast enough for a good score, but not so fast that the risk of a crash dominates. You need to reach the destination reliably.
    • Antifragile (Convex Response): Benefits increasingly from volatility/stress (within limits). Think muscles growing from lifting weights, options payoffs. Antifragile systems like variance. [4]
  • Averaging the input (X) is irrelevant when the output (F(X)) is non-linear. The sequence and distribution of inputs matter profoundly. Small, frequent stresses are vastly different from a single large shock. Non-linear functions can amplify noise and transform thin-tailed inputs into fat-tailed outputs.

3. Survival & Compounding – The Primacy of Not Going Bust

  • In processes unfolding over time (investing, life), survival is a prerequisite. Decisions must prioritize avoiding "ruin" – hitting an absorbing barrier from which recovery is impossible. Long-term success depends on compounded (geometric) growth, not average (arithmetic) returns. [5]
  • A single catastrophic loss destroys compounded returns. Path dependence is critical: the order of gains and losses drastically alters the final outcome and the risk of ruin encountered along the way (drawdowns). Optimizing the geometric mean naturally incorporates ruin avoidance.
  • "Risk aversion" is often a misunderstanding of the mathematical necessity of protecting the compounding process. Optimal strategy involves managing volatility and cutting tail losses, even if it means lower average returns.

4. Flawed Metrics & Illusions – The Tools Rebel

  • Standard metrics often fail or mislead outside their narrow, assumed domain (typically linear, thin-tailed). [6]

    • Std. Deviation/Variance: Distorted by outliers (squaring effect), poor measure for fat tails.

    • Simple Relationship Measures (like Correlation): Measures only linear association, breaks down when relationships are more complex. Conditional correlation is context-dependent.

    • Statistical Significance Measures (like P-Values / R-squared): P-values are stochastic, skewed, and easily gamed ("hacked"). R-squared can be misleading. They don't reliably confirm relationships, especially under model misspecification or data dredging.

  • These tools create an illusion of understanding and quantifiable risk. Acting on their outputs in inappropriate domains leads to errors like overestimating heritability (Twin Studies), misjudging portfolio risk (Ellipticality), or validating bogus science (IQ, Psychology Quackery).

5. Consistency & Arbitrage – The Discipline of Price

  • Forecasts and probabilities, when viewed as prices in a dynamic system (like betting odds or option prices), must adhere to mathematical consistency to prevent arbitrage (risk-free profit). [7]
  • Wildly fluctuating forecasts for bounded outcomes signal irrationality or lack of "skin in the game." The possibility of arbitrage enforces discipline and bounds forecast volatility based on underlying uncertainty. Uncertainty itself pulls probabilities towards maximum entropy.

6. The Unseen & Precaution – Dealing with Limited Knowledge

  • Empirical data (what we've seen) inherently underestimates tail risk (the Lucretius Problem). Absence of evidence is not evidence of absence when dealing with potentially catastrophic, irreversible harm. [8]
  • Relying solely on past data guarantees being surprised by the unprecedented extreme. For systems with potential systemic ruin (pandemics, ecological collapse, financial crises), the burden of proof must shift: demonstrate safety rather than demand proof of harm before acting cautiously.

Thoughts:

Instead of rejecting models outright, we should apply them carefully. Models are good approximates, but they are only that, approximates. The models themselves are not flawed, it's when people rely too much on them that's flawed.

Even understanding when models start to break down helps us better understand what they are modelling.

References & Further Reading

[1] Relying on concepts like standard deviation or variance becomes treacherous in fat tails. Policies based on average outcomes are prone to catastrophic failure because they ignore the disproportionate impact of the rare event. Focus shifts from the typical to the extreme.

[2] Understanding the shape of the response curve (F(X)) is crucial for risk assessment. Fragility isn't just weakness; it's specifically accelerated harm from volatility. Focus shifts from predicting the event (X) to understanding the response (F(X)).

[3] For Fragile (Concave) systems: Average(F(X)) < F(Average(X)). In the driving example: E[F(X)] < F(E[X]).

[4] For Antifragile (Convex) systems: Average(F(X)) > F(Average(X)).

[5] Focus shifts from maximizing average gain to maximizing long-term compounded growth via survival.

[6] Metrics are not reality. Understand their assumptions. Visualize. Be deeply skeptical of statistical "significance" without considering model validity, effect size, and potential for gaming. Focus shifts from metric values to the validity of the underlying model and assumptions.

[7] Dynamic consistency is a powerful test for forecast rationality. Treating forecasts as tradable instruments reveals flaws. Focus shifts from point prediction to the consistency and properties of the forecasting process over time. Mathematical consistency implies the Martingale property.

[8] Use statistical methods (like Extreme Value Theory) to extrapolate beyond the observed sample. Apply the Precautionary Principle rigorously when uncertainty is high and potential downsides are catastrophic and systemic. Focus shifts from reactive correction to proactive prevention of ruin.

Illustrations Referenced Above

(Corresponding to numbered sections)

  1. Fat Tails: Pandemic/War tail analysis, Bernoulli Simulation's tail extension, critiques of financial risk models, Mini-Lessons on Fat Tails & LLN.
  2. Non-Linearity: Fragility/Antifragility lectures, Kelly Criterion (bet sizing), Black-Scholes (gamma convexity), GMO critique (unforeseen system effects), Base Rate Fallacy.
  3. Survival & Compounding: Kelly/Shannon/Thorp video, Bernoulli Simulation, Drawdowns & Logs, Path Dependence, Bitcoin critique (value hinges on avoiding absorption).
  4. Flawed Metrics & Illusions: Mini-Lessons on StdDev/Correlation/P-Values/Metrics, IQ Swindle, Twin Studies critique, Psychology Quackery, Genetics Reports.
  5. Consistency & Arbitrage: Election Pricing, Black-Scholes derivations (no-arbitrage core), Bitcoin valuation arguments.
  6. The Unseen & Precaution: Hidden Moments, Bernoulli tail extension, GMO/Mao critique, COVID arguments (initial response, vaccine risk profile analysis), Base Rate Fallacy (ignoring prior probabilities).

Thanks to Yi Yao Tan for reading drafts of this.