Black swan

Black Swan

Overview

Taleb (2007): a black swan is (1) outside all prior expectations, (2) extreme impact, (3) obvious in hindsight only. Many domains (markets, careers, tech) are Extremistan (power-law / fat-tail), yet most models assume Mediocristan (Gaussian / thin-tail) — underestimating tail risk by orders of magnitude. The Turkey Problem: 1000 days of feeding creates confidence; day 1001 is Thanksgiving.

Composes with antifragile, probabilistic-thinking, inversion, first-principles.

When to Use

Use when: a risk model assumes normality in a fat-tailed domain; "never happened in N years" dismisses tail risk; strategy assumes stable environment; you're constructing a retrospective narrative; stress-testing against extreme scenarios; someone says "fat tails / Taleb / narrative fallacy / turkey problem"; a thesis rests on a one-directional trend like "AI demand can only go up," AI-capex payoff, or concentrated mega-cap / AI-bubble exposure.

Not when: domain is genuinely Mediocristan; "black swan" is being used to excuse a foreseeable planning failure.

Coaching Novices (Adaptive Front Door)

• Engine mode: user has a concrete case → run The Process directly.
• Coach mode: user is unfamiliar or has no concrete case → guide step by step.

In Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.

1. One-line: some domains have rare events that dominate everything (markets, careers, tech); models assuming "normal" distributions miss them — and we always invent stories afterward that make them look predictable.
2. Check fit: genuinely thin-tailed domain (heights, commutes)? Point elsewhere.
3. Elicit the real situation — what decision or system are we auditing?

[WAIT — do not advance until user responds]

4. One question at a time: is this Extremistan or Mediocristan? What's the tail-survival design? What narrative am I constructing post-hoc?

[WAIT — do not advance until user responds]

5. Close: name the specific tail-event preparation — not prediction — they've uncovered.

[WAIT — do not advance until user responds]

The Process

Step 1: Classify the domain (Extremistan vs Mediocristan)

Sample 30+ historical observations. Bell-shaped → Mediocristan. Power-law / long-tail → Extremistan. Specific test: does the largest observation dominate the sum of the others? If yes → Extremistan.

Step 2: Identify hidden Mediocristan assumptions in your Extremistan domain

For each assumption: does it imply "normal distribution"? Use "rarely happens" or "the average is X"? Ignore outliers dominating outcomes? If yes and domain is Extremistan — structurally vulnerable.

Step 3: Design for tail survival

Tail survival ≠ tail prediction. Options: Bounded downside (no single event breaks you), Optionality (many small bets, convex upside), Redundancy (multiple suppliers / revenue lines), Skin in the game (decision-makers bear tail consequences). See antifragile.

Step 4: Defend against the narrative fallacy

Resist "we should have seen it coming." Ask: what was the epistemic state of well-informed people the day before? The "obvious in retrospect" feeling is hindsight bias, not predictability. Design for the category, not the specific next event.

Step 5: Distinguish black swans from grey swans

• White swans: routine, well-modeled. Grey swans: rare but predictable (hurricane in a hurricane zone). Black swans: rare AND unpredictable.
• Grey swan failure = planning failure. Black swan loss = design failure. The distinction matters.

Output: Black Swan Audit

# Black Swan Audit: <system / decision>
## Domain: Extremistan / Mediocristan / mixed — Evidence: <…>
## Hidden Mediocristan assumptions: <list>
## Tail-survival design: bounded downside via <…> | optionality via <…> | redundancy via <…>
## Narrative-fallacy check: past event <…> | story told <…> | actual epistemic state <…>

→ Method in Action: Long-Term Capital Management Collapse, 1998 · Fukushima Daiichi Tsunami, 2011 → 2026 lens: Concentration risk in the AI trade (2023–2026)

Pack: Black Swan Domains

Domain	Extremistan evidence	Tail-survival design	Common failure
Finance / trading	single crash day dominates decades of returns; 1987 was 20+ sigma under Gaussian	leverage that survives correlations→1; no ruin-level position sizing	VaR under Gaussian assumptions; "diversification" across risks that converge in crises
Venture / startups	one fund-returner outweighs the rest of the portfolio combined	many small bets, convex upside; reserves for winners, not averages	modeling outcomes with expected values; concentrating on one "sure thing"
Infrastructure / engineering	worst recorded flood / quake / tsunami dominates all damage totals	design margin above the full historical record; backups with no common-cause failure	design basis anchored to recent-instrumental maximum (Fukushima's 5.7 m seawall)
Careers / creative work	a handful of hits, roles, or books produce most lifetime payoff	barbell: stable base + repeated cheap shots at unbounded upside	betting everything on one employer or project; reading a hot streak as a stable trend

Mediocristan (safe to use averages): human heights, daily caloric intake, commute times on normal days.

Contribute a pack for your domain — see the template at the repo root.

Applying It Well

• Goal is tail-survival design, not prediction — you cannot predict the specific next black swan.
• Most "diversified" portfolios are not: in crises, correlations jump toward 1. True diversification is across uncorrelated risks.
• Mathematical sophistication does not protect against domain misclassification (LTCM lesson).
• "This has worked for years" is the turkey's argument.

→ Primary sources: references/sources.md

Common Rationalizations

[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.

Fake move	Reality
[D] "It's never happened in N years, so it can't happen"	The turkey problem. N years of data is consistent with both impossible and "imminent."
[D] "We have sophisticated models"	Models that assume the wrong distribution are sophisticated in the wrong direction.
[D] Constructing a confident retrospective narrative	Hindsight bias. The actual epistemic state before the event was much less clear.
[D] "Insurance / hedges / risk management protect us"	If those mechanisms also use Mediocristan assumptions, they fail in the same event.
[D] Predicting the specific next black swan	Not possible. Predict the category; design for tail-event survival.
[D] Treating "black swan" as an excuse for poor planning	Grey swans get mislabeled as black swans by people who failed to plan.
[D] Diversification across correlated risks	True diversification is across uncorrelated risks. In crises, correlations jump toward 1.
→ Add [O] entries here after each real use — paste the actual failure pattern	What went wrong and why

Red Flags

• Risk model assumes Gaussian distribution in an Extremistan domain
• "Has never happened" used as evidence of impossibility
• Diversification claimed without analysis of correlation in crises
• Confident prediction of when the next tail event will occur
• Retrospective "obvious in hindsight" claims about past events
• Mathematical sophistication treated as protection against domain misclassification

Verification

• Domain classified (Extremistan / Mediocristan / mixed) with evidence
• Hidden Mediocristan assumptions identified
• Tail-survival design specified (bounded downside / optionality / redundancy / skin)
• Past "black swans" in domain checked for narrative-fallacy framing
• Grey swans distinguished from true black swans
• Plan does not depend on predicting specific tail events

---

Part of deciqAI Knowledge Skills — 233 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. See it run → https://www.deciqai.com/s/black-swan · Built by deciqAI · github.com/deciqAI · Contributions welcome.

Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/black-swan.json