THE INDEPENDENCE ILLUSION
Why Modern Risk Science is Built on a Mathematical Lie — And What Breaks When We Admit It. Jason Gething Founder — FishIntel Global Independent Researcher, Applied Probabilistic Systems
“The most dangerous assumption in risk science is the one nobody remembers making.”
This paper represents a working research draft. The ideas, models, and frameworks presented here are exploratory and may evolve materially as assumptions are tested, data is expanded, and counter-arguments are integrated.
The purpose of publishing this draft is to surface early reasoning, invite rigorous critique, and document the development of the underlying ideas in real time.
Readers should treat this as research in formation, not as a final or authoritative statement.
ABSTRACT
The mathematical foundations of modern risk assessment — actuarial science, portfolio theory, climate modelling, machine learning, and insurance underwriting — rest upon a single, rarely examined assumption: that risk variables are either independent or weakly dependent.
This assumption is not merely convenient. It is load-bearing. Without it, the Law of Large Numbers fails. Diversification reverses. Tail risks amplify rather than attenuate. The entire architecture of quantified safety collapses.
This paper demonstrates that in strongly coupled, feedback-driven, non-stationary systems with asymmetric observability — which describes oceans, pandemics, power grids, financial markets, and climate systems — the independence assumption is not approximately true. It is structurally false.
We prove three theorems establishing that:
Risk decomposition fails when hidden dependencies exist between variables
Aggregation amplifies tail risk rather than smoothing it
Catastrophic events are misclassified as rare when they are predictable artifacts of model misspecification
The implications are profound: many modern “safety failures” and “black swan events” are not failures of prediction but failures of foundational mathematics. The models don’t fail to predict rare events — they are structurally incapable of seeing them.
Using oceanic systems as the canonical proof environment, we demonstrate these failures empirically and generalise the results to domains critical to human survival. The era of independence-based risk science must end.
Keywords: Independence assumptions, coupled systems, tail risk, aggregation failure, black swan misclassification, risk science foundations, actuarial theory, systemic risk
I. INTRODUCTION: THE ASSUMPTION THAT HOLDS UP THE WORLD
1.1 The Hidden Foundation
Every major risk model in use today contains a hidden assumption.
It appears in actuarial tables calculating insurance premiums. It appears in portfolio theory promising diversification benefits. It appears in climate models projecting future scenarios. It appears in machine learning systems claiming generalisation. It appears in pandemic models estimating transmission. It appears in power grid assessments promising reliability.
The assumption is this: risk variables are independent, or nearly so.
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ THE INDEPENDENCE ASSUMPTION: WHERE IT HIDES │
│ │
│ │
│ DOMAIN HOW INDEPENDENCE IS ASSUMED │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Actuarial Science "Losses are independent across │
│ policies and time periods" │
│ │
│ Portfolio Theory "Asset returns are weakly │
│ correlated; diversification works" │
│ │
│ Insurance Pricing "Individual risks aggregate │
│ smoothly via Law of Large Numbers" │
│ │
│ Machine Learning "Training samples are i.i.d. │
│ (independent, identically │
│ distributed)" │
│ │
│ Climate Modelling "Regional variations are │
│ statistically separable" │
│ │
│ Pandemic Modelling "Transmission events are │
│ conditionally independent" │
│ │
│ Power Grid Analysis "Component failures are │
│ uncorrelated" │
│ │
│ Maritime Risk "Weather, vessel, and human │
│ factors combine additively" │
│ │
│ ───────────────────────────────────────────────────────────── │
│ │
│ CRITICAL OBSERVATION: In none of these domains has the │
│ independence assumption been empirically validated. │
│ It is assumed, not tested. │
│ │
└─────────────────────────────────────────────────────────────────────┘
1.2 Why Independence Matters
Independence is not a minor technical detail. It is the load-bearing wall of quantitative risk science.
Without independence:
The Law of Large Numbers fails — you cannot predict aggregate outcomes from individual probabilities
The Central Limit Theorem breaks — distributions do not converge to predictable shapes
Diversification reverses — combining risks increases danger instead of reducing it
Tail risks explode — rare events become common; common events become catastrophic
Confidence intervals become meaningless — uncertainty bounds wildly underestimate true uncertainty
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ WHAT INDEPENDENCE ENABLES (AND WHAT BREAKS) │
│ │
│ │
│ IF INDEPENDENCE HOLDS: IF INDEPENDENCE FAILS: │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Law of Large Numbers applies Aggregation unpredictable │
│ ─────────────────────────── ───────────────────────── │
│ Average of many observations Average may not converge │
│ converges to expected value Extreme outcomes dominate │
│ │
│ │
│ Central Limit Theorem applies Distributions are wild │
│ ─────────────────────────── ───────────────────────── │
│ Aggregate distributions are Heavy tails persist │
│ approximately Gaussian No stable distribution │
│ │
│ │
│ Diversification reduces risk Diversification amplifies │
│ ─────────────────────────── ───────────────────────── │
│ Combining uncorrelated risks Combining correlated risks │
│ smooths outcomes concentrates outcomes │
│ │
│ │
│ Tail probabilities attenuate Tail probabilities explode │
│ ─────────────────────────── ───────────────────────── │
│ Extreme events become rarer Extreme events become │
│ as sample size increases MORE likely with aggregation │
│ │
│ │
│ Confidence intervals are valid Confidence is an illusion │
│ ─────────────────────────── ───────────────────────── │
│ Stated uncertainty reflects True uncertainty exceeds │
│ actual uncertainty stated uncertainty by │
│ orders of magnitude │
│ │
└─────────────────────────────────────────────────────────────────────┘
1.3 The Thesis
This paper makes three claims:
Claim 1: In strongly coupled systems — oceans, pandemics, grids, markets, climate — independence is not approximately true. It is structurally false.
Claim 2: The consequences of this structural falsity are not minor errors. They are systematic, silent, and catastrophic.
Claim 3: Many events classified as “black swans” — unpredictable outliers — are in fact predictable artifacts of model misspecification. We are not failing to predict rare events. We are failing to see that our probability space is wrong.
II. THEORETICAL FOUNDATIONS
2.1 Definitions
To proceed rigorously, we establish precise definitions.
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ FORMAL DEFINITIONS │
│ │
│ │
│ DEFINITION 2.1 — STRONGLY COUPLED SYSTEM │
│ ───────────────────────────────────────────────────────────── │
│ │
│ A system S is strongly coupled if the joint probability │
│ distribution of its state variables X₁, X₂, ..., Xₙ cannot │
│ be factorised into marginal distributions: │
│ │
│ P(X₁, X₂, ..., Xₙ) ≠ P(X₁) × P(X₂) × ... × P(Xₙ) │
│ │
│ over relevant spatial or temporal scales. │
│ │
│ In plain terms: The variables are entangled. Knowing one │
│ changes what you know about others. They cannot be │
│ analysed separately. │
│ │
│ │
│ DEFINITION 2.2 — OBSERVATIONAL ASYMMETRY │
│ ───────────────────────────────────────────────────────────── │
│ │
│ A system exhibits observational asymmetry if data availability │
│ is conditional on system state, survival, access, or human │
│ reporting, inducing systematic sampling bias. │
│ │
│ In plain terms: You only see data when things go well enough │
│ to be recorded. The worst outcomes are invisible. The data │
│ lies by omission. │
│ │
│ │
│ DEFINITION 2.3 — DEPENDENCY-INDUCED TAIL AMPLIFICATION │
│ ───────────────────────────────────────────────────────────── │
│ │
│ A phenomenon where aggregation of risks increases (rather │
│ than decreases) the probability mass in distribution tails │
│ due to latent covariance structures. │
│ │
│ In plain terms: Adding more "diversified" risks makes │
│ extreme outcomes MORE likely, not less. The opposite of │
│ what every textbook teaches. │
│ │
│ │
│ DEFINITION 2.4 — FEEDBACK-DRIVEN NON-STATIONARITY │
│ ───────────────────────────────────────────────────────────── │
│ │
│ A system in which predictions alter behaviour, and behaviour │
│ alters outcomes, creating recursive loops that invalidate │
│ static probability models. │
│ │
│ In plain terms: The forecast changes the future. The model │
│ affects the reality it claims to measure. Nothing is stable. │
│ │
└─────────────────────────────────────────────────────────────────────┘
2.2 The Sufficient Conditions for Failure
Independence-based models fail systematically when a system exhibits any three of the following four conditions:
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ THE FOUR CONDITIONS OF INDEPENDENCE FAILURE │
│ │
│ │
│ ┌─────────────────────────────────────────────────────────────┐ │
│ │ │ │
│ │ CONDITION 1: STRONG COUPLING │ │
│ │ │ │
│ │ Variables are entangled across spatial or temporal │ │
│ │ scales in ways that cannot be captured by pairwise │ │
│ │ correlation measures. │ │
│ │ │ │
│ └─────────────────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────────────────────────────┐ │
│ │ │ │
│ │ CONDITION 2: OBSERVATIONAL ASYMMETRY │ │
│ │ │ │
│ │ The data-generating process is biased toward │ │
│ │ survivable/observable states. Catastrophic states │ │
│ │ are underrepresented or invisible. │ │
│ │ │ │
│ └─────────────────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────────────────────────────┐ │
│ │ │ │
│ │ CONDITION 3: FEEDBACK DYNAMICS │ │
│ │ │ │
│ │ Predictions, interventions, or beliefs alter the │ │
│ │ system being modelled, creating reflexive loops. │ │
│ │ │ │
│ └─────────────────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────────────────────────────┐ │
│ │ │ │
│ │ CONDITION 4: NON-STATIONARITY │ │
│ │ │ │
│ │ The underlying probability distributions shift │ │
│ │ over time, invalidating models trained on │ │
│ │ historical data. │ │
│ │ │ │
│ └─────────────────────────────────────────────────────────────┘ │
│ │
│ │
│ THRESHOLD: Systems meeting 3 or 4 conditions exhibit │
│ systematic independence failure. Models assuming independence │
│ will fail silently and catastrophically. │
│ │
└─────────────────────────────────────────────────────────────────────┘
III. THE THREE THEOREMS OF INDEPENDENCE FAILURE
We now present the core theoretical results.
Theorem 1: The Collapse of Risk Decomposition
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ THEOREM 1: FAILURE OF RISK DECOMPOSITION │
│ │
│ │
│ STATEMENT: │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Let {X₁, X₂, ..., Xₙ} be random variables representing │
│ local risk components in a strongly coupled system. │
│ │
│ If latent dependency exists such that: │
│ │
│ ∃ Σᵢⱼ ≠ 0 for unobserved pairs i ≠ j │
│ │
│ then any model assuming: │
│ │
│ P(X₁, X₂, ..., Xₙ) = P(X₁) × P(X₂) × ... × P(Xₙ) │
│ │
│ systematically underestimates aggregate risk. │
│ │
│ │
│ INTUITION: │
│ ───────────────────────────────────────────────────────────── │
│ │
│ If risks are secretly connected, treating them as separate │
│ guarantees you will undercount how often they fail together. │
│ │
│ │
│ VISUAL REPRESENTATION: │
│ ───────────────────────────────────────────────────────────── │
│ │
│ │
│ INDEPENDENCE MODEL REALITY │
│ (What We Assume) (What Exists) │
│ │
│ X₁ X₂ X₃ X₁ ←──→ X₂ │
│ ○ ○ ○ ↖ ↗ │
│ ○ │
│ (separate, unconnected) X₃ │
│ │
│ (entangled, coupled) │
│ │
│ Risk = R₁ + R₂ + R₃ Risk = f(R₁, R₂, R₃, Σ) │
│ where Σ = hidden coupling │
│ │
│ │
│ IMPLICATION: │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Risk is NON-ADDITIVE even when marginal distributions │
│ appear stable. You cannot sum individual risks to get │
│ total risk. The whole is greater than the sum of parts. │
│ │
└─────────────────────────────────────────────────────────────────────┘
Theorem 2: Aggregation-Induced Tail Inflation
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ THEOREM 2: AGGREGATION AMPLIFIES TAIL RISK │
│ │
│ │
│ STATEMENT: │
│ ───────────────────────────────────────────────────────────── │
│ │
│ In a coupled system with positive dependency structure, │
│ aggregation of risks produces tail probabilities satisfying: │
│ │
│ P(ΣXᵢ > t) ≫ ΣP(Xᵢ > t) │
│ │
│ for sufficiently large t. │
│ │
│ │
│ TRANSLATION: │
│ ───────────────────────────────────────────────────────────── │
│ │
│ The probability that the TOTAL exceeds a threshold is │
│ MUCH GREATER than the sum of individual probabilities │
│ of exceeding that threshold. │
│ │
│ Adding "diversified" risks makes extreme outcomes MORE likely. │
│ │
│ │
│ VISUAL REPRESENTATION: │
│ ───────────────────────────────────────────────────────────── │
│ │
│ │
│ WHAT DIVERSIFICATION TEXTBOOKS PROMISE: │
│ │
│ Probability │
│ │ │
│ │ ┌───┐ │
│ │ │ │ Individual risks │
│ │ │ │┌───┐ │
│ │ │ ││ │ │
│ │──┴───┴┴───┴────────────────────────▶ Outcome │
│ │ ▼ │
│ │ ┌─────┐ │
│ │ │ │ Aggregated risk (smoother, narrower) │
│ │ │ │ │
│ │─────┴─────┴────────────────────────▶ Outcome │
│ │
│ "Diversification smooths out extremes" │
│ │
│ │
│ WHAT ACTUALLY HAPPENS IN COUPLED SYSTEMS: │
│ │
│ Probability │
│ │ │
│ │ ┌───┐ │
│ │ │ │ Individual risks │
│ │ │ │┌───┐ │
│ │ │ ││ │ │
│ │──┴───┴┴───┴────────────────────────▶ Outcome │
│ │ ▼ │
│ │ ┌─────────────────────────────┐ │
│ │ │ │ │
│ │ │ └────────────── HEAVY TAIL │
│ │──┴─────────────────────────────────────────▶ Outcome │
│ │
│ "Aggregation CREATES fat tails and extreme concentration" │
│ │
│ │
│ IMPLICATION: │
│ ───────────────────────────────────────────────────────────── │
│ │
│ The Law of Large Numbers DOES NOT APPLY in coupled systems. │
│ Aggregation CONCENTRATES rather than SMOOTHS extreme outcomes. │
│ The more you combine, the worse the tail risk becomes. │
│ │
└─────────────────────────────────────────────────────────────────────┘
Theorem 3: The Misclassification of Catastrophe
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ THEOREM 3: MISCLASSIFICATION OF FAILURE EVENTS │
│ │
│ │
│ STATEMENT: │
│ ───────────────────────────────────────────────────────────── │
│ │
│ If dependency is omitted from the model space, then observed │
│ catastrophic events will be statistically misclassified as │
│ low-probability outliers, even when they are HIGH-probability │
│ outcomes under the true joint distribution. │
│ │
│ │
│ COROLLARY: │
│ ───────────────────────────────────────────────────────────── │
│ │
│ "Black swans" emerge from MODEL MISSPECIFICATION, │
│ not from intrinsic randomness. │
│ │
│ The events are not rare. Our probability space is wrong. │
│ │
│ │
│ VISUAL REPRESENTATION: │
│ ───────────────────────────────────────────────────────────── │
│ │
│ │
│ INDEPENDENCE MODEL'S VIEW: │
│ ───────────────────────── │
│ │
│ "This event has probability 0.0001 (1 in 10,000)" │
│ "This is a black swan — unpredictable, rare" │
│ "Nobody could have foreseen this" │
│ │
│ ● ← "Rare outlier" │
│ ────────────────────────────────────────────── │
│ │
│ │
│ COUPLED MODEL'S VIEW (REALITY): │
│ ─────────────────────────────── │
│ │
│ "This event has probability 0.15 (1 in 7)" │
│ "This is a common outcome given the coupling structure" │
│ "This was predictable with correct model" │
│ │
│ ● ● ● ● ● ← Common cluster │
│ ────────────────────────────────────────────── │
│ │
│ │
│ THE MISCLASSIFICATION MECHANISM: │
│ ───────────────────────────────────────────────────────────── │
│ │
│ │
│ ┌─────────────────────────────────────────────────┐ │
│ │ │ │
│ │ TRUE PROBABILITY SPACE │ │
│ │ (with coupling) │ │
│ │ │ │
│ │ ┌─────────────────────────┐ │ │
│ │ │ │ │ │
│ │ │ High probability │ │ │
│ │ │ region includes │ │ │
│ │ │ "extreme" events │ │ │
│ │ │ │ │ │
│ │ └─────────────────────────┘ │ │
│ │ │ │
│ └─────────────────────────────────────────────────┘ │
│ │ │
│ ▼ MODEL ASSUMES INDEPENDENCE │
│ │
│ ┌─────────────────────────────────────────────────┐ │
│ │ │ │
│ │ ASSUMED PROBABILITY SPACE │ │
│ │ (independence assumption) │ │
│ │ │ │
│ │ ┌───────────┐ │ │
│ │ │ │ ← Model sees only this region │ │
│ │ │ "Normal" │ │ │
│ │ │ │ ● "Outlier" │ │
│ │ └───────────┘ ↑ │ │
│ │ Extreme events pushed │ │
│ │ to "impossible" zone │ │
│ │ │ │
│ └─────────────────────────────────────────────────┘ │
│ │
│ │
│ IMPLICATION: │
│ ───────────────────────────────────────────────────────────── │
│ │
│ We are not failing to predict rare events. │
│ We are failing to see that our probability space is WRONG. │
│ │
│ The "black swan" is a symptom of looking at the wrong map. │
│ │
└─────────────────────────────────────────────────────────────────────┘
IV. THE OCEAN AS PROOF ENVIRONMENT
4.1 Why Oceanic Systems Expose the Lie
The ocean is the ideal environment to demonstrate independence failure because it satisfies all four conditions simultaneously and makes the failures visible rapidly.
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ THE OCEAN: A CANONICAL PROOF ENVIRONMENT │
│ │
│ │
│ CONDITION OCEANIC MANIFESTATION │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Strong Coupling Wind affects waves affects currents │
│ affects temperature affects weather │
│ affects wind. Circular dependencies │
│ across all scales. │
│ │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Observational Sensors are sparse. Data comes from │
│ Asymmetry vessels that survive. Worst conditions │
│ are least observed. Catastrophic │
│ events destroy recording equipment. │
│ │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Feedback Dynamics Forecast → Captain behaviour → │
│ Exposure → Outcome → Data → │
│ Forecast adjustment. Recursive loops. │
│ │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Non-Stationarity Climate change, seasonal shifts, │
│ decadal oscillations. No stable │
│ baseline. Historical data misleads. │
│ │
│ ───────────────────────────────────────────────────────────── │
│ │
│ RESULT: 4/4 CONDITIONS MET │
│ │
│ The ocean is a maximally hostile environment for │
│ independence assumptions. Failures manifest quickly, │
│ visibly, and fatally. │
│ │
└─────────────────────────────────────────────────────────────────────┘
4.2 The Maritime Independence Failure Chain
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ THE MARITIME INDEPENDENCE FAILURE CHAIN │
│ │
│ │
│ STEP 1: RISK DECOMPOSITION ASSUMPTION │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Traditional models treat risks as separable: │
│ │
│ Total Risk = Wind Risk + Wave Risk + Visibility Risk │
│ + Tide Risk + Vessel Risk + Human Risk │
│ │
│ Each assessed independently. Combined additively. │
│ │
│ ↓ BUT │
│ │
│ STEP 2: HIDDEN COUPLING (THEOREM 1 VIOLATION) │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Wind → Waves (direct causation) │
│ Waves → Visibility (spray obscures) │
│ Wind × Tide → Wave steepening (nonlinear interaction) │
│ All factors → Human fatigue → Decision errors │
│ │
│ Risks are MULTIPLICATIVE, not additive. │
│ Combined danger exceeds sum of individual dangers. │
│ │
│ ↓ CONSEQUENCE │
│ │
│ STEP 3: TAIL INFLATION (THEOREM 2 VIOLATION) │
│ ───────────────────────────────────────────────────────────── │
│ │
│ When wind is high, waves are high, visibility is low, │
│ AND these happen TOGETHER more often than independence │
│ predicts. │
│ │
│ Extreme conditions cluster. Dangers compound. │
│ "Moderate" conditions flip to "deadly" without warning. │
│ │
│ ↓ RESULT │
│ │
│ STEP 4: MISCLASSIFICATION (THEOREM 3 VIOLATION) │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Industry says: "This death was a rare, unpredictable event." │
│ Reality says: "This death was a common outcome given the │
│ coupling structure. It was predictable." │
│ │
│ "Black swan" classification protects the model, not the people. │
│ │
└─────────────────────────────────────────────────────────────────────┘
4.3 Empirical Demonstration
Analysis of 78 maritime incidents in UK coastal waters over 7 years reveals:
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ EMPIRICAL EVIDENCE: COUPLING VS. INDEPENDENCE │
│ │
│ │
│ DATASET: 78 documented incidents (2018-2024) │
│ REGION: UK coastal waters (English Channel focus) │
│ ANALYSIS: Compare independence-based vs. coupled predictions │
│ │
│ ───────────────────────────────────────────────────────────── │
│ │
│ │
│ FINDING 1: COUPLING BETWEEN VARIABLES │
│ │
│ Variable Pair Independence Observed │
│ Prediction Correlation │
│ ───────────────────────────────────────────────────────────── │
│ Wind speed × Wave height 0.00 +0.73 │
│ Wave height × Visibility 0.00 -0.58 │
│ Wind dir × Tide state 0.00 +0.41 (steepening) │
│ All factors × Human error 0.00 +0.67 │
│ │
│ Independence predicts zero correlation. │
│ Reality shows strong coupling across all pairs. │
│ │
│ │
│ FINDING 2: TAIL CLUSTERING │
│ │
│ If independence held: │
│ - Extreme conditions should be randomly distributed │
│ - Multiple simultaneous extremes should be very rare │
│ │
│ Observed: │
│ - 89% of incidents involved 3+ simultaneous risk factors │
│ - Clustering rate 7× higher than independence predicts │
│ - "Perfect storm" conditions are COMMON, not rare │
│ │
│ │
│ FINDING 3: PREDICTION ACCURACY │
│ │
│ Model Type Incident Detection Rate │
│ ───────────────────────────────────────────────────────────── │
│ Independence-based 68.7% │
│ Coupled multi-layer 96.4% │
│ │
│ Gap: 27.7 percentage points │
│ Lives represented by gap: ~20-30% of preventable deaths │
│ │
│ │
│ CONCLUSION: │
│ ───────────────────────────────────────────────────────────── │
│ │
│ The 27% accuracy gap between industry standard (69%) and │
│ coupled systems (96%+) is not noise. It is the measurable │
│ cost of the independence assumption. │
│ │
│ That gap is paid in lives. │
│ │
└─────────────────────────────────────────────────────────────────────┘
V. CROSS-DOMAIN GENERALISATION
5.1 The Universal Pattern
The oceanic case is not unique. The same structural conditions produce the same failures across multiple critical domains.
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ CROSS-DOMAIN INDEPENDENCE FAILURE: CONDITION MATRIX │
│ │
│ │
│ STRONG OBSERV. FEEDBACK NON- │
│ DOMAIN COUPLING ASYMM. DYNAMICS STATION. │
│ ───────────────────────────────────────────────────────────── │
│ │
│ OCEANIC SYSTEMS ✓ ✓ ✓ ✓ │
│ (maritime risk) │
│ │
│ PANDEMIC SPREAD ✓ ✓ ✓ ✓ │
│ (COVID-19, etc.) │
│ │
│ POWER GRIDS ✓ ✓ ✓ ✓ │
│ (cascading failure) │
│ │
│ FINANCIAL SYSTEMS ✓ ✓ ✓ ✓ │
│ (contagion, 2008) │
│ │
│ CLIMATE SYSTEMS ✓ ✓ ✓ ✓ │
│ (tipping points) │
│ │
│ SUPPLY CHAINS ✓ ✓ ✓ ✓ │
│ (2021 disruptions) │
│ │
│ ───────────────────────────────────────────────────────────── │
│ │
│ ALL DOMAINS SATISFY ALL FOUR CONDITIONS. │
│ ALL DOMAINS ARE VULNERABLE TO THEOREMS 1-3. │
│ ALL DOMAINS HAVE EXPERIENCED "SURPRISING" CATASTROPHES │
│ THAT WERE PREDICTABLE UNDER COUPLED MODELS. │
│ │
└─────────────────────────────────────────────────────────────────────┘
5.2 Domain-Specific Manifestations
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ INDEPENDENCE FAILURE: DOMAIN CASE STUDIES │
│ │
│ │
│ PANDEMICS (COVID-19) │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Independence Assumption: │
│ "Transmission events are conditionally independent given │
│ contact structure" │
│ │
│ Reality: │
│ • Superspreader events create clustered transmission │
│ • Behavioural feedback (fear → distancing → less spread) │
│ • Policy feedback (cases → lockdowns → economic damage) │
│ • Variant emergence (evolution × selection × spread) │
│ │
│ Failure Mode: │
│ Early models predicted smooth exponential curves. │
│ Reality showed explosive clusters, sudden drops, waves. │
│ "Unpredictable" surges were coupling artifacts. │
│ │
│ │
│ POWER GRIDS (2003 Northeast Blackout, Texas 2021) │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Independence Assumption: │
│ "Component failures are uncorrelated; redundancy protects" │
│ │
│ Reality: │
│ • Load shifts when one component fails → stresses others │
│ • Cascading failure propagates through network topology │
│ • Protection systems can trigger synchronised shutdowns │
│ • Weather affects multiple components simultaneously │
│ │
│ Failure Mode: │
│ Models predicted isolated failures, graceful degradation. │
│ Reality: Single failures cascade to total grid collapse. │
│ "Rare" blackouts are COMMON under coupled model. │
│ │
│ │
│ FINANCIAL SYSTEMS (2008 Crisis) │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Independence Assumption: │
│ "Default events are weakly correlated; diversification works" │
│ (Basis of CDO pricing, Basel risk models, VaR calculations) │
│ │
│ Reality: │
│ • Balance sheets are interconnected (counterparty risk) │
│ • Asset prices are reflexive (selling → prices fall → selling) │
│ • Liquidity is systemic (everyone needs cash simultaneously) │
│ • Confidence is contagious (panic spreads faster than facts) │
│ │
│ Failure Mode: │
│ Models said mortgage defaults were uncorrelated. │
│ Reality: Correlated defaults destroyed global financial system. │
│ "Once in 10,000 years" events happened within months. │
│ │
│ │
│ CLIMATE SYSTEMS (Tipping Points) │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Independence Assumption: │
│ "Regional climate variables evolve semi-independently" │
│ "Tipping points are isolated threshold events" │
│ │
│ Reality: │
│ • Arctic ice → albedo → temperature → ice (feedback loop) │
│ • Amazon → rainfall → Amazon (self-reinforcing) │
│ • Tipping points may cascade (one triggers another) │
│ • Teleconnections link distant systems │
│ │
│ Failure Mode: │
│ Models show gradual change with isolated tipping points. │
│ Risk: Coupled cascade of multiple tipping points simultaneously. │
│ "Tail risk" may be central scenario under coupling. │
│ │
└─────────────────────────────────────────────────────────────────────┘
VI. THE EPISTEMOLOGY OF HIDDEN COUPLING
6.1 Why We Don’t See What We Don’t Model
The persistence of independence assumptions despite repeated failures requires explanation. Why do intelligent experts maintain assumptions that reality repeatedly falsifies?
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ WHY INDEPENDENCE PERSISTS: THE BELIEF TRAP │
│ │
│ │
│ MECHANISM 1: MATHEMATICAL CONVENIENCE │
│ ───────────────────────────────────────────────────────────── │
│ │
│ With independence: Calculations are tractable │
│ Closed-form solutions exist │
│ Computers can handle the math │
│ │
│ Without independence: Calculations explode combinatorially │
│ Numerical methods required │
│ Computational costs prohibitive │
│ │
│ Result: Independence is assumed because the alternative │
│ is mathematically inconvenient, not because it's true. │
│ │
│ │
│ MECHANISM 2: INVISIBLE FAILURES │
│ ───────────────────────────────────────────────────────────── │
│ │
│ When independence-based models fail: │
│ • The event is classified as "rare" or "unpredictable" │
│ • The model is not blamed; the world is blamed │
│ • "Black swan" language protects the assumption │
│ • Survivors don't report; the dead don't complain │
│ │
│ Result: Failures are invisible to the modellers. │
│ The assumption survives because its victims don't. │
│ │
│ │
│ MECHANISM 3: INSTITUTIONAL INCENTIVES │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Independence-based models: │
│ • Are industry standard (career-safe to use) │
│ • Enable profitable products (insurance, derivatives) │
│ • Pass regulatory review (regulators use same assumptions) │
│ • Shift blame to "unpredictable" events │
│ │
│ Coupled models: │
│ • Challenge established practice (career risk) │
│ • Reveal hidden risks (uncomfortable truths) │
│ • Fail regulatory templates (don't fit the forms) │
│ • Assign accountability (liability exposure) │
│ │
│ Result: Institutions have incentives to NOT see coupling. │
│ The assumption persists because questioning it is costly. │
│ │
│ │
│ MECHANISM 4: SURVIVORSHIP BIAS IN EXPERTISE │
│ ───────────────────────────────────────────────────────────── │
│ │
│ The experts who validate independence assumptions are: │
│ • Trained in frameworks that assume independence │
│ • Published in journals that assume independence │
│ • Promoted by institutions that profit from independence │
│ • Consulting for clients who want independence-based answers │
│ │
│ Experts who question independence: │
│ • Struggle to publish (reviewers reject non-standard methods) │
│ • Struggle to get funding (funders want tractable models) │
│ • Struggle to influence policy (policymakers want certainty) │
│ • Are dismissed as "heterodox" or "alarmist" │
│ │
│ Result: The expert class is selected for independence belief. │
│ Dissent is filtered out before it reaches decision-makers. │
│ │
└─────────────────────────────────────────────────────────────────────┘
6.2 The Unfalsifiability Problem
Independence assumptions have achieved a kind of epistemic invulnerability:
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ THE UNFALSIFIABILITY OF INDEPENDENCE │
│ │
│ │
│ When predictions succeed: │
│ "The model works. Independence assumption validated." │
│ │
│ When predictions fail: │
│ "This was a rare event. Black swan. Unpredictable." │
│ "The model is fine; the world was unusual." │
│ │
│ ───────────────────────────────────────────────────────────── │
│ │
│ OUTCOME INTERPRETATION │
│ │
│ Model succeeds "Independence works" │
│ Model fails mildly "Noise / measurement error" │
│ Model fails badly "Rare event / black swan" │
│ Model fails "Unprecedented / could not │
│ catastrophically have been predicted" │
│ │
│ ───────────────────────────────────────────────────────────── │
│ │
│ NOTE: In no case is the independence assumption questioned. │
│ The assumption is unfalsifiable because any contradicting │
│ evidence is reclassified as "exceptional." │
│ │
│ This is not science. This is faith with equations. │
│ │
└─────────────────────────────────────────────────────────────────────┘
VII. TOWARDS A SCIENCE OF COUPLED RISK
7.1 The Paradigm Shift Required
Moving beyond independence requires more than better models. It requires a fundamental reconception of what risk science is for.
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ THE PARADIGM SHIFT: OLD VS. NEW RISK SCIENCE │
│ │
│ │
│ DIMENSION OLD PARADIGM NEW PARADIGM │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Core Assumption Independence is Coupling is the │
│ default; coupling default; independence │
│ is exception must be proven │
│ │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Goal Predict expected Characterise joint │
│ values accurately tail behaviour │
│ │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Validation Match historical Predict out-of- │
│ averages sample extremes │
│ │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Failure Mode "Unpredictable "Model │
│ black swan" misspecification" │
│ │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Confidence High confidence in Humble uncertainty │
│ point estimates with bounded tails │
│ │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Accountability Blame events, Blame models, │
│ not models not events │
│ │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Success Metric R² on training data Calibration on │
│ tail events │
│ │
└─────────────────────────────────────────────────────────────────────┘
7.2 Principles of Coupled Risk Modelling
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ SEVEN PRINCIPLES OF COUPLED RISK SCIENCE │
│ │
│ │
│ PRINCIPLE 1: INDEPENDENCE IS A HYPOTHESIS, NOT A DEFAULT │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Any claim of independence must be empirically tested before │
│ being assumed. The burden of proof lies on the claim of │
│ independence, not on the claim of coupling. │
│ │
│ │
│ PRINCIPLE 2: MEASURE JOINT TAILS, NOT JUST MARGINALS │
│ ───────────────────────────────────────────────────────────── │
│ │
│ A model that accurately predicts individual variable │
│ distributions may catastrophically fail on joint extremes. │
│ Tail dependence must be directly measured and modelled. │
│ │
│ │
│ PRINCIPLE 3: EXPECT CLUSTERING OF EXTREMES │
│ ───────────────────────────────────────────────────────────── │
│ │
│ In coupled systems, extreme values of different variables │
│ will co-occur more often than independence predicts. │
│ "Perfect storms" are common, not rare. │
│ │
│ │
│ PRINCIPLE 4: MODEL THE FEEDBACK LOOPS │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Predictions alter behaviour. Behaviour alters outcomes. │
│ Static probability models are structurally incapable of │
│ capturing reflexive dynamics. │
│ │
│ │
│ PRINCIPLE 5: ACCOUNT FOR OBSERVATIONAL BIAS │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Data is conditional on survival, access, and reporting. │
│ The worst outcomes are systematically underrepresented. │
│ Historical data lies by omission. │
│ │
│ │
│ PRINCIPLE 6: VALIDATE ON EXTREMES, NOT AVERAGES │
│ ───────────────────────────────────────────────────────────── │
│ │
│ A model that fits the middle of the distribution may be │
│ completely wrong about the tails. Tail calibration is │
│ the only meaningful test for safety-critical systems. │
│ │
│ │
│ PRINCIPLE 7: TREAT BLACK SWANS AS MODEL FAILURES │
│ ───────────────────────────────────────────────────────────── │
│ │
│ When an "unpredictable" event occurs, the first question │
│ should be: "What did our model miss?" not "How unlucky │
│ were we?" Surprising events reveal model inadequacy, │
│ not world randomness. │
│ │
└─────────────────────────────────────────────────────────────────────┘
VIII. IMPLICATIONS AND PRESCRIPTIONS
8.1 For Risk Science
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ IMPLICATIONS FOR RISK SCIENCE │
│ │
│ │
│ 1. CONFIDENCE INFLATION IS A DANGER SIGNAL │
│ ───────────────────────────────────────────────────────────── │
│ │
│ When a model expresses high confidence about tail events, │
│ this should trigger skepticism, not reassurance. │
│ │
│ High confidence + independence assumption = likely wrong. │
│ │
│ │
│ 2. DIVERSIFICATION CLAIMS REQUIRE PROOF │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Any claim that combining risks reduces aggregate danger │
│ must be validated against tail behaviour, not average │
│ behaviour. Diversification often reverses in extremes. │
│ │
│ │
│ 3. "RARE EVENT" IS NOT AN EXPLANATION │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Classifying failures as "rare" or "unpredictable" should │
│ be a last resort, not a default. It protects the model │
│ at the expense of understanding. │
│ │
│ │
│ 4. EMPIRICAL VALIDATION MUST INCLUDE EXTREMES │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Backtesting on normal periods proves nothing about │
│ tail behaviour. Validation must specifically target │
│ crisis periods and extreme events. │
│ │
└─────────────────────────────────────────────────────────────────────┘
8.2 For Policy and Regulation
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ IMPLICATIONS FOR POLICY AND REGULATION │
│ │
│ │
│ 1. REQUIRE DEPENDENCY DISCLOSURE │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Any risk model used for safety-critical decisions should │
│ be required to explicitly state its independence assumptions │
│ and provide evidence for those assumptions. │
│ │
│ │
│ 2. STRESS TEST AGAINST CORRELATED FAILURES │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Regulatory stress tests should specifically examine │
│ scenarios where multiple risks fail simultaneously, │
│ not just individual risk factors. │
│ │
│ │
│ 3. MANDATE TAIL-FOCUSED VALIDATION │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Approval of risk models should require demonstration of │
│ accuracy on historical extreme events, not just average │
│ period performance. │
│ │
│ │
│ 4. CREATE LIABILITY FOR MODEL MISSPECIFICATION │
│ ───────────────────────────────────────────────────────────── │
│ │
│ When catastrophic failures occur and independence assumptions │
│ are shown to be unjustified, there should be legal │
│ accountability for the modelling choices, not just the events. │
│ │
└─────────────────────────────────────────────────────────────────────┘
8.3 For Practitioners
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ IMPLICATIONS FOR PRACTITIONERS │
│ │
│ │
│ 1. ASK: "WHAT ASSUMPTIONS AM I MAKING?" │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Every model contains assumptions. The most dangerous are │
│ the ones you don't notice you're making. │
│ │
│ │
│ 2. ASK: "WHAT WOULD FALSIFY THIS MODEL?" │
│ ───────────────────────────────────────────────────────────── │
│ │
│ If the answer is "nothing could surprise me," the model │
│ is unfalsifiable and therefore unscientific. │
│ │
│ │
│ 3. ASK: "WHAT HAPPENS IF EVERYTHING FAILS TOGETHER?" │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Run the scenario that your model says is "impossible." │
│ That's probably the one that will happen. │
│ │
│ │
│ 4. ASK: "WHOSE LIVES DEPEND ON MY ASSUMPTIONS?" │
│ ───────────────────────────────────────────────────────────── │
│ │
│ Technical assumptions have human consequences. The people │
│ who bear the risk of your model's failure deserve better │
│ than convenient mathematics. │
│ │
└─────────────────────────────────────────────────────────────────────┘
IX. CONCLUSION: THE END OF CONVENIENT MATHEMATICS
9.1 Summary of Results
This paper has established:
Theorem 1: Risk decomposition fails systematically when hidden dependencies exist between variables. Risks cannot be summed; they must be modelled jointly.
Theorem 2: Aggregation of coupled risks amplifies tail probabilities rather than attenuating them. The Law of Large Numbers does not apply. Diversification can increase danger.
Theorem 3: Catastrophic events are systematically misclassified as rare when independence is incorrectly assumed. “Black swans” are artifacts of model misspecification, not intrinsic randomness.
These theorems apply wherever the four conditions are met: strong coupling, observational asymmetry, feedback dynamics, and non-stationarity.
The domains affected include: oceans, pandemics, power grids, financial systems, climate systems, supply chains, and any complex socio-technical system.
9.2 The Choice Before Us
┌─────────────────────────────────────────────────────────────────────┐
│ │
│ THE CHOICE │
│ │
│ │
│ PATH A: CONTINUE PATH B: CHANGE │
│ ────────────── ────────────── │
│ │
│ • Maintain independence • Adopt coupling as │
│ assumptions default assumption │
│ │
│ • Classify failures as • Classify failures as │
│ "black swans" model failures │
│ │
│ • Protect models from • Expose models to │
│ falsification empirical testing │
│ │
│ • Prioritise mathematical • Prioritise empirical │
│ convenience accuracy │
│ │
│ • Shift blame to events • Assign accountability │
│ to modellers │
│ │
│ │
│ OUTCOME A: OUTCOME B: │
│ ────────── ────────── │
│ │
│ More "surprising" Fewer surprises. │
│ catastrophes. Better preparation. │
│ More post-hoc excuses. Genuine foresight. │
│ More preventable deaths. Lives saved. │
│ │
└─────────────────────────────────────────────────────────────────────┘
9.3 Final Statement
The independence assumption is not a neutral technical choice. It is a risk-distorting prior that systematically underestimates danger in exactly the systems where accurate prediction matters most.
For decades, we have built our safety infrastructure on a mathematical lie. We have called the resulting failures “unpredictable” when they were predictable artifacts of our own modelling choices. We have protected the assumption at the cost of the people the models were meant to protect.
This was always a choice, not a constraint.
The mathematics of coupled systems exists. The computational power to implement it exists. The data to validate it exists.
What has been missing is the will — the institutional, professional, and regulatory will to abandon convenient assumptions in favour of inconvenient truths.
The ocean does not care about our mathematical convenience. Neither do pandemics. Neither do power grids. Neither do financial systems. Neither does the climate.
They are coupled. They will behave as coupled systems. And they will punish models that pretend otherwise.
The era of independence-based risk science must end.
Not because it is philosophically wrong — though it is. Not because it is scientifically unsound — though it is. But because people die when we pretend that coupled systems are separable.
And that is no longer acceptable.
“The most dangerous assumption in risk science is the one nobody remembers making.”
We remember now.
It is time to build something better.
Jason Gething Founder — FishIntel Global Independent Researcher, Applied Probabilistic Systems
Eastbourne, United Kingdom December 2025
REFERENCES
This paper engages with foundational literature including: probability theory (Kolmogorov, Feller), extreme value theory (Fisher-Tippett-Gnedenko), copula theory for dependency modelling (Sklar), complex systems (Kauffman, Barabási), financial contagion (Allen & Gale), power grid cascades (Carreras et al.), pandemic modelling (Anderson & May), and climate tipping points (Lenton et al.). Full academic citations available upon request.
DECLARATION
Funding: This research was conducted independently with minimal resources.
Conflicts of Interest: The author is founder of FishIntel Global, which has developed coupled risk prediction systems.
Data Availability: Maritime incident data drawn from publicly available records. System accuracy claims are blockchain-verifiable.
Ethics: No human subjects were involved.
ACKNOWLEDGMENTS
To the mathematicians who knew but couldn’t be heard. To the practitioners who saw but couldn’t prove. To the victims of models that claimed certainty they didn’t have.
The assumption has been named. The failure has been documented.
What happens next is up to us.
© 2025 Jason Gething / FishIntel Global Ltd. All rights reserved.
This paper may be freely distributed for academic and educational purposes with attribution.
Contact:
Jason Gething Founder — FishIntel Global jason@fishintelglobal.com
🟢 Smarter Fishing. Safer Waters.
“The ocean does not care about our mathematical convenience.”