The Gaussian Assumption Is Breaking Every Risk System
Every standard risk system in capital markets rests on a single assumption that is empirically false: that financial returns are normally distributed. The consequences compound across every risk metric in every system.
Mandelbrot identified this problem in 1963. Taleb formalized the fragility argument in 2007. The empirical evidence has been accumulating for 60 years. Yet every major capital adequacy framework — Basel IV, FRTB, SIMM, SA-CCR — remains built on Gaussian foundations.
The Risk Intelligence framework replaces the Gaussian assumption with two empirically calibrated parameters from Mandelbrot's Multifractal Model of Asset Returns (MMAR, 1997) and fuses this with LLM orchestration via Model Context Protocol for regime surveillance and regulatory narrative generation.
Why This Domain Matters for AI Labs
Institutional capital markets fractal risk offers four properties that almost no other real-world domain provides simultaneously for AI training:
Replacing the Gaussian Assumption
Two parameters derived from MMAR drive all measurement corrections:
- Hurst Exponent H — measures long-range dependence. H = 0.5 (Gaussian random walk). H > 0.5 = persistence (trending volatility). H < 0.5 = anti-persistence (mean-reversion). Estimated via Detrended Fluctuation Analysis (DFA) and Rescaled Range (R/S) analysis on 5-year rolling windows.
- Tail Index α — characterizes the weight of distribution tails under an α-stable (Lévy) distribution. α = 2 recovers the Gaussian. α < 2 implies heavier tails. For financial returns, α typically ranges 1.45–1.88. Estimated via Hill estimator or maximum likelihood.
Risk Metric Replacements
Q/P Separation and Regulatory Bridge
The framework maintains a rigorous separation between the risk measurement layer (P-measure) and the pricing layer (Q-measure):
- P-Measure (Fractal Engine) — MMAR path generation, α-stable Monte Carlo, Hurst-adjusted PFE scaling, Clayton copula WWR detection. Operates under real-world probability measure for risk capital calculation.
- Q-Measure (Pricing Layer) — Remains arbitrage-free for fair-value reporting. The fractal engine informs economic capital; Q-measure pricing remains independent for mark-to-market.
- Regulatory Bridge — Maps fractal outputs to FRTB Expected Shortfall, SIMM sensitivity-based IM, SA-CCR EAD, and ICAAP/ILAAP capital requirements. Enables coexistence with Gaussian-anchored regulatory standards.
- LLM Orchestration (MCP) — Model Context Protocol servers expose the fractal engine, regulatory data feeds, and document stores to the LLM agent. The LLM sequences tool calls, monitors H/α drift, and generates supervisory-quality regulatory narratives.
LLM Roles in the Capital Markets System
| Role | Trigger | Output | Regulatory Use |
|---|---|---|---|
| Regime Surveillance | H and α drift beyond thresholds | Alert + recalibration trigger | SR 11-7 model monitoring |
| Capital Narration | Month-end capital calculation | Risk committee memo | ICAAP Board pack section |
| FRTB Documentation | Model change / IMA application | Technical documentation | Regulatory submission |
| Stress Scenario Design | Regulatory stress testing cycle | Fractal-calibrated scenarios | FRTB stressed ES scenarios |
| Counterparty Alert | CCR book H-drift detection | Counterparty escalation brief | SA-CCR / CVA oversight |
Fractal XVA and the Wrong-Way Risk Problem
The full XVA suite is systematically understated under Gaussian assumptions. For a derivatives book with H=0.62 and α=1.65, the corrections are material:
| XVA Component | Classical Treatment | Fractal Enhancement | Uplift |
|---|---|---|---|
| CVA | Gaussian PFE × PD × LGD | Fractal PFE (H-adj) × PD × LGD | 25–50% larger |
| DVA | Own Gaussian exposure profile | Own α-stable exposure — heavier tails | 15–30% larger |
| FVA | Gaussian funding cost √T scaling | Hurst-adjusted funding horizon | 10–25% larger |
| KVA | SA-CCR regulatory capital | Fractal economic capital add-on | 30–50% larger |
| MVA | SIMM IM (Gaussian Greeks) | Fractal Greeks add-on per class | 12–35% larger |
Wrong-Way Risk (WWR) — where counterparty default probability and exposure at default are positively correlated — is dramatically underestimated under Gaussian copulas. The Clayton copula used in this framework captures the asymmetric lower tail dependence that defines WWR in stress periods: joint extreme losses are more correlated than joint extreme gains.
Federated Deployment & Framework Conclusion
The framework deploys across three tiers to accommodate institutions ranging from fintech sandboxes to air-gapped national regulators — all while maintaining complete data sovereignty at every tier and zero cross-tenant signal leakage.
Honest Gaps and Limitations
Intellectual honesty is a design principle of this framework. The following limitations define the boundary between what is claimed and what remains to be demonstrated:
- Default correlation modelling — single-counterparty exposure is handled well; a full fractal default correlation model for portfolio CCR (replacing Gaussian copula in CDO/credit portfolio models) is a planned extension.
- Backtesting requirements — regulators will require comparative backtesting demonstrating fractal outperformance vs Gaussian baselines during 2008, 2020, and 2022 stress periods. This roadmap is planned for subsequent publication.
- Parameter smoothing — a Bayesian smoothing layer with regime-conditional priors is required to prevent pro-cyclical capital swings from noisy H/α re-estimation in quiet markets.
- Regulatory acceptance timeline — SA-CCR and FRTB remain Gaussian-anchored. The regulatory bridge addresses this operationally, but full framework acceptance requires regulator engagement over a multi-year horizon.