A Universal Dependence Metric: Four Numbers That Replace Independence
When correlated components are modeled as independent, the probability of simultaneous failure is understated by a factor that grows exponentially in system size — 50,773× for five assets with heavy-tailed dependence, exceeding 10^89 at fifty. We define eight axioms a dependence measure must satisfy (independence baseline, comonotonicity ceiling, marginal invariance, tail sensitivity, monotonicity, dimensional scaling, decomposability, continuity), show that CDF-based divergences fail dimensional scaling and nonparametric mutual information estimation degrades beyond moderate dimension, and construct a four-number metric: D_G (Gaussian copula mutual information from pairwise ranks), D_T (tail excess ratio vs. Gaussian baseline), D_S (upper/lower tail persistence slope), and D_H (hidden dependence excess). We prove D_G is a lower bound on true mutual information, decomposes via Schur complements, and satisfies most axioms — with explicit counterexamples where it does not. The metric discriminates copula families with 100% power at n = 1,000, predicts out-of-sample tail losses (r = 0.82), and catches adversarial evasion attacks via D_H at fire rates as low as 0.1%. For the equicorrelated Gaussian copula, log R ≈ c(ρ) · D_G, confirming that independence mispricing is exponential in measured dependence. We derive a capital charge formula K from the four numbers and validate the metric across finance, engineering reliability, epidemiology, supply chains, and machine-learning ensembles — five domains where the independence assumption hides catastrophic compound risk.
Authors
- Tal Zisckind
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-18
- DOI
- https://doi.org/10.5281/zenodo.24357461
- Primary Topic
- Financial Risk and Volatility Modeling
- Type
- article
- Field-Weighted Citation Impact
- 0.00