Correlation budgets

Definition

A correlation budget caps how much of a book may load on any single common driver — sector, factor, macro regime, or crowding cohort — enforced with STRESS correlations rather than calm-period estimates. It operationalizes the two documented facts that break naive diversification: correlations are ASYMMETRIC (higher in down markets — Ang-Chen 2002) and EXTREME co-movement is worse than normal correlations imply (Longin-Solnik 2001). The budget’s question: what fraction of this book is secretly one trade?

How it works / structure

• The measurement stack (engine-executable): rolling pairwise correlation matrices (calm estimate), downside/ stress-conditional matrices (the budget’s binding input — estimated from drawdown periods or stress-scenario overrides), eigenvalue decomposition (the first principal component’s variance share = the book’s “one-trade-ness”), and named-driver mapping (each position tagged to sector / factor / regime drivers — port-exposure-netting supplies the netting).
• Budget structure: caps per driver (e.g. no more than X% of portfolio risk from any one sector or factor under stress correlations), crowding overlays (sent-13f-holdings hotel names budgeted as a driver — herded positions fail together, bias-herding), and a total concentration ceiling (first-PC share below a threshold).
• The asymmetry evidence: Ang-Chen — correlations conditional on downside moves exceed upside-conditional ones systematically; Longin-Solnik — extreme-tail co-movement exceeds what normal-distribution correlation predicts; both mean CALM MATRICES FLATTER every budget — hence the stress-conditional rule.
• Regime linkage: correlation level tracks the volatility regime (regime-volatility) — budgets tighten mechanically as the regime state escalates.

When it applies

Every multi-position book (the diagnostic runs standing); position-add decisions (the marginal question is the new position’s stress correlation to the EXISTING book, not its standalone merits); strategy mixing (strategy-level correlation budgets — two strategies with one driver are one strategy); crowding management (the 13F/flow overlays).

Risk profile & failure modes

• Calm-matrix flattery: the budget’s reason for existing — enforcing on trailing-year correlations approves exactly the concentrations that stress reveals.
• Unnamed drivers: budgets cover the drivers someone thought to name; 2007’s quant unwind correlated books through a crowding driver nobody budgeted — the first-PC ceiling is the catch-all for unnamed structure.
• Estimation noise: stress-conditional matrices use few observations by construction — wide error bars; the platform treats them as bounds, not points, and lets scenarios (risk-scenario-analysis) override upward.
• Budget gaming: reclassifying positions to free budget (the tech position tagged “consumer”) — driver tags are mechanical (classification data), not discretionary.

Evidence & limits

Longin-Solnik (2001) and Ang-Chen (2002) are the peer-reviewed asymmetry/extreme-correlation evidence; crowding-unwind episodes are documented (2007, 2021). Budget THRESHOLDS are policy choices, not empirical constants — the platform sets them per book and audits them against realized drawdown attribution.

Falsifiable-thesis examples

Illustrations only, not signals:

• “No single driver will contribute more than 40% of this book’s realized drawdown in the next 5%+ portfolio decline (budget effectiveness)” — falsified by the attribution.
• “The book’s first principal component explains under 50% of daily return variance this quarter (concentration ceiling)” — falsified by the decomposition.

Cross-references

• The per-position risk view: risk-correlation-exposure
• The arithmetic underneath: port-diversification-math
• The crowding overlays: sent-13f-holdings, bias-herding
• The regime driver: regime-volatility; the scenario override: risk-scenario-analysis