13 Systemic-Risk & Crowding: Existing Groundwork
Author: Todd B. Adams Reinforces: Proposal §2 — Theoretical Foundations and §3 (network layers) · Reading order: chapter 05 of this part; follows the factor layer as multiplex Layer 1, precedes the integrated target architecture.
What this chapter establishes. The proposal’s network idea is not foreign to this platform — a version of it is already running, and we have already tested the simplest form of its central hypothesis and found it wanting. The platform carries correlation-network, factor-network, and crowding machinery in production, and an empirical spike against a published connectedness methodology returned a negative result that both establishes precedent and sharpens the proposal’s distinct contribution.
13.1 Network structure already lives in the platform
The proposal treats assets, factors, and funds as nodes across topological layers. The platform already computes network statistics over two of those object types today, each grounded in a published method:
| Existing capability | What it is, in network terms | Grounded in |
|---|---|---|
| Comomentum crowding | Abnormal within-decile return correlation among a factor’s constituent names — a correlation-network statistic over the traded cross-section | Lou & Polk (2012), Comomentum |
| Factor correlation network | Pairwise absolute Spearman correlation among all active factors — a weighted factor–factor graph, collapsed to an effective trial count | Benjamini–Yekutieli false-discovery control over a dependency graph |
| Orthogonality / confound test | Residual information coefficient of a candidate factor against the validated book — an edge-removal / independence test on the factor graph | rank-space residualisation |
| Factor risk model | Factor covariance plus idiosyncratic variance — the dense dependency matrix the proposal proposes to denoise | Barra-style decomposition |
The network concept is therefore already present at the factor level and in crowding. What the proposal adds is an asset-level, multi-layer read and a spectral phase-transition signal — genuinely new, but built on a substrate that already speaks this language.
13.2 A prior empirical test — and its negative result
The proposal’s supervisor, Tiziana Di Matteo, co-authored the survey we set out to stress-test: Raddant & Di Matteo (2023), A look at financial dependencies by means of econophysics and financial economics. In a read-only research spike run in July 2026, we computed a Diebold–Yilmaz total connectedness index over the platform’s own 2004–2026 universe — 5,698 trading days across 11 GICS sectors, via a rolling vector-autoregression and generalised forecast-error-variance decomposition — to test whether network connectedness leads market stress.
We report the finding, and it is negative:
- Connectedness proved coincident, not leading. The correlation between the connectedness index and forward realised volatility fell monotonically with horizon, from roughly +0.42 contemporaneously to roughly +0.24 at a 42-day horizon, with a saturated dynamic range and visible sensitivity to universe composition. These are in-sample descriptive correlations from a single spike on one universe; we did not attach a bootstrap interval, so we read them as preliminary rather than established. The full write-up is the asset-dependency-networks findings.
- This falsified the simplest network-as-early-warning hypothesis before any production code shipped — the empirical discipline that doctoral-grade research rewards, and reported here with the same care a positive result would receive.
- It also localises the proposal’s contribution precisely. The proposal does not claim that single-layer connectedness predicts crashes. It claims that the eigenvalue spectrum of the Supra-Laplacian — algebraic connectivity and critical slowing down across a multi-layer graph — carries an early-warning signature. That is a different and stronger mathematical object than a single-layer vector-autoregression connectedness index, and this spike is the evidence that the weaker version was tested and found wanting. The proposed research picks up exactly where the spike left off.
13.3 How this maps onto the proposal’s layers
- Layer 3 (institutional ownership / crowding). The comomentum producer is a return-correlation proxy for crowding today; the proposal’s 13F ownership layer is the holdings-based ground truth. One is an approximation of the other — see the gap analysis in chapters 03 and 07.
- Physics evaluation engine (§3). The platform already has a report-only, evidence-first grammar — a research table feeding a per-run sidecar feeding a dashboard card, gating nothing at first — that a Supra-Laplacian spectral producer would inherit. The comomentum producer is the literal template (chapter 06).
13.4 A citation-capture note
The Raddant & Di Matteo (2023) survey is not yet captured in the reference library, though it is open-access under a Creative Commons licence; chapter 10 flags it for bundling. Lou & Polk (2012), Comomentum, is already in the library and grounds both the platform’s crowding producer and the proposal’s Layer-3 motivation.
Cross-links: proposal §2 literature · reference library · connectedness-spike findings