18 Asset-Dependency Networks (Econophysics) — Paper Review & Platform Fit Findings

Date: 2026-07-30 Reviewer: initial-thinking pass (operator-requested) Source paper: Raddant, M. & Di Matteo, T. (2023). A look at financial dependencies by means of econophysics and financial economics. Journal of Economic Interaction and Coordination 18:701–734. DOI 10.1007/s11403-023-00389-6. Open access (CC-BY 4.0). Question posed: Does the core idea of introducing a connected network of assets improve the platform’s model, and how much change would it require?

Update 2026-07-30: an empirical research spike was run against gold.fact_eod to pressure-test the one recommended path (Tier A — Diebold–Yilmaz connectedness as a leading regime signal). Result: the leading-signal hypothesis failed — see §6. This changes the recommendation: do not build the Tier-A time-series connectedness index expecting to anticipate stress. Details below.

⚠️ This is a reasoned-analysis + research-spike findings doc, not a producer or a validated edge. No production code was changed (read-only spike). It exists to inform a go/no-go on a future /design-brief.

18.1 1. What the paper actually is

A 2023 survey/review — not a single method with a backtested edge. It tours the econophysics + financial-economics toolkit for modeling the N×N dependency structure among assets as a network. Load-bearing techniques it catalogs:

Technique	What it produces	Maturity in the paper
Correlation matrix (Pearson, exp-weighted, partial, rank/tail)	dense N×N dependency	foundational
Random Matrix Theory / PCA	denoised covariance, “market mode” + group modes	foundational
Information filtering networks — MST, PMFG, TMFG + DBHT clustering	sparse graph (N−1 or 3(N−2) edges) + endogenous clusters	the paper’s centerpiece
Multivariate GARCH (DCC/BEKK/DECO)	dynamic conditional covariance	mature but hard to scale
Granger-causality / pairwise de-GARCHed regression	directed lead-lag network	explicitly weak — Billio et al.: significant links barely exceed the 5% false-positive rate except during crises
Diebold–Yilmaz variance decomposition / TVP-VAR	directed spillover / connectedness network → systemic-risk proxy	the most cited applied one

Crucial framing: none of these are presented as alpha. The paper’s own applications are risk topology, systemic risk, comovement/segmentation, and covariance denoising — descriptive structure, not return prediction.

18.2 2. Does the “connected network of assets” improve the platform’s model?

It depends entirely on which of three distinct value propositions you mean. Conflating them is where this kind of idea usually goes wrong. Two of the three are a poor fit for this platform specifically.

18.2.3 2.3 As a risk / crowding / systemic-observability layer — This is the real fit, and it’s a good one

Where the concept genuinely earns its place — and where the platform is already leaning:

sbcrowding (FC-1 comomentum) is already a correlation-network statistic — abnormal within-decile return correlation among the momentum family (comomentum_crowding_service.py + _comomentum_math.py, ~600 LOC). The paper is the general theory of exactly that move.
sbriskmodel already carries a factor covariance + idiosyncratic variance model (factor_covariance_service.py, idio_variance_service.py).
sbic already builds a factor-level pairwise-|ρ| network (pairwise_correlation_service.py, LH-6y / O-089).

So the network concept is not foreign — it exists at the factor level and in crowding. What’s genuinely missing is an asset-level dependency/connectedness read: a market-wide “how correlated/fragile is the whole book right now” regime signal, plus network-based clustering to (a) sanity-check GICS-sector neutralization, and (b) detect concentration/crowding in the traded cross-section that sector labels miss.

The single highest-value, lowest-regret extraction from this paper is Diebold–Yilmaz total connectedness as a report-only systemic-risk / regime diagnostic, optionally paired with TMFG + DBHT clustering of the universe. That fits the platform’s established grammar exactly: one-method-per-package, report-then-enforce, ops.research_* + sidecar → SPA, evidence-first, no gate at first (sbcrowding is the literal precedent).

18.3 3. Effort to bring it in

Scoped in platform units, anchored on the sbcrowding precedent (~600 lines of math+service + a close-out producer + ops.research_* table + sidecar + SPA card). Three tiers:

Tier	Scope	Rough effort	Risk
A — Diebold–Yilmaz connectedness diagnostic (recommended MVP)	One new producer (extend `sbcrowding` or new `sbconnectedness` leaf): rolling VAR + generalized FEVD over a sector/index return panel → total & directional connectedness index → `ops.research_*` + sidecar → one SPA regime card. Report-only, no gate.	~1 small epic / 4–6 tasks, comparable to FC-1. Math is standard statsmodels VAR.	Low — no portfolio/gate coupling.
B — TMFG + DBHT universe clustering	Add filtered-graph + clustering producer over the correlation matrix; surface network-clusters vs GICS sectors as an integrity/observability panel; feed a crowding/concentration read on the traded book.	+1 epic / 5–8 tasks. The TMFG + DBHT graph algorithms are the only genuinely new machinery — no currently-vendored library does them, so hand-rolled (Massara et al. 2017 is the scalable reference).	Medium — algorithmic surface; needs its own tests.
C — Dynamic conditional covariance / network-aware sizing	DCC-GARCH covariance feeding diversification-aware sizing in `sballocation`.	Large / multi-epic, and collides with X7.	High — don’t, unless the no-optimizer stance is being revisited deliberately.

18.4 4. Recommendation

Original recommendation (pre-spike): do Tier A only, as a report-only observability producer, and gate any move toward B/C on Tier A actually showing a connectedness signal that leads regime shifts.

Revised recommendation (post-spike, §6): the Tier-A leading-signal premise did not hold on this platform’s universe — DY total connectedness is coincident, not leading, has a saturated/narrow dynamic range at sector granularity, misses equity-specific selloffs, and is contaminated by composition drift. So:

Do not build the Tier-A time-series connectedness index as a regime predictor. Its marginal information over just watching realized vol is low, and it does not anticipate stress.
If a network producer is still wanted, the surviving rationale is the cross-sectional crowding/clustering angle (Tier B — TMFG/DBHT on the traded book, “which of my current holdings are one correlated cluster”), not the market-wide TCI. That is a different, heavier build and should get its own spike before commitment.
Everything worth having remains reachable without touching sbportfolio, sballocation, or X7; everything that would touch them is weak (alpha, §2.1) or against a settled decision (optimizer, §2.2). Unchanged.

Honest caveat (reaffirmed): this is a survey — no published edge to import. The spike confirms the concept’s value here is at best coincident risk legibility, not P&L or early warning.

18.4.1 Suggested next step (pick one)

Shelve the network work — the cheapest spike killed the strongest hypothesis; a coincident-with-vol index isn’t worth a producer. (Recommended default.)
Tier-B spike — if the crowding/clustering use still appeals, run a cross-sectional spike: TMFG/DBHT-cluster the current traded book and measure whether within-cluster concentration would have flagged real drawdown episodes. Only then consider a brief.
~~Tier-A /design-brief~~ — withdrawn; §6 removes its justification.

18.5 5. Reference-paper capture (convention reminder)

Per CLAUDE.md §7, any cited paper must be captured in docs/reference-papers/ (surfaced in-book as the Reference Library). This paper is open-access (CC-BY); if a /design-brief proceeds, add a row to docs/reference-papers/README.md and bundle the PDF as Raddant2023-financial-dependencies-econophysics.pdf. Not done here (no code/brief committed yet) — flagged so it isn’t missed at brief time.

18.6 6. Empirical research spike (2026-07-30) — does DY connectedness lead stress?

Read-only spike against gold.fact_eod. Goal: de-risk Tier A before any brief by testing its load-bearing premise — that a Diebold–Yilmaz total connectedness index (TCI) would give an early-warning regime signal in this platform’s own universe.

18.6.1 6.1 Method

Panel: sector-level daily equal-weight mean log return, built in DuckDB from gold.fact_eod.adj_close joined to dim_instrument.sector_sk. Cross-section restricted to liquid, non-ETF names (adj_close ≥ $5, adv_dollar_20d_f ≥ $1M, per-name return clipped to ±0.5 to guard split/data artifacts). Result: 5,698 trading days (2004-01-01 → 2026-07-29) × 11 GICS sectors.
Estimator: rolling 200-day VAR(2) → MA representation → generalized (order-invariant) FEVD (Diebold–Yilmaz 2012) at horizon H=10 → row-normalized → TCI = 100 × off-diagonal mass / N, stepped every 5 days (1,100 points).
Stress proxy: 21-day rolling annualized realized vol of the equal-weight market (mean sector) return.
Tests: (a) corr(TCI_t, forward vol_{t+h}) for h = 0/10/21/42d — rising with h ⇒ leads; falling ⇒ coincident/lagging. (b) lead/lag cross-correlation of ΔTCI vs Δvol. (c) crisis-window peak TCI vs full-sample median.
Reproduce: scratchpad/dy_extract.py (lock-polite extract) + scratchpad/analyze_dy.py (offline). Series: dy_tci_series.csv. Chart below.

18.6.2 6.2 Results

DY sector connectedness vs market stress

TCI is coincident, not leading. corr(TCI_t, forward vol_{t+h}) = +0.42 (h=0) → +0.38 → +0.31 → +0.24 (h=42d) — monotonically declining with horizon, the opposite of a leading indicator. Cross-correlation of changes peaks at lag 0 (+0.35) and goes flat/negative at positive (TCI-leads) lags (−0.17 at +5d). The ΔTCI change-form has a weak forward blip at h=+10d (+0.29) but nothing that would time a regime.
Saturated, narrow dynamic range. TCI mean 83.6%, band 69–90%. At sector granularity sectors are ~84% connected always; crises add only a few points.
Flags the big systemic crises, misses equity-specific ones. Peak-vs-median: GFC +4.5, Euro-2011 +6.0, COVID +5.7 — but Q4-2018 −3.0 (below median), 2015-16 +0.7, 2022 rate-shock +0.1. And the systemic peaks land at the crisis bottom (GFC peak 2009-03-26; COVID peak 2020-03-25) — coincident/lagging, not early.
Composition drift contaminates the level. Large multi-year TCI swings unrelated to stress (secular decline 2012→2018, a structural step ~2017-18, a sharp drop at the very end of sample) track changing cross-section membership/coverage more than market regime — a robustness problem for any thresholded use.

18.6.3 6.3 Verdict

The spike kills the Tier-A leading-signal hypothesis. DY connectedness here is a coincident echo of realized vol with low marginal information, blind to equity-specific selloffs, and sensitive to universe composition. It is not worth a producer as an early-warning/regime signal. This is the intended outcome of a cheap spike: the strongest recommended path was falsified before any code shipped.

Scope honesty: this tests the time-series/systemic use (Tier A) only. It does not test the cross-sectional crowding/clustering use (Tier B — network centrality of individual holdings), which remains unexamined and is the only surviving rationale for network work here — pending its own spike (§4, next-step 2). Also unswept: VAR lag order, window length, FEVD horizon, and alternative stress definitions (drawdown/VIX) — but the coincident signature is strong and consistent enough that tuning is unlikely to reverse it.