Numinor Co-Movement Graph — A News Co-Movement Graph Graded by a Structural Network, Identifying Which A-Share Co-Movements Are Structurally Grounded

Every correlation or covariance estimate starts from price history, and price history is unstable: a pair of stocks that looks correlated over the last quarter often is not over the next. We document a data layer built on a different substrate. We begin from the news co-movement graph — the pairs of China A-shares the market is actively co-mentioning — and grade each pair with a structural relationship network: shared products, supply chains, common ownership, and disclosed customer–supplier links. The grading forecasts which co-movements persist beyond what trailing correlation already contains, and, by its absence, marks which observed co-movements the graph cannot structurally explain — the "dark" pairs the market talks about together for no structural reason we can see. ("Dark" means no known structural link — not that the correlation is false; a pair may co-move for a macro or index reason this graph does not capture.)

The two layers do different jobs. News co-mention is the substrate — broad, timely, noisy; it tells you which pairs are in play right now. The structural network is the overlay — it identifies which of those co-movements are grounded in a durable link. News activates the network; the network explains the news. The product is the structure of their overlap: news ∩ network = structurally-confirmed co-movers, news − network = structurally-unconfirmed ("dark") pairs. The product is the graph and its correlation intelligence — a pairwise correlation overlay, not a return signal, not a portfolio, and not a covariance model. Every result below is out-of-sample on a fixed 2023+ validation window and is reproduced from the shipped data package by the accompanying code.

---

Headline empirical findings

Finding	Value	Construction
Deep product-peer link forecasts forward correlation, over trailing	+0.065 (t ≈ 22)	per-lamp Fama–MacBeth, holdout 2023+
SAM supply-chain link	+0.028 (t ≈ 9)	same
Disclosed customer–supplier (2-yr)	+0.013 (t ≈ 2.6)	same; marginal after overlap adjustment (§3.5)
Common-ownership (affiliate) link	+0.005 (t ≈ 2.2)	same; marginal after overlap adjustment (§3.5)
Connected pairs co-move above unconnected	100% of holdout months	grading gate
Grading ladder (forward correlation)	deep 0.52 > multi 0.41 > single 0.36 > random 0.28 > dark 0.27	tier means, holdout
Structurally-confirmed correlations retained forward	~92% vs ~75% random-pair price screen	retention, holdout (def. in WP §5)
10-name structure-selected hedge — forward residual-variance reduction	~45%	CSI 300, 11,262 stock-months
— vs a 10-name news-only basket (the overlay's value)	45% vs 15% (structure wins 89%)	size-matched, holdout
— vs a 10-name price-correlation basket	45% vs ~50% (comparable, not better)	size-matched, holdout
— vs a ~100-name industry basket / the index	beats 76% / 84% of months	CSI 300, holdout
Large-cap boundary	structure beats industry 76% (CSI 300) → 49% (CSI 1000)	the edge sits where the liquidity is
Tender/bid awards	excluded — unstable out-of-sample as a persistent link	distinct from C2C, where bids carry event-like return info

We report information content — correlation-forecast coefficients, retention, and variance reduction — not a portfolio Sharpe or backtested P&L; those are portfolio-construction-dependent and are for the buyer to measure on their own book. The effects are statistically stable and modest in size — the deep-peer and supply-chain lamps are the load-bearing forecasters, the realistic profile for a relationship dataset.

---

What the Co-Movement Graph is

A point-in-time, pairwise feed over China A-shares, built in two layers and delivered as two tables.

• The substrate — the news co-movement graph. Every pair in the feed is one the market is co-mentioning in news over the trailing 90 days (from ChinaScope's tagged-news corpus): ~54,000 A-share pairs over ~5,450 names in a typical month. Co-mention is a revealed-attention signal — when the market discusses two companies together, it is treating them as related. Broad, timely, and by itself noisy.
• The overlay — four structural "lamps." Each co-mentioned pair is annotated with deep product peer (the same specific product on the SAM ontology), SAM supply chain (a product-based input relationship), disclosed customer–supplier (financial-statement relationships, 2-year active window, bids excluded), and affiliate (common ownership). A pair with at least one lamp is confirmed; a pair with none — co-mentioned but structurally unexplained — is dark.

The two tables are an edge feed (one row per co-mentioned pair, carrying the substrate, the lamp flags, the confirmed/dark discount flag, and an expected forward correlation) and a derived per-stock peer set (each name's top-ten structural peers — the raw material for a hedge basket or a comparables universe).

This is distinct from a structural-relationship map. A network-only feed would enumerate every supply-chain or affiliate pair across ~6,000 names — most of them dormant at any moment, and silent on which observed co-movements have a structural basis. By scoping to the news substrate, the feed keeps only the links the market is currently acting on, and gains the one field a static map cannot have: a structural-confirmation flag on each observed co-movement. You can only confirm or leave unconfirmed a co-movement that has been observed, and observation is what the news substrate provides.

---

Mechanism: news activates, structure explains

The economic content is simple. Two firms co-mentioned because they make the same specific product, or sit on the same supply chain, share demand and cost drivers that produce persistent co-movement — whether or not last quarter's returns happened to show it. Two firms co-mentioned only because they appeared in the same headline last week do not. The structural lamps separate the two cases. In a typical vintage roughly two-thirds of co-mentioned pairs are confirmed by at least one lamp and one-third are dark.

The strongest lamp is the deep product peer: firms making essentially the same specific product add +0.065 of forward correlation over trailing alone, and roughly +0.11 in total once the shared-product base is included — about a quarter of the entire random-to-deep-peer range. SAM supply chain adds a smaller but firmly significant increment; disclosed and common-ownership links are positive but weak — marginal once you adjust for the overlapping forward windows (whitepaper §3.5) — so we lean on the two robust lamps.

Confirmed vs. unconfirmed correlations (the "discount list")

The most differentiated field in the feed is not a correlation we add but one whose durability we grade. Matching pairs on their trailing correlation and looking forward, structurally-confirmed pairs retain ~92% of their correlation a quarter later, versus ~75% for pairs selected on price history alone — a ~17-point durability gap that holds across correlation levels. This is the field a buyer cannot reconstruct from prices: which correlated pairs are correlated for a durable, structural reason and which only recently. Used operationally: trust the confirmed correlations, discount the unconfirmed ("dark") ones, whose level is no better than random and whose durability is materially lower.

The sharpest proof that the structure — not the news attention — is what works comes from hedging: a 10-name basket of a stock's structure-selected peers cuts ~45% of its forward residual variance, but a 10-name news-only basket — the stock's highest-co-mention peers with no structural lamp lit — cuts only ~15%, and the structure-selected basket wins 89% of the time. Against a strong price-only benchmark (top-10 trailing-correlation peers) the graph hedge is comparable, ~45% vs ~50% — it does not beat a good price selection, but it matches it from structure alone, with less than half the churn (~18% of names change month-over-month vs ~40% for a price-correlation basket), and decisively beats both the news-only basket and a ~100-name industry basket (~31%). The product implication: the graph is a structural prior — interpretable peers, low turnover, and a flag on which price-selected correlations deserve less trust — not a replacement for price-correlation selection where the only goal is maximum short-horizon hedge fit.

Why we report correlation forecasts, not a strategy Sharpe

A Sharpe depends on universe, weighting, neutralization, turnover, and costs that differ for every buyer; a correlation-forecast coefficient and a variance-reduction figure are portfolio-construction-independent and portable. You apply the graph to your own risk model and hedging book and measure your own contribution — the feed's corr_delta field shows exactly where the graph disagrees with the price screen. The decisive test is buyer-side replication on your stack.

Scope and limitations — stated plainly

• Large-cap. The hedging edge fades from CSI 300 (beats industry 76%) to a tie at CSI 1000 (49%). This follows from the substrate: small caps are not co-mentioned enough to enter the graph with signal, so the feed concentrates on well-covered names — exactly where the edge lives. The architecture and the boundary come from the same fact.
• Not a global risk-model replacement. The pairwise edge washes out in a whole-universe minimum-variance optimizer; the value is targeted — single-name hedging, pairs, concentrated-position risk.
• Correlation, not returns — not alpha. A return-spillover signal on the same graph is flat-to-negative out-of-sample. The graph forecasts how names move together, not which way.
• Tender/bid awards excluded. Unstable as a persistent correlation link out-of-sample. (In the C2C return paper, by contrast, bid awards still carry event-like directional information — a different use than grading co-movement.)
• Modest magnitudes, and the 2023+ holdout has been queried several times across validation — read the figures as confirmed-out-of-sample.

---

Next steps

The full evidence, figures, and feed schema are in the Co-Movement Graph Whitepaper v1.3, and every validation number is reproducible from the shipped data package by the accompanying code (verify_outputs.py → "✓ matches whitepaper"). To read the whitepaper, request a data trial, or discuss applying the graph to your own risk and hedging book, contact tyl@numinor.io.

---

Numinor Systems · ChinaScope · June 2026 · Companion to Whitepaper v1.3. A risk/correlation Construct Data product — information content, not a return signal. Contact: tyl@numinor.io