Observed Supply-Chain Networks and Customer-Momentum Spillover in Chinese A-Shares — Evidence from Disclosed and Materially-Sized Bid Company-to-Company Relationships

We document that an observed company-to-company supply-chain graph for Chinese A-shares — built by combining mandatory top-customer / top-supplier disclosures with publicly-recorded procurement-award records, and resolved to listed-company identity through ChinaScope's affiliate-ownership tables — carries measurable, orthogonal, industry-neutral cross-sectional information about future returns.

Validated through a Cohen–Frazzini (2008) customer-momentum test (a focal seller's recent customer returns predicting its own forward return), the headline union signal — disclosure, refined where a materially-sized procurement relationship also exists by that bid evidence — earns an orthogonalized information ratio of +0.47 full-sample / +0.39 out-of-sample (t = 4.0 / 2.8) at the 20-day horizon, across 4,863 listed companies (~85% of the listed universe; ~2,476 per rebalance), after residualization against a 22-factor Barra-style base. The effect strengthens under additional industry neutralization (out-of-sample +0.48), is positive across all rebalance phasings and every full calendar year 2020–2025, exceeds disclosure alone at every phasing, and is not concentrated in micro-cap or illiquid names. The product is the relationship graph; the customer-momentum signal is its validation, not the whole of its use. The research is published openly, and every result is reproducible from the underlying data and the accompanying code.

---

Headline empirical findings

Finding	Value	Construction
Union signal orthogonal ICIR (headline)	+0.47 full / +0.39 OOS	disclosed ∪ materially-sized bid, 22-factor residualized, 20-day forward
— significance	t = 4.0 / 2.8	full / OOS, 74 / 51 monthly rebalances
— coverage	4,863 companies (2,476 / rebalance)	~85% of listed universe; ~½ of tradable names per rebalance
Disclosed channel (precision core)	+0.41 full / +0.31 OOS (t 3.5 / 2.2)	4,828 companies, value-weighted, n ≥ 1
Material-bid channel (denoising overlay)	+0.36 full / +0.34 OOS (t 3.0 / 2.4)	1,483 companies; bid ≥ seller's median disclosed customer
Industry-neutral union ICIR	+0.56 full / +0.48 OOS	+ SYWG L1 one-hots — signal strengthens
Multi-offset stability (union)	100% positive (10 phasings), > disclosed at each	full and OOS
Random 12-factor subset stability	100/100 (disclosed and union)	re-orthogonalized each draw
Calendar-year stability (union)	positive every full year 2020–2025 (2026 YTD +)	20-day forward
Union vs. disclosure (paired test)	t = +2.1 full / +1.9 OOS; union ahead at all 10 phasings	a genuine refinement, not just coverage
Materiality-screen sensitivity	union beats disclosure across every band; unscreened bid does not (paired t ≈ 0)	the screen, not its level, is what matters
Diagnostic quantile spread (gross)	+51 bps full / +33 bps OOS per 20 days	descriptive economic magnitude, not a Sharpe

All figures are gross of transaction costs and regenerated from SHA-frozen inputs by a single reproduction script. We report information content (information ratios and coverage), not portfolio Sharpe or alpha — these are portfolio-construction-dependent and are for the buyer to measure in their own book.

---

What the company-to-company graph is

The graph is an entity-level, directional, valued, point-in-time edge set in which each edge is a supplier → customer relationship between two listed companies. It is assembled from two complementary ChinaScope sources:

• Disclosed edges — mandatory top-five customer and top-five supplier relationships, plus related-party transactions, from annual and interim financial statements. Sparse in relationship depth (only the largest counterparties), but reaching ~85% of the listed universe at the focal level.
• Bid-evidenced edges — awarded public procurement contracts (public tender-award records), where the winning bidder is the seller and the purchaser is the buyer, valued by contract amount. Continuous and transaction-specific. For the momentum signal we admit only the awards that are material to each individual seller — at least as large as that seller's typical (median) disclosed customer — so the validated bid layer refines disclosure rather than extending it (the full edge set, including the smaller awards, still ships in the data).

Both buyer-side and seller-side ChinaScope records are used and deduplicated; named parties are resolved to their controlling listco through ChinaScope's structured affiliate-ownership tables (not a language model), with ownership ratios carried on each edge. The result is distinct from a product-taxonomy network of the kind used in Numinor's SAM Amplifier study: those links are inferred from product proximity, whereas these edges record companies that actually transacted.

This relationship graph is delivered as Numinor Construct Data — a daily-refreshed, point-in-time, factor-model-agnostic edge table (supplier/customer tickers, ¥ relationship value, ownership rollup, source type, availability date, provenance). The customer-momentum signal validated here is one construction over the table; the same edges also support supplier-side spillover, concentration and counterparty-risk measures, network-centrality features, and shock-propagation studies.

---

Mechanism: precision core plus a per-company denoising overlay

The economic content is the Cohen–Frazzini customer-momentum effect, operating on an observed A-share graph, with a clear division of labor between the two channels. Disclosure is the precision core — the higher-quality, near-market-wide channel that carries most of the signal. Materially-sized bid awards are a per-company denoising overlay — screened, for each seller, to contracts at least as large as that seller's typical (median) disclosed customer, so only economically meaningful procurement relationships enter. The screen is what makes it work: a procurement award, unlike a disclosed top-five customer, carries no built-in materiality, and admitting every award dilutes the signal.

Calibrated to each company's own disclosed scale, the surviving bid evidence is information-rich — on the overlap segment, in this sample, the bid signal is actually the stronger of the two (disclosure on those names is statistically indistinguishable from zero), and the two are nearly uncorrelated there (rank correlation +0.07), so the bid carries independent information. A paired test shows the union genuinely improves on disclosure (paired t = +2.1 full-sample and +1.9 out-of-sample, with the union ahead at all ten rebalance phasings), not merely matching it with extra names; the union's edge is variance reduction — it has the lowest IC-volatility of any channel — rather than the highest mean IC. On its own the materially-sized bid channel earns a clean +0.36 full / +0.34 out-of-sample, comparable to disclosure itself. The effect's monthly-horizon strength and multi-week persistence are consistent with the limited-attention propagation mechanism documented for U.S. equities and, in prior broker research, for A-shares.

---

Why we report information content, not portfolio metrics

A deliberate choice governs every number here: we report the information ratio, not a Sharpe ratio or alpha. The information coefficient is a cross-sectional quantity that depends only on the signal and realized returns — it is the most portable diagnostic across buyers. A portfolio Sharpe, by contrast, is a function of universe, weighting, neutralization, leverage, and cost assumptions that differ for every buyer; the same signal combined two ways produces two different Sharpe figures. We therefore prove information and orthogonality and leave the portfolio contribution for the buyer to measure on their own stack — which is the appropriate, and only honest, decisive test.

---

Scope and limitations

This is diagnostic evidence on a data layer, candidly bounded:

• Information content, not a guaranteed alpha. The buyer's realized contribution depends on their factor library, universe, and construction; the orthogonal ICIR is necessary, not sufficient, evidence.
• High turnover. The signal refreshes monthly (basket turnover ~54–60%/month); net-of-cost value must be assessed under the buyer's own cost model — though, for a fund already rebalancing monthly or faster, only the marginal turnover matters.
• Out-of-sample window is 51 monthly rebalances. Adequate to resolve the information ratio (t ≈ 2.8), but short for portfolio-level inference; we present it as validation evidence rather than conclusive proof, make no Sharpe claim, and treat buyer-side replication as the decisive test.
• The bid channel is a denoising overlay, not a coverage extension. Because the materiality screen needs a disclosed footprint to calibrate against, the validated bid layer lives within the disclosed universe and does not reach undisclosed names; its role is to refine the disclosed signal, not to extend it.
• Topology-vintage convention. The company-financial and return layer is strictly point-in-time; the entity-identity rollup reflects the current-vintage mapping applied across the sample.

---

Next steps

• Read the full technical whitepaper (Numinor C2C Supply-Chain Whitepaper v3.0) for complete methodology, the channel decomposition, the industry-neutral and liquidity-bucket diagnostics, and the full robustness battery.
• For methodology or replication questions, reach out to Numinor Systems.

---

Numinor · ChinaScope · June 2026 Contact: tyl@numinor.io