Network-Effect Amplification of Quantitative Factor Models in Chinese A-Shares

We document that routing standard quantitative factors through a supply-chain peer network — using ChinaScope's SAM industry classification and inter-product relationship graph — adds measurable, regime-stable, orthogonal information to a 22-factor Barra-style model in the Chinese A-share market. Pure SAM contribution is +0.149 ICIR on top of the base at the canonical 20-day forward horizon, scales monotonically with base factor count (the operator behaves as an amplifier on existing factor books rather than a substitute for them), and translates to +0.54 to +0.72 Sharpe lift at standard long-short cutoffs out-of-sample on 2022–2026. Findings are based on a 10-year sample (2016–2026) of ChinaScope SAM data. Three data-vintage conventions are disclosed in the full whitepaper and should be read alongside these figures: filings are admitted on a fixed fiscal-period-end-plus-60-day schedule rather than on their actual publication dates, which places 62.7% of in-window company-period observations in the feature set before the filing was public; the factor layer is mixed, being point-in-time for realized returns and daily market data but fixed-schedule for anything drawn from filings; and the product-to-product topology follows the current-vintage ChinaScope graph. Numinor and ChinaScope are inviting selected institutional A-share quant funds to evaluate the SAM data layer on their own factor stack under NDA.

---

Headline empirical findings

Finding	Value	Construction
Pure SAM contribution, canonical	+0.149 ΔICIR	22-factor base, 20-day forward, individual factor allocation
At longest forward horizon	+0.265 ΔICIR	120-day; positive and monotonic across all five horizons tested
Portfolio Sharpe lift	+0.54 to +0.72	10–30% L/S cutoffs, joint Ridge CV-tuned, fixed hold-out 2022–2026
Walk-forward year stability	8 / 8 positive	2019–2026, every year
Random factor base stability	80 / 100 positive	Stratified random subsets of base, mean +0.068 ΔICIR; median +0.052
Family-drop stability	7 / 8 positive	Single-family drop tests; 8 / 9 including full-base reference
Standalone capacity at noise base	+0.22 ICIR	Base replaced with random noise; SAM still informative
Amplifier coefficient	+0.97 correlation	Between base factor count (1–22) and ΔICIR contribution
Turnover impact (long / short / combined)	−13.7% / −7.5% / −10.5%	Annualized one-way turnover, base 22 vs. base 22 + SAM, 30% L/S diagnostic portfolio

All numbers are gross of transaction costs. Net IR transferability requires the buyer's own cost, liquidity, borrow, and market-impact models.

---

What the SAM network is

SAM (Sector Analysis & Mapping) is ChinaScope's multi-level industry classification system for the Chinese A-share market, with a delivered taxonomy of 6,544 product nodes organized across a nine-level hierarchy — from 116 broad industry categories at Level 1 (such as Auto Parts and Equipment or Beverages) through progressively finer sub-categories down to highly specific product lines at the deepest levels. Each listed company is mapped to a primary product based on its filed segment-revenue disclosures. The empirical work in the companion whitepaper admits each filing 60 days after the fiscal period end (whitepaper §2.3.3); the shipped Construct Data product instead admits it 30 days after the filing's actual publish_date. SAM is paired with a complementary supply-chain dataset that defines directional product-to-product input-output relationships. For the empirical work in the companion whitepaper we use the core input subset (COREIN) — 34,481 directional edges across 5,219 of the SAM product nodes that capture supplier-customer production dependencies. Together these create a directed graph that links companies across both upstream-downstream production chains and shared peer-product memberships.

For each focal stock at each trading date, we identify three sets of related stocks: peers (companies sharing the focal stock's primary product), upstream (companies whose primary product is an input to the focal product), and downstream (companies whose primary product uses the focal product as input). For every base factor, we compute revenue-weighted averages over each set, producing peer-routed, upstream-routed, and downstream-routed variants of the original factor. After cross-sectional orthogonalization against the 22-factor base, these network-routed factors become the SAM amplification layer that the empirical findings above are measured against.

---

Mechanism

The SAM contribution is concentrated in portfolio volatility reduction of approximately 35–38% (annualized portfolio σ falling from ~12.0% to ~7.4% at the 30% L/S cutoff), not mean-return enhancement — consistent with the network-aggregation operator capturing systematic peer-shared exposures that the base model treats as idiosyncratic. The downstream-product-side channel dominates the total contribution (approximately 70%), consistent with the Cohen and Frazzini (2008) mechanism in which customer-side information can propagate to suppliers with a measurable lag. SAM captures this at the product-network level rather than through direct company-level customer links — companies sharing customer or supplier products are aggregated together, exploiting the same economic relationship at a more general resolution. The contribution scales monotonically with forward-return horizon (positive at 5-day, strongest at 120-day), confirming a multi-month economic-propagation interpretation rather than short-horizon trading noise.

---

Scope and limitations

This is diagnostic evidence, not a production-validated backtest. Three caveats are relevant for institutional readers:

• Diagnostic L/S construction. Long-short portfolios are equal-weighted at 10–30% cutoffs without sector neutrality, liquidity weighting, position limits, or borrow constraints. Production deployment under buyer-specific risk and execution constraints requires direct backtesting on the buyer's stack.
• Net IR transferability is buyer-dependent. Our turnover measurement indicates the SAM layer does not mechanically increase trading-cost burden (it reduces turnover modestly), but absolute net-IR estimates require the buyer's own cost, liquidity, borrow, and impact models.
• Data-vintage conventions. Company filings and segment revenues are admitted on a fixed fiscal-period-end-plus-60-day schedule rather than on their filing dates, so a majority of annual filings enter the feature set before they were public (whitepaper §2.3.3 and §7.1.8 quantify this). Factor gating varies by factor. The supply-chain topology (product-to-product edges) reflects the current ChinaScope graph applied throughout the historical sample. Whitepaper §7.2.5 specifies a buyer-side trial protocol for data-vintage sensitivity testing, including a re-run anchored on actual filing dates.

---

What the trial includes

We are inviting a small number of institutional A-share quant funds to evaluate the SAM data layer in their own environment under NDA. Each trial includes:

• The full ChinaScope SAM data package. Ten years of SAM industry classification, segment revenues, and the COREIN supply-chain edge set, with the filing dates needed to apply your own availability rule. Delivered as clean parquet files with full data dictionary and integration documentation.
• Plug-and-play feature construction code. Python implementation of the network-routing operators described in the whitepaper — peer / upstream / downstream aggregation, revenue weighting, cross-sectional orthogonalization. Drop the buyer's own base factor stack in; reproduce the analysis on the buyer's own universe and constraints.
• Methodology consultation. Direct technical engagement with Numinor's research team on integration, robustness testing against the buyer's specific factors, and production deployment considerations.

---

Next steps

• Read the full technical whitepaper (Numinor Whitepaper v1.7, 72 pages) for complete methodology, all eight robustness analyses, and the full supporting figure set.
• Request a data trial through your existing ChinaScope or Numinor relationship, or contact us directly at the address below.
• For technical methodology or replication questions, reach out to Numinor Systems directly.

---

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