Research Library
Numinor research
Original quantitative research on Chinese A-share markets. Each paper ships with the methodology written out in enough detail to implement, and (for paid tiers) with a runnable codebase you can drive against your own factor stack.
v2.3 · September 2026
SAM Product Momentum — A Product-Spillover Alpha Signal for Chinese A-Shares Under Two Residualization Regimes
We document a cross-sectional alpha signal for Chinese A-share equities constructed from ChinaScope's SAM (Sector Analysis & Mapping) Level-2 industry/product taxonomy and the accompanying per-stock segment-revenue panel. For each focal stock at each rebalance date, the signal captures the recent movement of the stock's product environment — the size-weighted average return of pure-proxy stocks (those booking at least 50% of their Level-2 SAM segment revenue to a single product) — weighted by the focal stock's own revenue mix across SAM products. The methodology is evaluated under two parallel residualization regimes: Construction R (raw daily returns into signal construction; orthogonality only at evaluation) and Construction S (returns residualized against the 22-factor base before signal construction). Both regimes produce positive, robust orthogonal incremental ICIR against a standard 22-factor institutional risk model: +0.352 (R) and +0.307 (S) with 100% of 20 grid alignments positive. Under the industry-included variant adding 31 SYWG L1 dummies, Construction R rises to +0.435 OOS multi-offset orth-ICIR. Both regimes produce positive annual mean orth-ICIR in every year of the 8-year sample (2019–2026); the weakest year is 2022 (still positive in annual mean under both constructions). A comprehensive robustness battery — factor quality shuffle, factor family drop, base factor count scaling, multi-horizon evaluation, diagnostic turnover, and pre-orthogonalized delivery equivalence — confirms standalone-alpha character, microstructure-space concentration, horizon-monotonic strengthening (peak orth-ICIR at 60d), and moderate-fast portfolio turnover (~4.4–4.8× annualized).
Read the paper →Headline results
OOS orth-22 ICIR · Construction R
+0.352
OOS orth-22 ICIR · Construction S
+0.307
OOS orth-22+industry · Construction R
+0.435
v1.7 · September 2026
Network-Effect Amplification of Quantitative Factor Models in Chinese A-Shares
We present a systematic empirical study of network-effect amplification applied to quantitative factor models in Chinese A-share equity markets. Across two complementary empirical designs — walk-forward testing across 10 years and a fixed train-test hold-out — applied to a universe of 3,500–5,000 listed companies, we document that network-routed factor features add measurable incremental information to standard factor models, deliver portfolio Sharpe lifts of +0.54 to +0.72 out-of-sample, and operate as a robust amplifier rather than a fragile parallel signal.
Read the paper →Headline results
ΔICIR at 20-day horizon
+0.149
Sharpe lift (long-short 10–30%)
+0.54 to +0.72
Portfolio volatility reduction
~38%
v1.4 · September 2026
Numinor Co-Movement Graph — A News Co-Movement Graph Graded by a Structural Network, Identifying Which A-Share Co-Movements Are Structurally Grounded
A pairwise relationship feed over Chinese A-shares that forecasts which names genuinely co-move and flags which observed co-movements have no structural basis it can see. The substrate is the news co-movement graph — the pairs the market is co-mentioning (broad, timely, noisy). The overlay is a structural network of four 'lamps' — product peer (any shared product), SAM supply chain, disclosed customer–supplier, and affiliate, plus a deep-product-peer refinement of the first — that grade which co-movements are structurally grounded; a pair with at least one base lamp lit is 'confirmed', a pair with none is 'dark' (the discount list). Out-of-sample on a fixed 2023+ window (42 monthly vintages, of which the 38 with a complete forward window enter the averages), the deep product-peer link adds +0.065 to forecast forward correlation over trailing correlation and the other lamps — +0.114 in total over trailing alone (t ≈ 22, Fama–MacBeth); structurally-confirmed correlations retain ~92% of their level a quarter forward versus ~75% for a price screen; and a 10-name structure-selected hedge cuts ~45% of a CSI 300 name's forward residual variance — 3× a news-only basket of the same names (~15%) and at less than half the turnover of a price-correlation basket. This is a risk/correlation overlay — information content, not a return signal, not a portfolio, and not a covariance model — delivered as two tables from which the accompanying code re-derives the headline validation numbers (the t-statistics, benchmark comparisons and robustness checks fall outside that automated check).
Read the paper →Headline results
Deep product-peer · incremental fwd-corr (holdout)
+0.065 (t ≈ 22)
SAM supply-chain · incremental fwd-corr
+0.028 (t ≈ 9)
Connected pairs co-move > unconnected
100% of holdout months
v1.1 · October 2026
Numinor Hidden Sector Risk: Finding the Industry Risk That Shenwan-Neutral A-Share Portfolios Still Carry
A-share managers neutralise industry risk on Shenwan labels, and Shenwan puts every company in exactly one industry. Using ChinaScope SAM's product-segment data, this paper measures the industry risk such label-neutral portfolios still carry. In 204 long/short portfolios a month from 2016, each exactly neutral on Shenwan level 1, with factors estimated on one half of the market and portfolios built from the other, SAM uncovers nearly as much hidden risk again on top of Shenwan level 3 as level 3 itself does (18.7 against 20.0 bps a year, t 13.2, positive in 10 of 10 years). It holds on a model with 14 styles and 10 statistical factors, and 65% of it is lost when product mixes are shuffled among companies in the same Shenwan level-3 industry. Adding SAM to a level-3 neutralisation lowers realised volatility in 9 of 10 years while shuffled SAM raises it, but taking the risk out costs more signal than it saves, in diversified books and in concentrated books of 100 to 300 names. SAM's value is in seeing, reporting and weighing this risk in the risk model. The whole test reruns in the Numinor Matrix.
Read the paper →Headline results
Hidden risk SAM uncovers beyond Shenwan L3
18.7 bps a year
What Shenwan L3 uncovers beyond L1
20.0 bps a year
SAM incremental adjusted R²
+0.65pp
Read the full papers — free.
Headline results and abstracts are open. Sign up with a work email (or LinkedIn / Google / GitHub) to read the full methodology, send your agent into the Matrix to replicate the papers, and chat with the research assistant.