Why Fine Print & Deterministic Code Win
No Crowded Headlines.
No LLM Guesswork.
When an earnings release hits the wire, the headline number is priced in milliseconds, while a generative language model prompt anchors on management's press-release spin and produces two different portfolios if you run it twice. Real edge lives in structured footnotes, statutory exhibits, and uncrowded data streams parsed by deterministic code.
Priced-In Wires & Prompt Drift
Every quant pod and retail screen sees the headline beat at the exact same millisecond, and language models are trained to summarize commentary rather than price balance-sheet risk. Neither catches when a clean headline hides receivables outgrowing cash flow, a customer concentration cliff in Note 18, or a stock that already gapped +6% before the open.
Many Narrow Scouts. One Risk Kernel.
Whether a signal comes from a graduated SEC filing scout or a canary alt-data scout, no single scout ever places a trade on its own. Every reading is deduplicated to its primary document, rank-calibrated, stripped of sector and style beta, checked against hard accounting and credit vetoes, and sized by volatility parity into one reproducible book.
From Scattered Filings to Coordinated Flight
Four Practical Questions.
Four Mathematical Answers.
Turning a swarm of archival filing scouts and live ephemeral feeds into one disciplined portfolio curve requires answering four practical questions on every candidate: when did the public actually see the disclosure, how do we combine scouts without double-counting, is the edge independent of the sector, and how big should the bet be.
Down-to-the-Second SEC acceptanceDateTime & Survivorship-Free Panels
Backtests cheat when they use calendar filing dates that ignore whether a report dropped at 8:00 AM or 4:15 PM after the close, or when they only test on companies that survived to 2026. We join every SEC XBRL number to its exact down-to-the-second acceptanceDateTime across 3,607 historical Russell 3000 issuers (retaining 555 delisted companies with Shumway -30% delisting haircuts), keyed by canonical 10-digit SEC CIK, composite point-in-time OpenFIGI identifiers, and One-Document-One-Witness accession provenance alongside automated bad-tick and unadjusted split scrubbing.
OOS Blom Quantiles, Concave Pooling & One-Way Canary Glass Wall
Corporate events have fat tails, and three parsers reading the same filing are not three independent bets. We deduplicate signals sharing a primary filing accession (1 Document = 1 Witness), map events onto Out-of-Sample Blom quantiles and zero-anchored MAD0 scales, and pool them with diminishing returns across four orthogonal pillars (P1–P4) behind a One-Way Canary/Shadow Glass Wall so probationary Darwinian arms never inflate core conviction. Empirical Bayes shrinkage pulls noisy scouts to zero (IC0 = 0.0), while Form 8-K Item 4.02 restatements, Form NT notices, acute HYG/IEF credit-spread impulses, and opposing-pillar collisions enforce hard one-way vetoes to zero (with zero single-stock inverse-ETF routing).
Six-Dimension Ridge Residualization, Purged CV & Sub-Arm Governance
Making money because tech went up is not stock-picking edge. We strip out each stock’s GICS Sector ETF return and five OHLCV style factors (Market, 20d Dollar-Volume Size, 12–1M Momentum, 5d Short-Term Reversal, and 14d ATR Volatility) via trailing 252-day Ridge regression with zero pre-entry intercept subtraction. Scouts and isolated sub-arms must survive 5-fold Purged Walk-Forward CV (10d horizon + 2d settlement embargo = 12d purge), a frozen 45-trial Earned-Seat Moratorium, Dual-Maturity Cohort Gating, a 50/50 Drop-1 leave-one-out mean check, and continuous out-of-sample veto counterfactual attribution.
Volatility-Parity Sizing, Crowding Guards & Volume-Clock Decay
Even a real signal loses money if trading friction or high-turnover regime switching eats the edge (Kaminski & Lo, 2014; Lo & Remorov, 2017). Candidates must clear round-trip spread and slippage hurdles (ηcost · RTCosti,t, ηcost > 1) via Barra-implied fair-midpoint collars before sizing by Grinold active risk-aversion, inverse-ATR stop parity, and Ledoit-Wolf GICS cluster plus market-residualized pairwise crowding guards, paired with volume-clock subordinated alpha decay, volatility-scaled profit ratchets, 2-sweep stop confirmation, and 10-day regime dwell + VIX contango hysteresis inside a 1.00x cash-funded core.
Four Independent Corners of the Fine Print
Four Corners of the Public Record.
Zero Overlapping Bets.
A single footnote in isolation can be a fluke, and three signals saying the same thing in different words are still just one bet. We group both backtested core scouts and Darwinian ephemeral scouts into four independent research domains (corporate accounting, federal regulatory dockets, forced-selling dislocations, and supply-chain footnotes) so conviction only builds when genuinely different parts of the public record agree.
ASR) dealer-covering contracts, Schedule 13D stakes, Item 1A novel numeric/covenant density, and XBRL Non-Cash Fortress versus GAAP accrual divergences, gated by Form 8-K Item 4.02 and Form NT distress vetoes.
T+3) crash-low stabilization gate, paired with options volatility skew, 2-sweep stop confirmation, and volatility-scaled profit ratchets.
HYG/IEF) vetoes.
$0 settled margin debt, managed with symmetric [-30d, +30d] wash-sale partner rings, partial whole-share Treasury trims, ex-dividend guards, and beta-aware funding waterfalls) alongside the active idiosyncratic sleeve, protected by 10-day regime dwell and VIX contango term-structure shields.
Published System Framework
Disciplined Governance.
Audited Outcomes.
While proprietary parser definitions, feature weights, and execution microstructure remain confidential trade secrets, our econometric validation protocol and forward execution ledger are documented openly.
Point-in-Time Alternative Data & Survivorship-Free Panels
SEC acceptanceDateTime alignment, 10-digit CIK & OpenFIGI mapping, 1-Doc-1-Witness provenance & Tier A/B schemas.
Non-Parametric Calibration & Orthogonal Pillar Pooling
OOS Blom quantiles, floored MAD, concave 4-pillar pooling, One-Way Canary Glass Wall & statutory/credit vetoes.
Six-Factor Residual Attribution & Purged Walk-Forward Governance
6-dimension Ridge residuals, 5-fold Purged CV, sub-arm governance, crowding guards & volume-clock decay.
Research & Institutional Correspondence
Built for Our Own Book.
Selective Institutional Discourse.
Starling Quantitative is engineered primarily to compound Biwada’s internal capital and advance our proprietary research agenda across liquid U.S. equities and ETFs.
Starling Quantitative
A Biwada Company • Raleigh, North Carolina • Cryptographic Witness Chain (.ots)
Institutional Research & Counterparty Correspondence
We do not offer retail signal subscriptions, consumer model marketplaces, or discretionary asset management. Because several filing and event sleeves are capacity-constrained, external correspondence is limited to quantitative research discourse or selective, capacity-bounded institutional evaluation under mutual NDA, where we provide unredacted OpenTimestamps (.ots) receipts, NDA-salted blotter preimages, and regulated clearinghouse statements.