Answers are grounded in this paper and cite their sources.
Naive alternative-data backtests routinely overestimate live out-of-sample alpha due to three structural data leaks: (1) aligning corporate filings by fiscal period-end (fp) or calendar filingDate, which introduces lookahead bias whenever disclosures arrive after the 16:00:00 ET market close and silently overwrites historical states upon subsequent restatements; (2) evaluating signals over surviving index constituents rather than historical registrant unions; and (3) passing historical text through commercial large language models whose pre-training cutoffs post-date the evaluation window. We document the deterministic, zero-LLM ingestion architecture of Starling Quantitative (https://starlingquant.com) across 3,607 historical Russell 3000 SEC Central Index Keys (CIKs) from January 2019 through November 2026, explicitly retaining 555 non-surviving and delisted issuers (15.4%) with automatic -30.0% Shumway (1997) performance-delisting haircuts. Every structured XBRL item and primary filing is joined to its exact down-to-the-second SEC EDGAR acceptanceDateTime, mapped to canonical 10-digit CIK and composite point-in-time OpenFIGI identifiers, stamped with One-Document-One-Witness (source_artifact_id) provenance, and quarantined to live forward shadow testing whenever a vendor feed relies on mutable watermarks. Finally, we specify our dual institutional delivery schemas: Feed Tier A (pre-quorum ScoutPitch feature envelopes) and Feed Tier B (post-quorum calibrated MasterDecision consensus). Canonical Web Monograph: https://starlingquant.com/guides/point-in-time-alternative-data-and-survivorship-free-panelsComplete Technical Specification: https://starlingquant.com/llms-full.txt
No takes yet. Share an insight, caveat, or question.
Starling Quantitative Research (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: