Research Paper — Precrime: Structural-Metadata Screening of Pre-Financial Securities Issuers on EDGAR
Mark Phillips · June 2026. The position/survey paper describing the design and methodology of the precrime / fraud-heuristic system this site runs on.
Download / view the full paper (PDF, ~780 KB) →
Abstract
This paper publishes the design and methodology of Precrime, a software system that screens securities issuers from raw EDGAR structural metadata — before any analyzable financial statement exists. Securities fraud is its most consequential application, but the screen's scope is broader: the same structural signals serve as predictors of corporate manipulation, of abnormal price trajectories (the engineered distribution and subsequent collapse), and of entity duration — how long a shell or microcap issuer is likely to persist before going dark or being absorbed.
The method is a feedback loop running in reverse of the academic literature: rather than fitting a classifier on filed financials, we begin from the outcome — the corpus of published DOJ and SEC enforcement pleadings — and work backward to the structural fingerprints those matters left in the filing record, encoding them as weighted heuristics that then run forward as a predictor over live filing feeds. Postcrime labels train precrime flags. The labeled, enriched learning corpus is published openly so that third parties can inspect, reproduce, and extend the pipeline.
The heuristic catalog is grounded most deeply in a primary-source anchor, Exemplar 1 — an enterprise running schemes in parallel across more than 650 entities and CIKs — drawn from sealed federal pleadings that afford a paragraph-level view of a sustained manipulation operation that drew no regulatory or criminal intervention over decades.
Key figures
| Metric | Value |
|---|---|
| Labeled corpus | 37,619 DOJ-SDNY + SEC litigation-release documents |
| Hand-traced cases | 3 (RTSL, SGR Energy, GP Solutions) |
| Mean lead time, first structural signal → enforcement | 6.6 years |
| Exemplar 1 scale | 650+ entities / CIKs run in parallel |
| Detection families surveyed | 4 (Precrime occupies the empty pre-financial position) |
| Watchlist re-evaluation cadence | 18 months |
All quantitative figures in the paper are illustrative projections on synthetic data, not measured results. The paper is a position/survey publication of the design, not a benchmark of deployed performance.
Scoring model
The screen sums weighted heuristic firings into a score S(e) for each entity e, gated on a
minimum number of distinct signals and a score threshold, then bucketed into tiers:
S(e) = Σ wₕ · 𝟙[h fires]- Gate:
nsig ≥ 2andscore ≥ τ(τ = 3) - Tiers: T1
≥ 10· T26–9· T33–5
See the precrime overview, heuristics detail, and scheme-detection methodology for the full catalog and the case walk-throughs.