Event classifiers on SEC filings

We ask 529 questions of every sentence a public company files.

Each classifier answers one precise, economically meaningful question — is demand softening? did a trial miss its endpoint? are reserves being upgraded? Every answer carries a direction and a historical information coefficient, point-in-time across 15 years. Read them one at a time to build a thesis, or in bulk as model features.

One 10-Q, read by every classifier signal stack · acceleration setup

▲is_backlog_building“Contracted backlog rose to a record, with bookings outpacing revenue.”0.94

▲is_pricing_power_expanding“We implemented list-price increases that customers absorbed without volume loss.”0.91

▲is_market_share_gaining“Wins against incumbent suppliers drove share gains in our core segment.”0.88

▲is_cost_structure_improving“Facility consolidation lowered fixed cost per unit versus the prior year.”0.86

Four independent events, one filing — a coherent acceleration read the market hasn't priced from a routine quarterly. The same stack of negatives — softening demand, margin compression, aging receivables — is a deterioration read.

The questions, by family.

Classifiers are grouped by the economic story they tell. These are real production classifiers — the arrow is the alpha direction we assign each one.

Demand & pricing

Revenue quality, ahead of the print

▼Is this company's demand softening — in the MD&A, before guidance?is_demand_softening
▲Is it successfully raising prices?is_pricing_power_expanding
▲Is backlog / contracted revenue building?is_backlog_building
▲Is it taking share from incumbents?is_market_share_gaining

Margins & efficiency

Where operating leverage turns

▼Are operating margins compressing?is_operating_margin_declining
▲Is the cost structure structurally improving?is_cost_structure_improving
▲Is utilization / billable capacity rising?is_utilization_improving
▼Are receivables aging, collections slowing?is_dso_deteriorating

Distress & red flags

The signals that lead a re-rating

▼Did they just announce a restatement?is_restatement_announced
▼Did the auditor change?is_auditor_change
▼Are key executives heading for the exit?is_management_exodus
▼A Nasdaq minimum-bid delisting notice?is_nasdaq_bid_price_delisting

Capital & governance

Balance-sheet and control events

▲Was debt refinanced on better terms?is_debt_refinanced_favorably
▼Did they adopt takeover defenses / a poison pill?anti_takeover_poison_pill
▲Are institutions building a position?is_institutional_investment
▼Did they effect a reverse stock split?is_reverse_stock_split_effected

Native to every sector.

A generic sentiment score can't tell a dry hole from a discovery. Because the classifiers are trained on the language of each industry, they read the drivers that actually move a name — a reserve upgrade for an E&P, a missed endpoint for a biotech.

Biotech & pharma

is_trial_endpoint_missed · is_clinical_hold · is_interim_data_positive · is_orphan_drug_designation · is_fda_breakthrough_device

Energy & E&P

is_proved_reserves_upgraded · is_reserve_replacement_positive · is_reserve_life_improvement · is_drill_results_positive

Metals & mining

is_maiden_resource_estimate · is_resource_estimate_upgraded · is_resource_expansion_drilling

Banks & financials

is_allowance_for_loan_losses · is_non_accrual_loan_policy · is_loan_deposit_growth

Consumer & retail

is_same_store_sales_declining · is_pricing_pressure · is_demand_softening

Software & SaaS

is_bookings_deceleration · is_pipeline_softening · is_same_client_revenue_growing

Two ways to use them.

Mode 01

Build the narrative

Pull every classifier that fired on a name over a window and read the stack. Four positive events on one quarter is an acceleration thesis; a cluster of red flags is a warning you can act on before the restatement 8-K.

Every hit deep-links to the exact sentence in the primary EDGAR filing — full provenance, so an analyst can verify the read in one click.

Mode 02

Feed the model

Each classifier is a feature with a sign (direction) and a weight (ic, historical information coefficient). Aggregate sentence scores to the filing, then across a window, for a point-in-time company factor.

Scores are frozen and versioned — a backtest returns exactly what production returned then. No lookahead.

The corpus.

529

Production classifiers

270M

Sentences scored

744M

Classifier hits

5.5M

Filings covered

15YR

History, point-in-time

~47S

Median accept → scored

159 built from investment theses — a named question with a name you can read. 370 discovered unsupervised — recurring events surfaced from the data that no analyst thought to ask about. Both meet the same bar: they have to work on real sentences the model never saw, not just separate a training set. Meaning, not keywords — three sentences with no shared vocabulary that describe the same event all fire the same classifier.

Straight about what it is.

Features, not forecasts. Each classifier ships a direction and a historical IC. Individually the ICs are small — that's expected. They're inputs you weight, not a return promise.
Score is confidence, not magnitude. It's the probability the event is present in that sentence — never a price target.
Coverage is US SEC filings. Not news, not transcripts. Every hit resolves to a primary EDGAR document.