News, events and company disclosures

Document-to-event normalization

A document is a source artifact; an event is a structured interpretation linked to one or more artifacts. Clustering groups related documents while preserving their original identities. Syndicated copies belong to one evidence family rather than counting as independent confirmation.

An event records entity, type, supporting spans, numerical facts with units, source times and interpretation uncertainty. Corrections create new revisions linked to the earlier interpretation. A classifier label does not become a source assertion unless the cited text supports it.

Event schema fragment

type FinancialEvent = {
  id: string;
  revision: number;
  type: string;
  entity_refs: string[];
  event_at: string | null;
  available_at: string;
  evidence_refs: string[];
  facts: Array<{ concept: string; value: string;
    unit: string; period: string | null; evidence_ref: string }>;
  supersedes: string | null;
  coverage: "metadata_only" | "excerpt" | "full_text";
};

Standardized release surprise

The statistic uses consensus captured before release and a historical surprise standard deviation estimated from earlier eligible releases. Values, periods and units match. Missing consensus or a zero denominator produces an undefined statistic. Later revised figures remain separate from the original release.

Filings and financial facts

A reported fact is indexed by issuer, filing, acceptance time, taxonomy concept, unit and reporting period. Instant facts describe a point in time; duration facts describe an interval. Amendment and context-selection rules determine which revision is eligible for a calculation.

The filing reader accepts an identifier returned by the filing-list operation rather than an arbitrary URL. Text extraction reports truncation and coverage. A headline-only observation therefore remains distinguishable from a complete filing or article.

Collection and exposure mapping

Continuous collection, event subscriptions and portfolio-exposure joins record source coverage, refresh behavior and mapping uncertainty. A geopolitical headline has no universal trade direction: the impact depends on the instrument, current price, costs and existing exposure.

Exposure-weighted event representation

E_t is the set of eligible event records, x_k is a declared numerical event feature, and M is the time-valid event-to-instrument exposure map. z is an exposure-weighted feature, not a causal return estimate. Binary membership, revenue share and model-estimated sensitivity have different meanings and are not interchangeable weights.

Event time and revision-aware surprise

A release event separates the period being reported from the instant when its information became available. Consensus is frozen before that release; the first reported value and subsequent corrections have distinct revisions. This prevents a revised historical value from being compared with a forecast that could only observe the original release.

The surprise denominator uses earlier eligible releases with compatible units and measurement definitions. A change in taxonomy, seasonal adjustment or reporting basis can break comparability. Such a break changes the feature’s eligibility or method revision instead of being absorbed as an ordinary market surprise.