How a news story becomes a
measured vector.

Content is ingested from three paths. Named sources are polled directly by RSS: central-bank and statistical-agency releases (Federal Reserve, ECB, BLS), financial trade press and broadcast business news (CNBC, MarketWatch, Investing.com, FT, Seeking Alpha, OilPrice, The Register), and one aggregator retained deliberately and classed as such (Yahoo Finance). Company filings are read from SEC EDGAR's daily index, and enforcement actions from the SEC's administrative-proceedings feed. Both SEC paths are described below.

Each article is resolved to a sector by exact theme match and headline keyword, and to a source class — wire, official, filing, transcript, trade press, broadcast, aggregator — which drives the independence weighting. An article from a domain outside the classified set is excluded rather than filed under a guess.

Items are then collapsed into events by MinHash clustering over headlines with IDF verification, and each event is scored on five measures: momentum on a -100..100 scale, sentiment, influence, novelty and signal dispersion. Scores are normalised against each sector's own recent distribution, so a +60 in one sector is comparable to a +60 in another rather than reflecting differences in baseline news volume.

Each EDGAR filing carries an ACCEPTANCE-DATETIME recorded by the regulator to the second, which is the publication time, not the time we happened to fetch it. That is what makes the series replayable: re-running any past day reproduces the same events with the same timestamps.

Regulation draws on two distinct sources. Company filings flag their own compliance events — delisting notices, auditor changes, non-reliance. SEC administrative proceedings, published with the regulator's own timestamp, record enforcement brought against a company. Only what the source states is carried: the respondent, the statute and the release number, never an allegation or an outcome, which appear only in the underlying document.

• The premium wires are absent. Reuters, Bloomberg, AP, the Wall Street Journal and Barron's are not in the ingestion set. Reuters retired its public RSS feeds. GDELT's Global Knowledge Graph was evaluated as a way to reach them and then dropped: across 5,252 GDELT articles in a three-hour window it carried zero articles from any of those publishers. Coverage skews to trade press, central-bank releases, company filings and the open web.
• Influence is a partial estimate. It combines source-class credibility with a reach assumption. Observed downstream citation, which the definition mentions, is not yet measured.
• Sentiment is a lexicon, not a model. A finance-specific word list scored over headlines, in the Loughran-McDonald tradition. Where no term matches, sentiment reports as unmeasured rather than neutral. In practice it resolves on a small minority of events. Company filings dominate the corpus and their headlines are formulaic, carrying almost no sentiment-bearing vocabulary, so most events report sentiment as unmeasured.
• Momentum requires a warm baseline. Each sector is normalised against its own recent distribution, and below the minimum observation count no score is published at all.
• SEC EDGAR follows the trading calendar. A window ending at a weekend, or before the filing day opens, carries no filings for its most recent hours. So each sector vector is published with the timestamp of the interval it was actually measured over, which may be the previous trading day.
• A sector whose most recent interval holds fewer than three events reports no vector at all, rather than a number resting on a single item.
• The SEC administrative-proceedings feed carries only its most recent releases, so it cannot be reconstructed for an arbitrary past date the way company filings can. Civil actions filed in federal court are not covered at all.

Momentum (-100..100)

The direction and rate of change of coverage intensity for an entity or theme, relative to its own recent baseline. It measures news, not price.

Sentiment

The polarity of the language used about the entity in the resolved item, aggregated across the window. It is a property of the text, not a judgement about value.

Influence

A model estimate of how far an item is likely to propagate, based on source class and reach. It is an estimate, not a measurement of impact. Observed downstream citation is part of the intended definition but is not yet measured, so it contributes nothing to the current figure. In its current form influence resolves to one of a small number of values, because both terms are properties of the source class rather than of the individual item: a company filing scores the same figure every time. It is closer to a class label than to a continuous estimate, and should be read that way until downstream citation is measured.

Novelty

How much new information an item adds relative to items already ingested on the same event. High novelty means the event is not already represented in the stream. Where an event is carried by a single source — which at present is almost all of them — there is nothing for it to be redundant against, so novelty reports as fully novel by default rather than by measurement. It discriminates only in a corpus where the same event is reported repeatedly.

Signal dispersion

The degree of agreement across independent sources contributing to a composite. Low dispersion means independent sources agree on the direction of the measure. It is not a probability of a price move and carries no forecast content. Dispersion is only defined where an event was carried by more than one independent source. Where every event behind a composite came from a single source there is nothing to agree or disagree, and the measure reports as not measurable rather than as zero.

Time window

The period over which the composite was aggregated. Every composite is only meaningful together with its stated window; the same inputs over a different window will produce a different value.

• Coverage is not exhaustive. Events reported only in sources outside our ingestion set are absent from the measures entirely.
• Source-language bias. Coverage skews towards English-language and major-market reporting, which under-weights regional events.
• Latency between publication and ingestion. Measures reflect what has been ingested, not everything that has been published at that instant.
• Influence weighting is a model estimate and can be wrong for any individual item, particularly for new or infrequent sources.
• No measure is a forecast. Momentum, sentiment, influence, novelty and dispersion describe observed news, and none of them predicts the future value of any instrument.
• TrendVectors is in development. The console renders measured data from the pipeline above, with its generation timestamp and coverage shown on the page. Other surfaces remain illustrative and are still labelled as such.

The ingestion set is named in full above, because a measurement is only checkable if you know what it was measured over. Signals on the console show the source class — wire, official, trade press, broadcast, aggregator — since that is what the independence weighting uses, with a link to the original article. The set comprises: general and financial wire services; regulatory disclosure and company filing feeds; earnings, investor-day and official-briefing transcripts; central-bank and statistical-agency releases; and sector trade press.

Source composition changes over time. A measure is always computed against the sources available in its own window.

TrendVectors SL publishes quantitative measurements of publicly available news. It does not provide investment advice, personal recommendations, or portfolio management, and it is not authorised to do so.

TrendVectors SL holds no position in any instrument referenced anywhere on this site. Figures on the console and home page are measured output from the pipeline described above, each carrying its generation timestamp and coverage. Surfaces still using illustrative data are labelled as such.

Questions about this methodology: contact@trendvectors.com.