How news events about companies are classified into event types

News events about companies are classified into event types by two routes: a rule applied to a disclosure that already carries a code, and a reading of the free text of an announcement. This post sets out both, shows how two official schemes sort disclosures, lists what a taxonomy has to settle, and gives a test with two readers to run on a sample of any feed.

How are news events about companies classified into event types?

News events about companies are classified into event types in one of two ways. A rule maps a code the publisher attached, such as the item number on a regulatory filing, to a type. Or a model or a person reads the title and text of an announcement and chooses a type from a fixed list.

The routes differ in who makes the call. On a coded disclosure the filer picks the code, under a regulator's definitions. On free text the classifier picks the type, under the taxonomy's definitions. A newsroom post or a private company's release arrives without a code, so a feed that reads both kinds of source uses both routes, and each label should say which route made it.

Company news event classification: event types for earnings, a merger or acquisition, a product launch, regulatory and management changes

A working list of event types for company news covers earnings, mergers and acquisitions, product launches, regulatory or legal matters and management changes, and adds funding, partnerships, expansion, restructuring and incidents. Official schemes name part of that list: the SEC's item titles name results, completed acquisitions, changes in control and officer changes, and none names a product launch.

The SEC's scheme: the items of a current report

In the United States a listed company reports specified events on Form 8-K, the current report, and the form's items are the classification. The scheme is the Commission's own: Release 33-8400, effective 23 August 2004, added eight items to the events that require a current report and grouped the required items by topic, in numbered sections. The titles below are those in the interpretations of the form published by the staff of the Division of Corporation Finance. The left column is this post's wording.

Two points matter when you map items to event types. An item is a legal category and can be wider than the event: the title of Item 2.01 takes in a disposition as well as an acquisition, and that of Item 1.01 names no kind of agreement. And no title names a product launch. The 2004 release describes Item 8.01 as the item permitting voluntary disclosure of other events, so a rule that sees only that number learns that something was reported, and no more.

The EU's scheme: classes of regulated information

In the European Union the list sat until 2026 in Commission Delegated Regulation (EU) 2016/1437. Its Article 9, which applied from 1 January 2017, had each officially appointed mechanism classify all regulated information by the list in the annex: three classes, periodic, ongoing and additional regulated information, with sub-classes that followed the legal basis of the disclosure. Commission Delegated Regulation (EU) 2026/971, which applies from 10 July 2026, repealed it and draws the common list from the types in the annex to Implementing Regulation (EU) 2025/1338 that relate to Directive 2004/109/EC. Among them are the annual financial report, the major shareholdings notification and a single type for inside information.

Whatever the business event behind it, a disclosure of inside information carries that one type. Both schemes are built around what a company has to disclose, which is the regulator's question. A revenue team looking for a trigger event, or a research desk, asks what happened, and reads the answer from the text.

A taxonomy of event types for financial news classification with NLP: what it has to settle

A taxonomy of event types for classifying financial news with NLP has to settle five things before a model reads a word: whether one announcement can carry several types, how fine the types are, what happens to an item that fits none, whose event counts, and the version under which every label is made.

• Several types on one announcement. The regulator's own record allows it: the 2004 release says General Instruction D to Form 8-K lets a company meet several items with one report, so long as it names each applicable item by number and caption and gives the disclosure each requires. A press release can likewise announce an acquisition and its funding. Decide whether a record holds one type or a list, and if one, the order of precedence.
• The level of detail. A type has to be narrow enough to act on and common enough to test. Splitting acquisitions by direction, the company buying or the company bought, thins the sample: the example below shows how little sixteen graded items settle.
• A residual class and an empty outcome. Official schemes keep a residual: Item 8.01 is titled Other Events, and the EU list keeps a type for additional regulated information. A taxonomy for text needs one too, and an outcome of no type for an item that reports no event.
• Whose event it is. A newsroom also carries items about customers, partners and the industry. Assign a type only when the item is about the company itself, and say which side of a deal the company is on.
• A version. Freeze the list, the definitions and the rule or model that applies them under a name, and store that name on every record as its label version.

One rule sits above the five: an event type describes the company. A management change is a company event, a role filled or vacated on a date. Record the role, the direction of the change and the link to the company's statement, and collect nothing about any individual.

What changes when the taxonomy changes

Counts by type break at the change, and history has to be read under two schemes. The 2004 release moved the voluntary disclosure of other events from the former Item 5 to Item 8.01, and noted that a search of EDGAR spanning the change has to look for both numbers. It kept the new numbers clear of the old single-digit ones, to avoid confusion about an item's subject. On 26 July 2023 the SEC adopted rules creating Item 1.05 for a cybersecurity incident a registrant determines to be material, with disclosures due beginning the later of 90 days after publication in the Federal Register or 18 December 2023. A count of filings under that item begins when the item does.

Three practices follow, as this post's advice. Never give a retired name or number a new meaning. Store the version beside every label and count records by version before comparing two periods. When a type is split or added, expect a step in the series at that date, and relabel the earlier records or compare within one version.

How to test an event classification on a sample

Grade a sample by hand and report agreement type by type, with an interval. The statistics come from official guidance on another yes or no call: the FDA's statistical guidance on reporting results from studies evaluating diagnostic tests, issued by its Center for Devices and Radiological Health on 13 March 2007 as nonbinding recommendations. It includes the case of a test compared with a non-reference standard, which is where a classifier graded by a human reader stands. Applying it to event types is this post's advice.

Put the feed in the place of the new test and the reader in the place of the standard. For one event type, count four groups of items: a, where both gave the type; b, the feed alone; c, the reader alone; d, neither. On those counts the guidance says this.

• Overall percent agreement is 100% × (a + d) / (a + b + c + d).
• Positive percent agreement is 100% × a / (a + c), and negative percent agreement is 100% × d / (b + d).
• All agreement measures, kappa included, have two major disadvantages: two methods can agree and both be mistaken, and agreement moves with prevalence, the relative frequency of the condition in the group.
• In some situations overall agreement can be good while positive or negative agreement is very low, so the FDA discourages using it alone.
• Results altered by discrepant resolution, a second stage applied only where the two disagree, should not be used to estimate agreement.
• The FDA recommends reporting measures of agreement with their two-sided 95 percent confidence intervals.

The test itself is this post's advice.

1. Fix the sample. Take a closed period and the companies you track, and draw items at random, with a type or without.
2. Use two readers. Give each the definitions and the record's evidence: title, text and link. Hide the feed's label and the other reader's answers.
3. Grade events. For a management change a reader marks the type and the role. No name goes on the sheet.
4. Count a, b, c and d for each type you rely on, feed against reader and reader against reader.
5. Compute the three agreement figures, and a / (a + b), the share of the feed's calls a reader confirmed.
6. Put an interval on each share. Section 7.2.4.1 of the NIST/SEMATECH e-Handbook of Statistical Methods gives the Wilson interval, whose lower limit cannot be negative.
7. Keep every grade as first given, disagreements included.
8. Repeat with the same design for each feed and label version, since agreement moves with the mix of items.

Read the feed's figures beside the readers' own. Where the two readers part ways, mend the definitions first.

Example: two readers grade a sample for Acme Robotics

This example is illustrative. Acme Robotics is an invented company, and every figure below is an assumption, not a measurement of any feed.

Assumptions:

• The data team at Acme Robotics draws 200 items at random from the announcements a feed delivered in one closed quarter for the companies it tracks.
• Readers A and B grade every item for one event type, acquisition made, without seeing the feed's label or each other's answers.
• Against reader A, the feed and the reader both gave the type to 12 items, the feed alone to 6, the reader alone to 4 and neither to 178.
• Reader B and reader A both gave the type to 13 items, B alone to 3, A alone to 3 and neither to 181.
• Intervals are Wilson intervals at 95%, with z = 1.96, p the observed share and n the number of items behind it.

interval = ( p + z^2/(2n) ± z × sqrt( p(1-p)/n + z^2/(4n^2) ) ) / ( 1 + z^2/n )

1. Check the counts. 12 + 6 + 4 + 178 = 200 and 13 + 3 + 3 + 181 = 200. Reader A gave the type to 12 + 4 = 16 items and withheld it from 6 + 178 = 184; the feed gave it to 12 + 6 = 18. Reader B gave it to 13 + 3 = 16.
2. Interval for the feed's positive agreement, p = 0.75 and n = 16. z^2/n = 3.8416 / 16 = 0.2401, so the divisor is 1.2401. The centre is (0.75 + 0.12005) / 1.2401 = 0.7016. The half-width is 1.96 × sqrt(0.011719 + 0.003752) / 1.2401 = 0.1966. The interval runs from 0.505 to 0.898.
3. Interval for the feed's overall agreement, p = 0.95 and n = 200. The same steps give 0.910 to 0.973.
4. Read the two together. Overall agreement of 95% rests on the 178 items neither gave the type. On the 16 items reader A called acquisitions the feed matched 12 and reader B matched 13, so the feed stands one item from a second reader. The interval on 75% runs from about 51% to 90%: sixteen items settle little.
5. Sample by type to narrow it. If the team grades 100 further items the feed typed as acquisitions and a reader confirms 80, the share confirmed is 80%, with an interval from 0.711 to 0.867. That sample measures the feed's calls only: missed items still come from the random draw.

The 13 event types in Company News

Company News delivers company announcements and regulatory disclosures classified into 13 event types: product launch, partnership, acquisition made, acquired or merged, funding round, leadership change, layoffs or restructuring, expansion, incident, financial results, award or recognition, regulatory or legal, and other.

Both kinds of source are in the dataset: announcements from each company's own newsroom and regulatory disclosures. An event type describes an event of the company itself, and a record can carry more than one. Every record carries a named version, and a breaking change ships as a new version with at least 90 days' notice.

Each record carries the evidence two readers need: the title as published, the date the company published it, an excerpt of the company's own text and the address of the original. A regulatory disclosure adds its item numbers and its accession number. A leadership change is recorded as a company event, with the role concerned, such as chief financial officer or board, and the direction: appointed, departed or succession.

The data dictionary defines each field and each event type, and event-driven research with Form 8-K filings and newsrooms puts the types to work.

Frequently asked questions

The queries and code on this page are examples to adapt. Test them in your own environment before you rely on them.

What this page says about the products it names was checked against their public documentation on 7 October 2026. Product and company names are trademarks of their owners. Fokals is not affiliated with them or endorsed by them.