How is an earnings call different from a news article about the same quarter? This note separates the two texts, using a Microsoft quarterly call as the public example, and explains why a model should not treat them as one pile of words.
An earnings call is a scheduled conversation between a company and the people who cover it. Prepared remarks come first. Questions come second. The company chooses much of the wording, and the speakers return every quarter to a familiar script. A news article is a later document, written by someone outside the company, that selects, compresses, and sometimes disputes that conversation. Both documents are about the quarter. They are not the same evidence.
The timing differs as well. The call has a start time. A transcript may appear after a delay. A news article may appear during the call, after it, or, in the case of a preview, before it. A study that dumps every paragraph mentioning the company into one bag has mixed a primary statement, a delayed transcript, and a second writer's summary. Any score built from that bag is partly a score of the writers, not only of the company.
An earnings call is the company's scheduled remarks and answers. A news article is a later selection of that event, written by someone else, and it is a different document.
Take a Microsoft quarterly call as a concrete public document. The prepared remarks usually restate revenue categories the company has already defined, then add color a filing does not. The question period is less scripted. An analyst asks about a segment, and the answer may confirm, narrow, or dodge the question. A researcher who wants the company's own categories should read the remarks and the filing together. A researcher who wants unplanned language should mark the question period and keep it separate. Averaging them produces a third text that nobody spoke.
A news article published after that call does another kind of work. It may lead with the number that moved the stock, quote one sentence, and ignore the rest. That selection is informative about what the writer thought mattered. It is a poor substitute for the call if the question is what management said. It is also a poor substitute for the filing if the question is what the company reported under the accounting rules. The article can be its own series. It should not be silently labeled as the call.
Numbers in the two documents deserve the same caution. The call may emphasize a figure the filing presents differently, or speak about a constant-currency comparison the financial statements do not lead with. A language model asked to pull figures from the call will find some of them. Those figures still need a check against the filing before anyone treats them as the record. The call is speech. The filing is the document the company submits. A note that quotes the call should say so.
A practical file has one row per document, not one row per company-day. The row names the type, the source, the event time, and the time the text became available. Microsoft's call, the transcript, and a later article are three rows. A feature built only on transcripts can be compared with a feature built only on articles. If they disagree, the disagreement is a finding about the sources. If they are blended first, the disagreement is invisible.
None of this requires a judgment about whether the quarter was good. The separation is clerical and then analytical. Clerical, because the documents are different objects. Analytical, because a model that reads text will imitate the mix it is given. Give it articles, and it learns the habits of articles. Give it calls, and it learns the habits of calls. Say which one the series is. The Microsoft example is only a way to keep that sentence specific.
This piece does not claim that calls are truthful, that articles are misleading, or that a text score should change a position. It claims only that the two sources are not interchangeable, and that a study has to say which one it used.