SEC Filing Text Analysis

A 10-K annual report runs anywhere from thirty to several hundred pages, and the financial statements at its core are only a small fraction of that. The rest is prose: a description of the business, a long "Risk Factors" section, management's discussion and analysis of results, and legal disclosures. Most of that text is boilerplate, copied nearly verbatim from the prior year's filing, which is exactly what makes it useful — because the parts that change stand out, and those changes tend to matter more than they should to a market that mostly skims the document.

The core technique is a year-over-year text comparison. A model diffs this year's 10-K against last year's, sentence by sentence or section by section, and measures how much language has changed. A company whose Risk Factors section barely changes from one year to the next is telling you, implicitly, that nothing new has come up. A company that rewrites a third of that section — adding new risks around supply chains, litigation, or customer concentration — is disclosing something material through the act of rewriting, even before you read what the new language says. Research going back over a decade has shown that filings with large amounts of textual change tend to be followed by weaker returns and higher volatility than filings that look nearly identical to the year before, and that this effect is strongest for change buried in Risk Factors and MD&A rather than in dry legal boilerplate elsewhere in the document.

A second technique scores the tone of the language itself using finance-specific word lists (ordinary sentiment dictionaries misclassify accounting words like "tax," "cost," and "liability" as negative, so finance NLP uses lists built specifically from filings). A rising count of negative or uncertainty-flagged words, or an increase in complexity (longer sentences, more passive voice, more hedged phrasing like "may" and "could"), has been linked to subsequent weaker performance, on the theory that management's own choice of words leaks information that the numbers alone don't capture.

For example, a retailer's 10-K adds a new risk factor paragraph about "increasing dependence on a small number of suppliers in a single region" that wasn't present the year before. A text-change model flags this section as newly added; a human analyst reading only the income statement might not notice it at all, since it doesn't show up as a line-item change until a disruption actually hits.

What this means in practice

SEC filing text analysis works best as a screening layer — flagging filings with unusual textual change or rising uncertainty language for closer human review, or as one input feature among many in a broader quantitative model, rather than a standalone trading signal. It requires careful handling of filing structure, since section headers and formatting vary across companies and change over time, and any pipeline needs to be re-validated whenever the SEC's filing system (EDGAR) changes its format.