New ML study reveals which parts of 10-K filings predict stock moves best

A new machine learning paper from Sanggyu Sean Choi tackles a gap in financial sentiment extraction: how much of a 10-K filing actually matters for predicting market outcomes. Most prior work focused on news text or used return labels alone, ignoring volatility—the very thing risk disclosures are designed to inform. Choi extends a supervised lexicon-learning approach to both the full 10-K text and its Item 1A risk-factor sections, training sentiment scores against both return and volatility labels at three levels of aggregation: sector, portfolio, and individual firm. The dataset covers 1,383 filings from 94 Nasdaq-100 technology constituents spanning 2006 to 2023, and produces twelve distinct sentiment metrics evaluated on classification accuracy, correlation with realized market outcomes, and qualitative lexical content.

Results show a clear aggregation-dependent pattern: full-filing text generates more accurate sentiment at the sector and portfolio level for both returns and volatility, but this reverses at the individual-firm level, where the focused Item 1A section performs better. Choi attributes this to the interaction between document volume and the amount of independent training signal available at each aggregation level. Notably, the widely-used Loughran-McDonald dictionary consistently shows a strong negative correlation with price at every level, underscoring that a supervised approach is essential for regulatory disclosure text. These findings establish the sentiment-generation methodology behind a subsequent, larger-scale multi-source system, giving quants and analysts a data-driven roadmap for extracting signals from SEC filings.