10-K Sentiment Analysis: Full Text vs. Risk Factors for Financial Prediction

Traditional financial sentiment analysis often focuses on news articles, overlooking the rich data within 10-K filings, particularly for predicting volatility. This research introduces a supervised lexicon-learning approach tailored for 10-K documents, including their Item 1A risk-factor sections. The sentiment scores are trained against both return and volatility labels across three aggregation levels: sector, portfolio, and individual firm. Analyzing 1,383 filings from 94 Nasdaq-100 technology companies between 2006 and 2023, the study evaluated twelve sentiment metrics. Key findings indicate that full-filing text yields more accurate sentiment for sector and portfolio-level predictions concerning both returns and volatility. However, at the individual firm level, the more focused Item 1A section proves superior, a phenomenon attributed to the interplay between document volume and available training signals. The research also highlights the limitations of a standard Loughran-McDonald dictionary baseline, which consistently showed a strong negative correlation with price, underscoring the necessity of a supervised approach for regulatory disclosure text. These results provide a foundational methodology for developing more sophisticated, multi-source financial sentiment systems.