Unveiling themes in 10-K disclosures: A new topic modeling perspective
本文提出一种基于词嵌入的主题建模方法，分析10-K文件中管理层讨论与分析部分，提取主题流行度和情感两个指标，并发现它们与商业周期、历史事件及公司基本面相关，且股票市场对不同主题情感的反应存在异质性。
We analyze the topics in the Management’s Discussion and Analysis (MD&A) section of 10-K filings. Based on word embeddings, our approach identifies MD&A topics by clustering words around anchor words that broadly define potential themes. The resulting topics are interpretable, distinct, and minimally affected by noise. From the MD&As, we extract two loading series: topic prevalence and topic sentiment, both of which exhibit substantial temporal variation and heterogeneity across topics. Examining the link between MD&A topics and stock returns, we reveal that the market response to topic sentiment varies across topics and time horizons. • We introduce a topic modeling framework specialized for the Management’s Discussion and Analysis section of 10-K filings. • We extract two indicators: topic prevalence and topic sentiment. • Topic prevalence and sentiment indicators systematically vary with the business cycle, historical events, and firm fundamentals. • We unveil heterogeneous responses of stock market returns to topic-specific sentiment across different time horizons.