Predicting Firm Financial Performance from SEC Filing Changes Using Automatically Generated Dictionary

Textual data are increasingly used to predict firm performance, however extracting useful signals towards serving this goal with a continuously growing repository of financial reports and documents is challenging, even by the state-of-the-art machine learning and natural language processing (NLP) techniques. We propose a novel approach to automatically create a word list from SEC filings (10-K and 8-K reports) using advanced deep learning and NLP techniques and compare their performance against the widely used Loughran–McDonald sentiment dictionaries. We additionally analyze a corpus of 8-K and 10-K documents to evaluate their relative informativeness for firm performance prediction. Since 8-K filings provide corporate updates along a fiscal year, we compare their content against changes in 10-Ks between consecutive years to assess the incremental value of information provided in these regulatory filings. Information effectiveness is examined by predicting six key financial indicators for a set of US banks using ridge regression. Our results positively support sentiment dictionaries expansion by automatically extracting meaning from text and highlight the benefits obtainable from utilizing update filings.