Abnormal returns and volatilities around the filing of 8K reports and the performance of large language models in finance

Abstract

Public companies listed in the US must report specific events determined by the SEC in an 8K filing. The events that must be reported on are what the SEC considers “material” to investors, meaning that the SEC believes knowledge of these events is required for investors to be able to accurately value a company. Examples include “1.03 Bankruptcy or Receivership,” “2.01 Completion of Acquisition or Disposition of Assets,” and “4.01 Changes in Registrant’s Certifying Accountant.”

In this paper I demonstrate that 8K reports can be used to predict the filing company’s abnormal returns and abnormal volatilities during the first two weeks after a filing. I also show that newer transformer-based large language models perform better than previously used bag-of-words models in textual analysis on this dataset.

To do this I gather a novel dataset of 551 501 8K reports spanning the years 2000-2021 for 4753 listed companies. For this I make a program in the Python programming language that automatically gathers, cleans, and parses reports from the SEC’s EDGAR database. I combine the data in the reports with return data from CRSP for each filing company for the same period.

I find that companies tend to earn abnormal returns after making a filing regardless of the content of the filing. A filing company also has abnormal volatility during the days around the filing. I also find that certain event categories cause predictably negative or positive abnormal returns. Most event categories predict decreased abnormal volatility.

When utilizing textual analysis of the reports instead of just the event category I find that bag-of-word models can pick up on some of the sentiment in a report, but a transformer-based model can do this better. The sentiment from textual analysis provides less information about abnormal volatility than the first two models.