Bedrock AI whitepaper shows underutilization of 10-Ks
Companies with high Bedrock AI Risk Scores significantly underperform
The average 10-K is longer than Shakespeare’s Hamlet. It’s a shocking observation. A three hour time commitment means that even the most expert investors are unable to read these documents in detail each year. Consequently, information disclosed in 10-Ks is not incorporated into stock price until months after the filing deadline.
The Bedrock AI whitepaper demonstrates that companies with “risky” 10-Ks underperform the rest of the market significantly and consistently, over the mid to long-term. High risk 10-K filers (top 7 percent by Bedrock AI Risk Score) underperformed the remaining filers in the test by 11 percent after 24 months based on median ROI across all years. Consistent performance differentials of 5 to 11 percent are evidenced beginning 9 months after the 10-K is filed. Performance differentials are consistent each year in the test, with the exception of 2020. In 2020, performance differentials were not evident until 15 months after the filing of the 10-K, due in part to the impact of the timing and volatility of the stock market crash in February 2020. Performance differentials between high risk and lower risk companies magnify as the time horizon increases. Refer to the attached whitepaper for details on approach and testing.
The fact that such significant performance differentials can be derived from publicly available information contradicts efficient markets theory. These differentials exist because impactful information in securities filings is hard to find, hard to understand and buried within an extraordinary amount of noise.
The Bedrock AI models are trained to find associations between content in unstructured text and downside outcomes like SEC enforcement actions. Because we use deep learning based language models, our algorithms do not have to rely on key words but rather have a more human understanding of content. Our models rely on textual indicators, like accounting policy choices, aggressive adjustments to non-GAAP metrics, management turnover, related party transactions, regulatory troubles and more, to predict long-term outcomes. Companies that take an aggressive approach to earnings management often outperform in the short-term but, earnings management is unsustainable, and eventually performance will reverse, often dramatically. In the early years of aggressive earnings management or outright fraud, it is close to impossible to detect the risk using financial metrics. Manipulated financial metrics purposefully look normal or normal enough to evade detection early in the earnings management cycle. Textual indicators are essential to the early detection of troubled financial fundamentals. Often these textual indicators are buried deep in a 10-K and embedded in dense accounting verbiage.
Our models are able to understand the interaction of risk in multidimensional space, whereas humans struggle to effectively quantify risks when there are many variables at play and instead tend to rely on heuristics or “gut feel”. Additionally, our models process and understand a 10-K in a matter of seconds, whereas an expert human would take approximately three hours to perform a similar task. These two capabilities allow us to a) read information that most capital market participants never see and b) understand how disparate risk-based information comes together to predict the likelihood of crisis.
The results of our whitepaper demonstrate that valuable information contained in securities filings is underutilized and that this information can be leveraged using advanced natural language processing techniques.
Bedrock AI Whitepaper 12-2021.pdf