In investing as in life, numbers are important—but context matters too. Research using large language models and deep learning could help investors squeeze more value out of company financial statements while also advancing the application of machine learning, write Chicago Booth PhD student Alex Kim and Booth’s Valeri Nikolaev.
The US Securities and Exchange Commission requires publicly traded companies to include a management discussion and analysis of quarterly and annual financial statements. The researchers focused on analyzing what value the MD&A narrative content adds to the raw numbers.
Kim and Nikolaev used BERT, a large language model developed by Google and designed to capture the context-based meaning of words and sentences, to analyze thousands of annual MD&As from 1995 to 2020. By then using an artificial neural network to combine the extracted narrative context with the numbers from the company financial statements, the researchers were able to quantify how important the interactions between the numbers and the context were for projecting earnings, cash flows, and returns.
Using the numbers on their own to make these predictions had value, find the researchers, who also found value in projections that incorporated only the contextual information. But a model they developed that allows the full interaction between context and numbers achieved a higher accuracy, significant enough to move the needle for investors. This model, says Kim, is analogous to how the interconnected neurons in our brain process information. And the model’s superior performance, he explains, confirms that considering numbers with their corresponding contextual information—as many investors may already do, albeit in a more manual and less sophisticated way—is worthwhile in prediction tasks.
With this finding as a stepping stone, the researchers then incorporated the idea of number-context interactions into an asset pricing model. They did this by homing in on profitability, which is one of the factors foundational to investing identified by Chicago Booth Nobel Laureate Eugene F. Fama and Dartmouth’s Kenneth R. French. In 1993, Fama and French named three factors (market beta, size, and value) that could explain average excess stock returns, and then added profitability and investment as factors in 2015. The Fama-French factors underpin quantitative investing, and extensive scholarship and real-world experimentation has gone into fine-tuning them. (For more about investable factors, read “The 300 Secrets to High Stock Returns.”)
Their earlier insights top of mind, Kim and Nikolaev developed a context-adjusted profitability measure for their model. This refined version, when applied to the 1995–2020 data, yielded more accurate share-value forecasts than traditional profitability measures did, including for forecasts that predicted returns several years into the future.
The findings may help explain a shortcoming of the five-factor model, the researchers write. Fama, French, and subsequent scholars and practitioners have struggled to explain why the profitability factor at times loses its predictive power, especially for small companies. Context can help explain why smaller businesses could record positive returns regardless of their low profitability, or vice versa, says Kim. When a company is small, it may be young and less profitable because it is investing in its future.
Ideally, investors would use a company’s future profitability to predict its future stock returns. But as future profits are of course unknown in the present, they use current profitability as a rudimentary proxy—and may recognize the inherent disconnect. “Investors in early-stage companies may read through their low profitability,” Kim explains. “They might believe the future projects are likely to succeed, and invest in such companies despite low profitability values. It could be the case that small companies with low profitability have good market returns because of potential”—a concept that’s hard to pick up from numbers alone.
With narrative information creating a richer profitability measure, the researchers’ model is able to produce better forecasts. It’s a big leap forward from the days when intrepid analysts would scour quarterly reports manually as they searched for usable information, and it’s one that the researchers suspect will lead to investors developing other sophisticated models tailored to meet specific investing needs.
An expert panel discusses the yet-unknown and steps policy makers should take.
How to Design Policy in the Age of AI
Artificial intelligence is proving adept at technical investing.
How Self-Driving-Car Technology Can Help Machines Trade Stocks
End the requirement for quarterly reporting.
How to Curb Short-Termism and Boost the US EconomyYour Privacy
We want to demonstrate our commitment to your privacy. Please review Chicago Booth's privacy notice, which provides information explaining how and why we collect particular information when you visit our website.