Quick Buzz Feed
LLMs can account for 20 percent of moves following earnings announcements . . . so far.
For many decades, investors and market participants have relied on the 'earnings surprise' as a primary tool to understand how stock prices react to a company's financial results. This metric, calculated by comparing reported earnings per share against analysts' consensus forecasts, is intuitively appealing as it indicates whether a company exceeded or fell short of expectations. However, despite its widespread use and straightforward nature, this traditional measure has proven to be quite limited in its explanatory power, accounting for only a meager 5 percent of the same-day stock movements observed following these crucial earnings announcements. This highlights a significant gap in our ability to fully comprehend market dynamics based solely on numerical earnings data.
Groundbreaking new research conducted by Chicago Booth’s Ralph S. J. Koijen and Bradford Levy has unveiled the remarkable potential of artificial intelligence, specifically Large Language Models (LLMs), to dramatically improve our understanding of why stock prices fluctuate. Their study demonstrates that advanced AI models can explain approximately 17 percent of same-day stock price movements, which is more than three times the explanatory power of the traditional earnings surprise metric. This significant boost in performance is attributed to the AI models' sophisticated ability to 'learn' and predict market reactions. Crucially, the AI's insights are recorded in detailed, human-readable digital notebooks, offering a transparent window into their decision-making process. The researchers conducted a rigorous live test involving nearly 2,000 earnings announcements from diverse industries in late 2025, rather than relying on historical data simulations. This methodology was specifically chosen to eliminate the risk of 'lookahead bias,' ensuring the validity and real-world applicability of their findings.
A key advantage of the AI models developed by Koijen and Levy is their capacity to produce comprehensive, written explanations for their predictions, which can be easily analyzed and verified by humans. An in-depth review of these AI-generated explanations revealed that the models are adept at identifying and interpreting subtle yet impactful details from earnings calls that go far beyond mere numerical data. These crucial details often encompass qualitative factors such as the forward guidance provided by company executives, the nuances within management commentary, and even the specific language and tone used by speakers when discussing financial results. Intriguingly, many of the insights derived from the AI models resonated strongly with the intuitive understanding of seasoned financial professionals. For instance, the models frequently recognized that positive news was often already factored into a stock's price, and consequently, they did not overreact to strong results, instead treating them as an anticipated baseline. This implies that extraordinary achievements like record profits or substantial 20 percent growth were often not sufficient on their own to trigger further significant stock movement. Koijen and Levy anticipate that these transparent, human-readable explanations will serve as invaluable resources for researchers, enabling them to formulate more precise hypotheses to rigorously test what truly drives price movements in the market. Furthermore, as these AI models continue to evolve and become more powerful, there is a promising potential for their generated explanations and underlying methodologies to be adapted and applied to smaller, less resource-intensive models, paving the way for the widespread and cost-effective scaling of these advanced analytical systems across the financial industry.
Despite the significant advancements presented by their research, Koijen and Levy prudently caution that these are still early results, acknowledging that a substantial portion of the variation in stock movements following earnings releases remains largely unexplained. To address this enduring challenge and foster further innovation, the researchers have taken a proactive step by launching an open competition. This initiative, generously sponsored by the prominent trading firm Optiver, invites individuals and teams from around the world to submit their own sophisticated models. The objective of this competition is to accurately predict how stock prices will respond to the rich array of information conveyed during earnings calls, mirroring the research conducted by Koijen and Levy. Full details and rules for model submission are available on the dedicated platform, explainingmarkets.ai. Koijen underscores the inherent difficulty of accurately forecasting these intricate market movements, labeling it as a 'surprisingly challenging task,' even for experts equipped with cutting-edge AI tools. He highlights that decades of cumulative research in this field have historically only been able to account for about 8 percent of the observed variation in market behavior. While their current research has successfully elevated this explanatory capability to an impressive 20 percent, Koijen stresses that there is still considerable room for improvement and further exploration, indicating an ongoing commitment to unraveling the complexities of financial markets with the aid of advanced artificial intelligence.