Research

A new NBER working paper by Ralph Koijen and Bradford Levy finds that optimised agentic AI systems extracting signals from earnings call transcripts more than double the explained variation in stock returns around earnings announcements, pushing R-squared from roughly 8% to approximately 20%

The paper builds a real-time, out-of-sample benchmark that evaluates AI systems using only information available at the time, avoiding look-ahead bias, and the authors argue the approach's real value is producing "human-readable economic mechanisms" that can be compared against existing asset pricing models to identify previously unexplained sources of return variation.

· Source: National Bureau of Economic Research


The jump from roughly 8% to approximately 20% explained variation sounds abstract until you translate what it means: for every dollar of return movement around an earnings announcement, standard models were previously accounting for less than a tenth of it, and the optimised agentic system accounts for a fifth. That is still a long way from fully explaining returns, most of the variation remains unexplained by either approach, but more than doubling the explained share in one methodological step is a meaningfully large gain in a field where incremental improvements are usually measured in fractions of a percentage point.

The real-time, out-of-sample design is the part that gives the result credibility rather than just novelty. A huge amount of quant research overstates predictive power by testing on historical data the model could, directly or indirectly, have learned patterns from in hindsight. Koijen and Levy's benchmark evaluates the system using only information contemporaneous with the announcement itself, which is a genuinely harder test and the more relevant one for anyone considering deploying something like this against live markets rather than a backtest.

The emphasis on "human-readable economic mechanisms" as a stated goal, not just a side benefit, is worth noting given how much of the agentic-AI trading conversation elsewhere this week has been about opacity and accountability. The House Democrats' letter to the SEC and the Bank of England's concentration-risk review, both covered separately on this site, are fundamentally worried about AI systems making trading decisions nobody can audit after the fact. A methodology that produces interpretable mechanisms comparable to existing asset pricing models is a direct, if narrow, answer to exactly that concern, at least for the earnings-announcement use case this paper studies.

Whether this style of interpretable, benchmarked agentic system becomes the industry standard, or gets outcompeted by less transparent but higher-performing black-box approaches, is the open question. Given regulators on both sides of the Atlantic are actively asking who is accountable when an AI trading system gets it wrong, a system that can show its work has an advantage that pure predictive accuracy alone might not provide once oversight requirements catch up with deployment.


Read the original: National Bureau of Economic Research - A new NBER working paper by Ralph Koijen and Bradford Levy finds that optimised agentic AI systems extracting signals from earnings call transcripts more than double the explained variation in stock returns around earnings announcements, pushing R-squared from roughly 8% to approximately 20%. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.