Luka Vulicevic

This paper evaluates whether large language models (LLMs) improve or impair market efficiency. Exploiting plausibly exogenous outages at major LLM providers, we find that LLM availability eliminates at least 46% to 61% of post-news drift into the next trading day. The absence of reversal alongside the magnitude of this reduction suggests that LLMs play a first-order role in short-horizon price discovery. Because LLM supply is highly concentrated, outages expose a single point of failure in market-wide information processing. Consistent with this interpretation, investors insulated from major LLM outages can earn more than twice the day t + 1 returns of the long-short news-sentiment strategy following outage events.

Investors' computing power ("compute") governs how much signal they can extract from high-dimensional data. We show that forecast performance follows scaling laws—stable power-law relationships between accuracy and training compute—that pin down both an irreducible bound on return predictability and the rate at which additional compute improves accuracy. In firm-level cross-sectional return prediction, scaling laws explain over 80% of performance variation across models. Treating compute as a primitive yields sharp economic implications: scaling laws quantify predictability limits and the value of new data, distinguish problems with strong versus weak returns to scale, and deliver a market-efficiency metric by mapping computational advantage into a certainty equivalent. Measured this way, computational superiority earns sizable rents: a 25% increase in compute would have raised an investor's Sharpe ratio by about 10% over the last 30 years, indicating substantial returns to computational sophistication.