Replicating Financial Market Research with an LLM

Can large language models (LLM) rapidly replicate findings of financial market research? In his August 2026 paper entitled "Does Empirical Finance Replicate?",  Dmitriy Muravyev tasks an autonomous LLM pipeline with reproducing the main result of 5,430 empirical papers published in the Journal of Finance, Journal of Financial Economics and Review of Financial Studies from 2000 to 2020, both for the original data (in-sample) and then for later years (our-of-sample). Limited by data availability, the LLM is able to generate 1,328 in-sample replications, for which 1,005 have enough post-study data for out-of-sample tests. Using models specified in published papers and replication data as available, he finds that: