𝐀𝐈 𝐣𝐮𝐬𝐭 𝐫𝐞𝐭𝐞𝐬𝐭𝐞𝐝 𝟐𝟎 𝐲𝐞𝐚𝐫𝐬 𝐨𝐟 𝐭𝐨𝐩 𝐟𝐢𝐧𝐚𝐧𝐜𝐞 𝐫𝐞𝐬𝐞𝐚𝐫𝐜𝐡. 𝐌𝐨𝐬𝐭 𝐨𝐟 𝐢𝐭 𝐰𝐨𝐫𝐤𝐞𝐝. 𝐓𝐡𝐞𝐧 𝐭𝐡𝐞 𝐝𝐚𝐭𝐚 𝐫𝐚𝐧 𝐨𝐮𝐭. Dmitry Muravyev (University of Illinois Urbana-Champaign) had AI agents rebuild the main finding of 𝟏,𝟑𝟐𝟖 papers from the three top finance journals, 2000 to 2020. Doi

𝐀𝐈 𝐣𝐮𝐬𝐭 𝐫𝐞𝐭𝐞𝐬𝐭𝐞𝐝 𝟐𝟎 𝐲𝐞𝐚𝐫𝐬 𝐨𝐟 𝐭𝐨𝐩 𝐟𝐢𝐧𝐚𝐧𝐜𝐞 𝐫𝐞𝐬𝐞𝐚𝐫𝐜𝐡. 𝐌𝐨𝐬𝐭 𝐨𝐟 𝐢𝐭 𝐰𝐨𝐫𝐤𝐞𝐝. 𝐓𝐡𝐞𝐧 𝐭𝐡𝐞 𝐝𝐚𝐭𝐚 𝐫𝐚𝐧 𝐨𝐮𝐭. Dmitry Muravyev (University of Illinois Urbana-Champaign) had AI agents rebuild the main finding of 𝟏,𝟑𝟐𝟖 papers from the three top finance journals, 2000 to 2020. Doing that by hand takes about a week per paper, he says. Roughly 𝟐𝟓 𝐲𝐞𝐚𝐫𝐬 of work. For the 𝟏,𝟎𝟎𝟓 with enough later data, the agents then reran the exact same test on the years after each paper's data ended. ✅ Original data: 𝟕𝟓% of findings held up and the typical effect kept 𝟗𝟑% of its published size. Only 1.3% came out pointing the opposite way. Far better than psychology's famous 36%. 📉 New data: only 𝟒𝟐% held up and the typical effect shrank to 𝟒𝟓% of its published size. 🎯 The twist: the shrinkage happens all at once, 𝑒𝑥𝑎𝑐𝑡𝑙𝑦 𝑤ℎ𝑒𝑟𝑒 𝑡ℎ𝑒 𝑜𝑟𝑖𝑔𝑖𝑛𝑎𝑙 𝑑𝑎𝑡𝑎 𝑒𝑛𝑑𝑠. Not gradually as markets change. Not when the paper is published. That is the pattern you would expect from 𝑑𝑎𝑡𝑎 𝑠𝑛𝑜𝑜𝑝𝑖𝑛𝑔, choices that fit the original sample a little too well. It is much larger in asset pricing, the field that studies returns. 💪 What predicts survival? Mostly one thing: how strong the original evidence was. Findings with very strong statistical evidence survived 𝟕𝟐% of the time in new data and kept about 𝐭𝐰𝐨 𝐭𝐡𝐢𝐫𝐝𝐬 of their effect. Borderline ones survived 𝟑𝟏% of the time and kept about 𝐚 𝐭𝐡𝐢𝐫𝐝. How often a paper is cited, or how famous its authors are, matters surprisingly little. Why it matters: these estimates feed trading models, risk models and policy