The viral "GPT-4 beats stock analysts" paper was withdrawn by its own authors

If you have seen a post claiming that "AI predicts earnings better than human analysts," there is a good chance it traces back to one paper: "Financial Statement Analysis with Large Language Models" by Alex Kim, Maximilian Muhn and Valeri Nikolaev of the University of Chicago Booth School of Business. It was among the most widely shared AI finance papers of 2024. It is also no longer a valid source.

What the paper claimed

First posted on SSRN in May 2024 and on arXiv in July 2024, the paper gave GPT-4 standardized, anonymized balance sheets and income statements, with no company names and no narrative, and asked it, using chain-of-thought prompts written to mimic an analyst's process, whether next year's earnings would rise or fall.

The version posted in November 2024 described a full sample of 150,678 firm-year observations from 15,401 companies over 1968 to 2021, and an analyst comparison sample of 39,533 observations from 3,152 companies over 1983 to 2021. Its headline results:

• GPT-4 with chain-of-thought reached an accuracy of 60.35% in predicting the direction of earnings, about 7 percentage points above analysts' consensus forecasts issued within a month of the prior earnings release, which the paper put at 53%. nContext: n=39,533 firm-years in the analyst sample · 1983 to 2021 · Kim, Muhn and Nikolaev, arXiv:2407.17866 v2 (withdrawn) · margin: not disclosed
• GPT-4's accuracy on the full sample, 60.31%, was described as on par with a specialized neural network trained for the task. nContext: n=150,678 firm-years · 1968 to 2021 · arXiv:2407.17866 v2 (withdrawn) · margin: not disclosed
• An equal-weighted trading strategy based on GPT-4's predictions reached a Sharpe ratio of 3.36, against 2.54 for the neural network and 2.05 for logistic regression. nContext: n: portfolio months not stated · 1968 to 2021 sample · arXiv:2407.17866 v2, Table 8 (withdrawn) · margin: not disclosed
• The abstract stated that the model's predictions did not stem from its training memory.

All four bullets are claims from a withdrawn paper, reported here only so readers can recognize them. nContext: n=150,678 firm-years (39,533 in the analyst sample) · 1968 to 2021 · Kim, Muhn and Nikolaev, arXiv:2407.17866 v2, now withdrawn · margin: not disclosed

What happened

On February 20, 2025, the arXiv record was replaced by a withdrawal notice. The authors' comment reads: "A co-author identified inconsistencies in the data and analyses while attempting to replicate past analyses from the working paper. Accordingly, we have temporarily withdrawn the working paper from circulation while we review the research findings."

Valeri Nikolaev's Chicago Booth faculty page gives more detail. He writes that he has independently confirmed the inconsistencies in the underlying data and analyses, and that the team decided to review certain other LLM-based working papers for similar issues. An update on the same page says that a second paper, "Bloated Disclosures: Can ChatGPT Help Investors Process Information?", has also been withdrawn, because a co-author's attempt to replicate key results using the approach described did not yield results supporting the reported findings.

The authors themselves raised the problem and withdrew the papers. That is how research is supposed to correct itself. It does not tell us what a corrected analysis would show; as of our check on October 14, 2026, no revised version had been posted on arXiv.

Why the claim spread so far

The result fit a story people wanted to hear: a general chatbot, given only numbers, outperforming professionals. It came from a top business school, had a large sample and a clean headline number. And it was easy to compress into a single sentence.

Even before the withdrawal, the design carried a risk that later research made sharper. The sample ran from 1968 to 2021, years inside GPT-4's training data. The 2025 "Memorization Problem" paper from the University of Florida argues that when a model has seen the realized values, its forecasting skill cannot be identified from its outputs, and that anonymizing companies does not fully prevent models from reconstructing who and when. We cover that research in our lookahead bias explainer. The withdrawn paper's statement that predictions did not stem from memory was exactly the kind of claim those tests later showed to be hard to establish.

What to do with the headline number

• Do not cite it as evidence. The authors have withdrawn the paper and confirmed inconsistencies in the data and analyses. nContext: n=39,533 firm-years in the analyst sample · 1983 to 2021 · arXiv:2407.17866 v2 (withdrawn); withdrawal notice Feb 20, 2025 · margin: not disclosed
• Check the arXiv page, not the screenshot. A withdrawn paper keeps its old abstract on many mirror sites and citation databases.
• Look for post-cutoff tests. Studies that use data after a model's training cutoff, such as the ChatGPT headline study, are built to avoid the memory question.
• Expect replication to matter. In this case, the problem was found by one of the authors trying to reproduce the results.