AI/Fin Research
Analysis

A finance pipeline turned 31,460 Compustat signals into 288 complete papers in about a day

Novy-Marx and Velikov published every stage, every prompt and every paper in January 2025. One of them now helps run the AFA session for AI-written work.

AI Fin ResearchCovers US
Finance has had a published pipeline from idea to manuscript since January 2025. It wrote 288 papers.

Robert Novy-Marx of the University of Rochester and Mihail Z. Velikov of Penn State describe “a process for automatically generating academic finance papers” and run it on stock return predictability. The code, the prompts and every generated paper are in a public repository under an MIT license.

31,460candidate signals built from Compustat accounting items
96signals left after every test
288complete papers, three per signal

Seven steps run from raw accounting data to a compiled PDF

Step The script does this Result
1. Build candidates Combines Compustat items and their changes over time into ratios and differences 31,460 signals
2. Clean Drops duplicates, cross-sections under 30 stocks and histories under 360 months 17,074 signals
3. Test Sorts stocks into portfolios and checks the return spread five ways 183 signals
4. Benchmark Runs the authors’ “Assaying Anomalies” protocol against more than 200 published anomalies and prints a PDF report for each signal 96 signals
5. Name GPT-3.5-turbo turns variable codes into an academic name “Operating Liquidity Margin” for current assets over EBITDA
6. Write Claude 3.5 Sonnet writes the introduction, data section and conclusion from the report Three versions per signal
7. Assemble Scripts fill a LaTeX template, merge the references and compile 288 PDFs

The authors describe the 288 papers as “fully programmatically-generated”. The appendix prints the prompts.

Six tests cut 17,074 candidates to 96

Test Signals left
Candidates after cleaning 17,074
Significant return spread, equal-weighted deciles 7,102
Equal-weighted quintiles 1,249
Quintiles on NYSE breakpoints 808
Value-weighted portfolios 640
Alpha against the Fama-French six-factor model 183
“Assaying Anomalies” protocol 96

Each signal got three theories, written after the result was known

The introduction of each paper runs to roughly 1,100 words in four fixed parts: 200 on motivation, 300 on hypothesis development, 300 on results and 300 on contribution. The model wrote three versions for every signal, each with a different economic mechanism for the same return spread. The paper’s reference list credits the generated papers to an author named I. M. Harking.

The repository has moved since the paper. Its README now lists 95 signals with four versions each: slow diffusion of information, production-based asset pricing, consumption-based asset pricing and an unrestricted one.

The authors report that the model “occasionally generates fictitious references” when it cites specific or recent work, while most citations to foundational papers in top finance journals are accurate. They had not yet measured the rate. Their abstract calls the exercise “a cautionary tale, illustrating how it can be abused to industrialize HARKing”, hypothesizing after results are known.

The profession waited sixteen months to ask for more

NBER posts the paper.The repository with the code and the generated papers had been created on 16 December 2024.
The AFA asks for papers "predominantly generated by AI".Novy-Marx is one of the four organizers of the session at the 2027 annual meeting.
Submissions close.Each one carries every AI conversation and a time log of the human work. The rules.

A reader of one of these papers cannot tell from the text that the mechanism was chosen after the sort. Open the repository, pick one signal and read its versions side by side. Then date your own hypothesis before you run the test: a research plan committed to git ahead of the data pull is the evidence a referee can check.

Sources

Related