Streamlined, Focused AI and Stock Return Prediction

Can relatively modest large language models (LLM), pretrained with diverse financial information, effectively rank stocks? In their September 2024 paper entitled “Re(Visiting) Large Language Models in Finance”, Eghbal Rahimikia and Felix Drinkall introduce base and small versions of FinText, LLMs that are: (1) kept compact compared to state-of-the-art LLMs to allow practical use with personal computers; and, (2) pre-trained by calendar year with diverse financial information. They test FinText stock-ranking ability by each day:

• Asking it to review Dow Jones Newswires articles available by 9:00AM and tagged as having significant news about cited firms (available since February 2013).
• Measuring the accuracy of its resulting predictions for individual stock price directions.
• Tracking performance of an equal-weighted or value-weighted portfolio that is each day at the market open long (short) the fifth of stocks with the highest (lowest) probabilities of positive daily returns.

They use articles from 2013 through 2016 for FinText training and those from 2017 through 2023 for testing. They compare accuracy and performance of FinText to those of unspecialized LLMs much larger than base FinText. Using pretraining information, daily Dow Jones Newswires articles and daily returns for a broad sample of U.S. stocks during 2013 through 2023, they find that: (more…)