Supervised Fine-tuning of Large Language Models on Financial Statements to Predict Future Earnings Changes
Professor Yiwei Dou
Professor of Accounting
Leonard N. Stern School of Business
New York University
We fine-tune an open-source LLM to predict the direction of one-year-ahead earnings changes based on standardized, anonymized financial statements paired with next year’s realized earnings change directions. Our model shows significant out-of-sample predictive power in the period after the LLM’s release date (i.e., no leakage of target earnings numbers), with an area under the Receiver Operating Characteristics curve (AUC) of 67.88 percent. It outperforms a random guess, gradient boosted decision trees trained on the same financial statement items, and machine learning models in Chen et al. (2022). Extensive analyses suggest that the predictive power comes primarily from the LLM’s prior knowledge, particularly its understanding of the temporal structure and accounting terms in financial statements, which informs the specification search in ways tabular machine learning models often miss. The findings suggest that properly fine-tuned LLMs are powerful tools for earnings forecasting.
