Generating long-horizon stock "buy" signals with a neural language model

TL;DR

This study fine-tunes a small language model to predict long-term stock buy signals from 10-K reports, achieving notable out-of-sample accuracy at 6-9 months, with implications for sector-specific reporting styles.

Contribution

It introduces a method for forecasting long-horizon stock movements using narrative text from financial reports with improved accuracy over random models.

Findings

Buy signals are most precise at 6 and 9 months.

F1 score of 0.62 indicates good out-of-sample performance.

Buy signals outperform random selection by 4.8-9%."

Abstract

This paper describes experiments on fine-tuning a small language model to generate forecasts of long-horizon stock price movements. Inputs to the model are narrative text from 10-K reports of large market capitalization companies in the S&P 500 index; the output is a forward-looking buy or sell decision. Price direction is predicted at discrete horizons up to 12 months after the report filing date. The results reported here demonstrate good out-of-sample statistical performance (F1-macro= 0.62) at medium to long investment horizons. In particular, the buy signals generated from 10-K text are found most precise at 6 and 9 months in the future. As measured by the F1 score, the buy signal provides between 4.8 and 9 percent improvement against a random stock selection model. In contrast, sell signals generated by the models do not perform well. This may be attributed to the highly imbalanced out-of-sample data, or perhaps due to management drafting annual reports with a bias toward positive language. Cross-sectional analysis of performance by economic sector suggests that idiosyncratic reporting styles within industries are correlated with varying degrees and time scales of price movement predictability.