> ## Content Index
> Fetch the complete content index at: https://fsc.stevens.edu/llms.txt
> Use this file to discover other available public pages before exploring further.

# Large Language Models and Financial Narrative Signals: Behavioral Feature Extraction from Hedge Fund Investor Letters for Return Prediction
- URL: https://fsc.stevens.edu/large-language-models-and-financial-narrative-signals-behavioral-feature-extraction-from-hedge-fund-investor-letters-for-return-prediction/
- Published: 2026-11-20T15:06:29.000Z
- Updated: 2026-11-20T15:06:29.000Z
- Author: Malavika Loka
- Tags: FE 800 Projects, 2026 Spring

**Researcher:** Arosh De Silva  
**Course Instructor:** Professor Zhenyu Cui 

## Introduction

Hedge fund investor letters describe managers’ views and how confidently they communicate them. This project explores whether a large language model can turn those cues into measurable features and whether the resulting signals relate to subsequent stock returns. 

### Methodology

The researcher extracted and cleaned text from investor letters issued by Greenlight Capital, Praetorian Capital, and Third Point. A large language model scored features including sentiment, conviction, uncertainty, complexity, unusual communication, and forward-looking tone. The project combined selected features into a behavioral signal and compared it with subsequent returns over multiple horizons. 

[ ![Worked example showing how uncertainty, complexity, and unusual communication reduce a positive sentiment and conviction contribution to a final signal of negative 1.25](https://fsc.stevens.edu/content/images/2026/11/DeSilva_ExampleCalcTab.jpg) ](https://fsc.stevens.edu/content/images/2026/11/DeSilva%5FExampleCalcTab.jpg) 

Example signal calculation. A sentiment and conviction contribution of +5 becomes −1.25 after the three adjustments. 

### Results & Analysis

**In several cases, bullish letters with elevated uncertainty, complexity, or unusual communication were followed by weaker short-term returns.** The worked example illustrates how those features can change the interpretation of a letter with positive sentiment. 

### Conclusion

The project demonstrates how investor letters can be converted into structured features for return research. The small sample makes the findings exploratory, and a larger set of letters would help test whether the patterns hold across funds and market periods.