Forecasting Stock Returns from Earnings Sentiment

Overview

This project explores the relationship between the language used in earnings call transcripts and subsequent stock price movement. Using NLP-driven sentiment extraction and market data, it attempts to forecast 5-day forward returns after a company’s quarterly earnings release.

Goals

• Extract meaningful sentiment signals from large-scale earnings transcripts
• Correlate sentiment and linguistic cues with post-call stock movement
• Build a regression model to predict short-term return magnitude and direction

Data & Processing

• Textual Data: Full transcripts of earnings calls from public companies
• Market Data: 5-day forward and historical stock returns retrieved via financial APIs
• Preprocessing: Text tokenization, sentiment aggregation across call sections, length normalization
• Additional Features: Call verbosity, polarity intensity, past price performance

Modeling & Experimentation

The modeling pipeline uses multiple engineered features to forecast the direction and magnitude of short-term returns. Historical returns and textual features are used to reduce variance and account for prior momentum. Model training included cross-validation and evaluation on a held-out test set.

Evaluation Metrics

• Mean Absolute Error (MAE): 0.0695 – suggesting average prediction errors of around 7%, which is considered reasonable for short-horizon return modeling
• R² Score: -0.1581 – a negative R² is common in short-term market prediction due to high noise levels and limited signal

Key Challenges

• Market noise and the inherently weak signal of short-term returns
• Ensuring that sentiment is extracted consistently across diverse transcript formats
• Dealing with illiquid tickers or inconsistent earnings dates in market data

Takeaways

Even with modest R², the model's structured approach to combining language sentiment and market behavior demonstrates practical potential. The system quantifies unstructured financial speech into measurable signals and shows promise for downstream applications like risk monitoring, volatility forecasting, or thematic investing. Further improvements may include fine-tuned large language models, event-time alignment, and integration of option flow or macroeconomic context.

Code

If you want access to the codebase, reach out to schilamkur@gmail.com.