Situation

The client aimed to identify predictive signals for stock performance by analyzing large-scale earnings call transcripts. Traditional financial analysis often overlooked qualitative aspects such as communication style, sentiment, and information disclosure, leading to incomplete insights.

Task

Develop a scalable NLP and ML pipeline capable of processing tens of thousands of earnings call transcripts, extracting structured features from qualitative communication patterns, and linking them with financial fundamentals to predict stock returns.

Action

• Implemented semantic similarity analysis (cosine similarity on vector embeddings) to evaluate precision of answers and proactive disclosure in presentations.

• Applied sentiment analysis on speaker-level and section-level granularity to capture nuanced tones in communication.

• Leveraged topic modeling to classify content, enabling segment-specific scoring.

• Combined extracted signals with company fundamentals using ML algorithms for predictive modeling of stock price performance.

• Built scalable data pipelines with Spark and deployed modular services via FastAPI for efficient experimentation and integration.

Result

Processed and analyzed over 90,000 transcripts, generating structured communication-based features. The resulting equity trading strategy achieved an annualized return of 20%, demonstrating the predictive power of communication-driven signals in financial markets.