Linking Earnings Call Sentiment to Market Performance
Earnings calls carry information that doesn't show up neatly in the numbers — tone, hedging, confidence, and how executives handle tough questions all shape how the market reacts afterward. This project builds an NLP pipeline that converts unstructured earnings call transcripts into a structured sentiment signal, using GPT-4 alongside custom ML processing, and studies how that signal relates to subsequent stock performance. A conversational query interface sits on top, letting a user ask questions and pull insights across many filings at once instead of reading each transcript individually.
Combined a large language model with a custom scoring pipeline to turn unstructured transcript text into a consistent, comparable sentiment signal — then paired it with a retrieval-based conversational interface so a user can query insights across many filings at once, instead of reading each transcript individually.
Sentiment & tone extraction
Transcript-level analysis
Query insights across filings
The pipeline processes raw earnings call transcripts and extracts a sentiment/tone score for each one, capturing language cues that go beyond the reported financial figures. Those scores are then structured so they can be compared across companies, sectors, and time — and lined up against subsequent stock price movement to study the relationship between what executives say and how the market responds.
Rather than searching transcript by transcript, the conversational interface lets a user ask a question in plain language — about a company, a sector, or a theme — and get an answer synthesized from across the relevant filings, cutting out the manual read-through that investment research usually requires.
The platform was built as an end-to-end pipeline, from raw transcript to queryable insight:
The combination of a sentiment signal and a cross-filing query interface has a few natural uses in an investment research workflow:
Ask questions across many filings at once instead of reading each transcript in full.
Surface companies or sectors with notably positive or negative call tone for further review.
Add a language-based signal alongside traditional financial metrics for a fuller picture.
How something is said in an earnings call carries information the reported numbers alone don't — a systematic way to capture that tone at scale is a useful complement to traditional financial analysis.
This project shows how a large language model can turn unstructured, high-volume text like earnings calls into a structured research asset, paired with a natural-language interface for exploring it.