Quantifying Earnings Momentum with AI
This guide explains how to fine-tune an LLM to synthesize unstructured earnings calls and structured financial data into a single "Earning Call Momentum Score." This score provides an actionable signal for predicting post-earning stock performance.
The Momentum Score Lifecycle
From raw documents and data to a predictive score, the process requires careful data integration and specialized training. Hover over each stage to understand its role.
Combining Unstructured & Structured Data
A robust model needs to understand both the narrative (what management says) and the numbers (what the company reports). We combine qualitative text with quantitative historical data.
Unstructured Data
Earnings Call Transcripts & Analyst Reports: The primary source for management's tone, sentiment, guidance, and responses to analyst questions. This provides the qualitative context.
Structured Data
Last 3 Quarters of Financials: Historical performance data provides the quantitative baseline for the current quarter's results.
| Quarter | Actual EPS | Est. EPS | Actual Rev | Est. Rev |
|---|---|---|---|---|
| Q3 '25 | $1.55 | $1.50 | $25.2B | $25.0B |
| Q2 '25 | $1.48 | $1.49 | $24.5B | $24.3B |
| Q1 '25 | $1.42 | $1.40 | $23.8B | $23.9B |
Teaching the Model to Synthesize
Data curation is where we create training examples that teach the LLM to connect the words to the numbers. The model learns to perform calculations and classifications based on the combined data.
Instruction Formatting Example
The model is instructed to analyze the current quarter's numbers in the context of the earnings call text and produce a structured analysis.
Raw Data (Inputs)
Q4 EPS: { Actual: $1.65, Est: $1.60 }
Q4 Rev: { Actual: $26.1B, Est: $26.0B }
Call Snippet: "Despite macro headwinds, we delivered another record quarter... Our AI cloud services are seeing exponential demand, and we are confidently raising next year's guidance."Formatted for LLM
{
"instruction": "Analyze the structured data and earnings call text. Determine the status (Beat, Miss, Meet) and percentage surprise for EPS and Revenue. Classify the call sentiment.",
"input": "Current EPS: {Act: 1.65, Est: 1.60}, Current Rev: {Act: 26.1B, Est: 26.0B}, Transcript: '...confidently raising next year's guidance.'",
"output": "{ 'eps_status': 'Beat', 'eps_surprise_pct': 3.13, 'rev_status': 'Beat', 'rev_surprise_pct': 0.38, 'sentiment': 'Very Positive' }"
}The Earning Call Momentum Score
The Momentum Score is a composite metric calculated from the LLM's output. It combines the magnitude of financial beats with the qualitative sentiment from the call to produce a single number from 0 (very bearish) to 100 (very bullish). Adjust the sliders below to see how each component affects the score.
Score = (EPS_Component + Rev_Component + Sentiment_Component)
Defining Success: From Score to Profit
The ultimate test is whether the Momentum Score has predictive power. Evaluation focuses on the correlation between high scores and positive future stock returns.
Key Metric Types
-
✓
Score-to-Return Correlation
The most critical metric. Does a portfolio of high-scoring stocks outperform a portfolio of low-scoring stocks? (e.g., Information Coefficient).
-
✓
Financial Backtesting
Simulates a trading strategy based on the score (e.g., "buy above 75, sell below 25") to calculate Sharpe Ratio and Alpha.
-
✓
LLM Output Accuracy
Assesses if the LLM correctly calculates surprise percentages and classifies sentiment against a human-labeled test set (F1-score, RMSE).
Navigating the Pitfalls
Creating a reliable momentum score from complex data introduces new challenges beyond standard LLM fine-tuning.
Score Calibration & Weighting
Determining the optimal weights for EPS surprise, revenue surprise, and sentiment is a major challenge, often requiring extensive historical testing and optimization.
Handling "Whisper Numbers"
The market often reacts to unofficial "whisper" estimates, not just the official consensus. A stock can drop despite a beat if it doesn't beat by enough. The model must learn this context.
Look-Ahead Bias
Ensuring that the model's training data for a specific earnings report ONLY contains information that was publicly available *before* the report was released.