Earnings calls are the most important recurring event in investment research. Every quarter, thousands of public companies discuss their financial results, business strategy, and forward guidance. Investment banks and asset managers employ teams of analysts to listen to these calls, extract key information, and produce research notes. A single earnings call can generate 4-6 hours of analyst work — and with hundreds of calls per quarter, the research burden is enormous.
The Analyst Workflow
A typical earnings call analysis workflow involves several steps: listening to or reading the earnings call transcript, extracting key financial metrics (revenue, EPS, margins, guidance), identifying changes in management's tone and language, comparing results against consensus estimates, identifying key themes and strategic shifts, and writing a research note for internal or client use. Each step requires domain expertise and attention to detail.
Fine-tuned LLMs can automate most of this workflow. The model is trained on historical earnings call transcripts paired with analyst research notes, learning to identify the most important information in each call and produce structured summaries, sentiment assessments, and key metric extracts.
Fine-Tuning Approach
The training dataset consists of thousands of earnings call transcripts paired with corresponding analyst research notes. Each training example includes the full transcript as input and the analyst's research note as the expected output. The model learns to extract forward guidance statements, identify sentiment shifts in management language, calculate key financial metrics, and compare results against consensus expectations.
Using a smaller model like Claude Haiku or GPT-4o mini with full fine-tuning provides a cost-effective solution. These models are fast and inexpensive to run at scale, making them suitable for the high-volume demands of investment research. The fine-tuned model can process a single earnings call transcript in 30-60 seconds, compared to 4-6 hours for a human analyst.
Key Capabilities
The fine-tuned model extracts forward guidance statements — management's expectations for future revenue, earnings, and business conditions — and classifies them as positive, neutral, or negative relative to consensus. It identifies sentiment signals by analyzing changes in management's language, tone, and word choice compared to previous quarters. It extracts key financial metrics from the call transcript, including revenue, earnings per share, segment performance, and margin data, and compares them against analyst consensus estimates.
The model also identifies key themes and strategic shifts discussed during the call, such as new product launches, market expansion plans, cost-cutting initiatives, or M&A commentary. These themes are summarized in a structured format that analysts can use directly in their research notes.
Validation and Accuracy
Accuracy is validated against three benchmarks: extraction accuracy for financial metrics measured against actual reported figures, sentiment classification accuracy measured against analyst consensus, and research note quality measured through blind A/B testing with human analysts. The fine-tuned model typically achieves 90%+ accuracy on metric extraction and 85%+ agreement with analyst sentiment assessments.