kevinkakkad168/earnings-alpha-llm
kevinkakkad168/earnings-alpha-llm is an 8 billion parameter language model, fine-tuned from Meta-Llama-3.1-8B-Instruct, specifically designed to extract quantitative signals from earnings call Q&A transcripts. This model scores management's confidence and evasiveness, providing an alpha signal for financial analysis. It is optimized for identifying discrepancies between management's communication style and reported financial results, offering a unique tool for quantitative finance strategies.
Loading preview...
Overview
kevinkakkad168/earnings-alpha-llm is an 8 billion parameter model, fine-tuned from meta-llama/Meta-Llama-3.1-8B-Instruct, specialized in analyzing earnings call Q&A transcripts. Its primary function is to quantify management's communication style by assigning two scores:
- Confidence Score (1-10): Measures how direct and assured management's responses are.
- Evasiveness Score (1-10): Assesses how much management hedges, deflects, or avoids direct answers.
Key Capabilities
This model's unique value lies in its ability to generate an alpha signal for financial markets. By comparing these communication scores with actual financial outcomes (like EPS surprise), it identifies statistically significant predictive patterns. The model was fine-tuned using LoRA on approximately 2500 earnings calls, with data labeled by Llama 3.3 70B, leveraging the Rogersurf/earnings-call-transcripts dataset.
Performance & Use Cases
Backtest results over the 2023-2026 period show a significant Information Coefficient (IC) of +0.076 (t-stat +2.00) and a Sharpe ratio of 0.74, indicating its potential for generating profitable trading signals. This model is ideal for quantitative analysts, hedge funds, and financial researchers looking to incorporate linguistic cues from earnings calls into their investment strategies. It provides a novel approach to extracting actionable insights beyond traditional fundamental analysis.