kevinkakkad168/earnings-alpha-llm

Earnings Call Alpha Extractor

Fine-tuned Llama 3.1 8B for extracting quantitative signals from earnings call Q&A transcripts.

What it does

Scores earnings call Q&A sections on two dimensions:

• Confidence score (1-10): How confident and direct is management?
• Evasiveness score (1-10): How evasive, hedging, or deflecting is management?

The mismatch between these scores and reported financial results (EPS surprise) forms an alpha signal with statistically significant IC.

Training

• Base model: meta-llama/Meta-Llama-3.1-8B-Instruct
• Fine-tuning: LoRA on Together AI
• Training data: ~2500 earnings calls labeled by Llama 3.3 70B (Groq)
• Dataset: Rogersurf/earnings-call-transcripts (9k+ calls, 2023-2026)

Backtest results

Metric	Value
IC (5-day)	+0.076
IC t-stat	+2.00 (significant)
Sharpe	0.74
N events	700
Period	2023–2026

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
import torch

model = AutoModelForCausalLM.from_pretrained(
    "kevinkakkad168/earnings-alpha-llm",
    torch_dtype=torch.float16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("kevinkakkad168/earnings-alpha-llm")

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)

See the project GitHub for the full pipeline.

References

• Mayew & Venkatachalam (2012): The Power of Voice
• Li (2008): Annual report readability and earnings persistence