Strux Transcripts Dataset

Overview

Our dataset consists of 11,950 quarterly earnings call transcripts from major companies listed on the NASDAQ 500 and S&P 500, sourced from the Motley Fool website between 2017 and 2024. Each transcript contains an average of 10,187 tokens, making it a valuable testbed for investment decision prediction and market sentiment analysis. The dataset includes transcripts from 869 companies across 11 financial sectors. For training, we constructed a balanced subset of 1,100 transcripts (100 from each sector), while our test set comprises 587 transcripts from 2024, specifically selected to avoid overlap with LLM pretraining data (cutoff: December 2023). The dataset is annotated with ground-truth investment labels (Strongly Buy, Buy, Hold, Sell, or Strongly Sell) based on 30-day post-earnings stock performance, focusing solely on textual content without acoustic features.

Acknowledgement

If you use this dataset, please cite the following paper:

STRUX: An LLM for Decision-Making with Structured Explanations
Yiming Lu, Yebowen Hu, Hassan Foroosh, Wei Jin, Fei Liu
In main conference of the Nations of the Americas Chapter of the Association for Computational Linguistics (NAACL’25), Albuquerque, US.

Bibtex

@article{lu2024strux,
  title={STRUX: An LLM for Decision-Making with Structured Explanations},
  author={Lu, Yiming and Hu, Yebowen and Foroosh, Hassan and Jin, Wei and Liu, Fei},
  journal={arXiv preprint arXiv:2410.12583},
  year={2024}
}

Usage

from datasets import load_dataset
transcripts = load_dataset("BUILDERlym/STRUX-Transcripts")

train_data = transcripts['train']
test_data = transcripts['test']

def generator(data_split):
  for instance in data_split:
    yield instance['ticker'], instance['date'], instance['participants'], instance['prepared_remarks'], instance['questions_and_answers']

Python Scripts Useful scripts and guidance can be found in github repo STRUX