Dataset

earnings_call_transcript_autograder

Earnings Call LLM Insights 📚 Read the Full Story: For a deep dive into the methodology, the wildest moments we found, and key takeaways, check out the blog post:KnowTrend.ai: Auto-Grading Ten Years of Earnings Calls for Prescience and Delusion This dataset contains LLM-generated analysis of ~70,000+ earnings call transcripts. The analysis was performed using Kimi k2-0905-preview, focusing on extra

Open on Hugging Face

Data preview

Subset

Loading preview...

Dataset card

The earnings_call_transcript_autograder dataset contains LLM-generated analyses of roughly 70,000+ earnings call transcripts, scored along financial dimensions and tagged with memorable analyst and management quotes. It is distributed in Parquet format on Hugging Face under an MIT license tag, with a flattened columnar schema suited to pandas or polars workflows.

What this dataset contains

Each row corresponds to a single earnings call, identified by a composite id built from ticker and date (e.g., WPC_2021-02-12). Core columns include ticker, date, company_name, market_cap, model_used, and processing_time, alongside composite scores such as score_company, score_macro, and score_time with free-text reasoning fields. Forward-looking returns are attached as return_1mo, return_1yr, and return_2025, allowing retrospective evaluation of the LLM’s foresight.

Structure and fields

The schema is a single train split of 34,282 rows in a flattened tabular format. Four award-style groups each repeat a uniform set of fields: a _speaker string, a _quote string, an _analysis string, and a numeric _rating. For example, the nostradamus group exposes nostradamus_speaker, nostradamus_quote, nostradamus_analysis, and nostradamus_rating, while the joker group mirrors this with joker_speaker, joker_quote, joker_analysis, and joker_rating on a 1–10 scale.

How analysts use it

Quantitative investors can join the per-call scores and award ratings to subsequent return_1mo and return_1yr outcomes to test whether prescient commentary or management missteps have predictive value for stock performance. Event-driven or sentiment-oriented funds may filter on fields such as joker_rating or nostradamus_rating to surface calls with unusually high-confidence calls or embarrassing quotes for further manual review. The dataset’s public-records provenance and ticker-level coverage also make it usable for systematic long/short screens built around transcript-derived signals.

Documentation & references 1