Grounded Event Extraction from SEC 8-K Filings with a Fine-Grained Taxonomy

Battle Count: 75

Highly relevant for quantitative trading. The paper demonstrates that fine-grained event tags carry reaction-relevant information that SEC item codes lack (tags explain 1.3pp more variance than item codes add on top of tags, a 3:1 ratio). The event study shows economically distinct reactions within single item codes (e.g., CEO departures vs. routine appointments under Item 5.02). The quality-score filtering dial enables practitioners to trade precision for coverage. The taxonomy identifies that the most market-moving events (clinical trials, merger completions, regulatory decisions) are filed under the uninformative Item 8.01 catch-all. The released dataset of 601,088 grounded tags provides a ready-made event calendar for signal generation. However, the paper explicitly makes no return-prediction claims and measures only reaction magnitude, not direction.

Moderate complexity. The pipeline requires: (1) a compact instruction-tuned LLM with structured output support, (2) schema validation with retry logic, (3) fuzzy n-gram quote matching (4-word sequences, 40% threshold), (4) a second LLM call for quality scoring, (5) a three-tier taxonomy with 119 definitions in the prompt. The core logic is straightforward but requires careful prompt engineering, robust error handling for retries, and access to commercial LLM APIs. The event study component uses standard financial econometrics (market-adjusted returns, variance standardization). No model training or fine-tuning is required.

The taxonomy and processed dataset (601,088 tags) are publicly released at massive.com/docs/rest/stocks/filings/8-k-disclosures. The methodology is described in detail including schema validation, fuzzy n-gram matching parameters (4-word sequences, 40% threshold), and scoring rubric. However, the specific extraction model is described only as 'a compact, commercially available instruction-tuned model' without naming it, and the exact prompt templates are not provided. The LLM judge (Opus 4.8) is named. Reproduction would require access to the same commercial model and knowledge of the exact prompts.