There is a quiet but significant problem at the heart of U.S. financial disclosure: the system designed to inform markets about major corporate events is, in many ways, too blunt to be useful. SEC 8-K event extraction has long been constrained by item codes that lump wildly different occurrences into the same bucket — a routine board committee update and a CEO resignation might share a single category. A new research system from Rian Dolphin aims to fix that, and the results suggest large language models can do far more than summarize text — they can build reliable, granular maps of corporate events at massive scale.
Key takeaways
- Form 8-K filings are the primary disclosure channel for U.S. public companies reporting material events, but existing SEC item codes are too coarse to distinguish economically distinct events.
- A new two-stage system applies a three-tier taxonomy of 119 event types to 8-K filings, anchoring every tag to a verbatim quote from the source text.
- Applied to 292,984 filings from 2022 to 2026, the system generated 601,088 grounded event tags, now released publicly.
- Tagging precision rises from 12% to 96% as quality scores increase, with unsupported tags falling to near zero at the highest thresholds.
- An event study confirms the taxonomy separates economically distinct events — without using any language model — validating the approach through abnormal return analysis.
Limitations of Existing SEC 8-K Disclosures
The Role of Form 8-K in U.S. Public Company Disclosures
Form 8-K is the mandatory reporting mechanism U.S. public companies use when something material happens — a merger, an executive departure, a bankruptcy filing, a significant contract. These filings hit the SEC’s public database and move markets. Investors, analysts, and automated trading systems all pay close attention to them.
The problem is not the filings themselves. It is the classification system sitting on top of them.
Why Current SEC Item Codes Fall Short
The SEC attaches item codes to 8-K filings to indicate the type of event being disclosed. In theory, this should make filtering and analysis straightforward. In practice, the codes are coarse enough to be misleading. A single item code can simultaneously cover routine administrative changes and the departure of a chief executive — two events with very different market implications. Some of the most market-moving disclosures end up in catch-all categories that tell analysts almost nothing specific about what actually happened.
This is not a minor inconvenience. For anyone trying to study how particular types of corporate events affect stock prices — or build automated systems that flag disclosures by event type — the existing classification scheme creates serious noise. The granularity simply isn’t there.
Introducing a Fine-Grained Taxonomy for Event Tagging
The Three-Tier Taxonomy of 119 Event Types
Dolphin’s system addresses this gap by building an entirely new classification layer. Rather than relying on SEC item codes, it applies a fine-grained event taxonomy with three tiers and 119 distinct event types to 8-K filings. The depth of that taxonomy is what makes this approach different from prior efforts. A three-tier hierarchy allows both broad categorization and highly specific labeling, depending on what a researcher or analyst needs.
The practical challenge with any fine-grained labeling system at scale is reliability. Large language models can assign labels quickly, but without mechanisms to verify those labels against the source text, errors accumulate silently. This is where the system’s architecture makes its most important design choices.
Two-Stage Tagging Process with Quote Anchoring and Quality Scoring
The tagging pipeline operates in two distinct stages. In the first stage, the model’s output is constrained to valid entries within the taxonomy — preventing the system from inventing or hallucinating categories — and every assigned tag is anchored to a verbatim quote from the filing itself, validated through fuzzy n-gram matching. This grounding step is critical: it means every label can be traced back to actual language in the source document, not just to the model’s interpretation of it.
The second stage re-grades each cited quote against the category definition to produce a quality score. This is more than a simple confidence measure. Ablation testing shows the quality score is only properly calibrated when assigned in this dedicated second pass — running it as part of the first stage produces miscalibrated scores that do not reliably predict accuracy.
The implication matters: the architecture is not arbitrary. The two-stage design is functionally necessary for the quality scoring to work as intended.
Application Results and Dataset Release
Scale of Application and Tagging Volume
The system was applied to 292,984 filings spanning 2022 to 2026, producing a total of 601,088 grounded event tags. That is a substantial corpus — roughly four years of live SEC disclosures processed through a consistent, auditable classification pipeline. The resulting dataset has been made publicly available, which means researchers and practitioners working on financial event analysis, market microstructure, or NLP can build directly on this foundation without replicating the infrastructure.
Improvements in Tagging Precision with Quality Scores
The headline accuracy result is striking. Evaluated over 5,125 stratified tags by a large language model judge, tagging precision rises monotonically with the quality score — from 12% at the low end to 96% at the high end. Unsupported tags, where a label cannot be traced back to the source text, drop from 8% to near zero as quality thresholds increase.
What this means in practice: researchers and analysts using the dataset can filter by quality score to control the trade-off between coverage and accuracy. A high-precision subset of the dataset — filtered to only the highest-quality tags — would carry 96% precision. A broader, lower-threshold filter would increase coverage while accepting some noise. That flexibility is a meaningful feature for real-world applications where different use cases demand different accuracy standards.
Public Availability of Grounded Event Tags Dataset
Beyond the technical architecture, the public release of the grounded event tags dataset may be the most consequential output of this work. Financial event studies have historically depended on hand-coded samples, proprietary data, or coarse SEC classifications. A large-scale, publicly available dataset of 601,088 verified event tags across four years of 8-K filings opens up new possibilities for reproducible research on how specific event types affect market behavior.
The event study included in the paper reinforces this point. By analyzing unsigned abnormal returns without using any language model — relying purely on the taxonomy labels — the study confirms that the fine-grained classification genuinely separates economically distinct events that share a single SEC item code. That is a meaningful empirical validation: the taxonomy is not just a labeling exercise, it is capturing real differences in how markets respond to different types of corporate disclosures.
For the broader field of financial event tagging using large language models, this work sets a methodological benchmark. The combination of constrained output, verbatim grounding, two-stage quality scoring, and large-scale validation produces a system where accuracy claims are verifiable — not just asserted. As LLMs become more embedded in financial analysis pipelines, the question of how to trust their outputs at scale may be just as important as the outputs themselves.
FAQ
What role do Form 8-K filings play for U.S. public companies?
Form 8-K filings are the primary channel through which U.S. public companies disclose material events, such as executive changes, mergers, or significant financial developments.
Why are current SEC item codes insufficient for event categorization?
Existing SEC item codes are coarse, grouping routine administrative changes and major events like executive departures under single categories, making it difficult to distinguish economically distinct disclosures.
How does the proposed two-stage tagging system work?
It first tags disclosures with taxonomy-constrained outputs anchored to verbatim quotes from the source filing, then regrades each citation against category definitions in a dedicated second pass to assign calibrated quality scores.
How accurate is the event tagging system?
Tagging precision improves from 12% to 96% as quality scores increase, according to large language model evaluations conducted over 5,125 stratified tags.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
