There’s a quiet however vital drawback on the coronary heart of U.S. monetary disclosure: the system designed to tell markets about main company occasions is, in some ways, too blunt to be helpful. SEC 8-Okay occasion extraction has lengthy been constrained by merchandise codes that lump wildly completely different occurrences into the identical bucket — a routine board committee replace and a CEO resignation may share a single class. A brand new analysis system from Rian Dolphin goals to repair that, and the outcomes counsel giant language fashions can do excess of summarize textual content — they’ll construct dependable, granular maps of company occasions at huge scale.
Key takeaways
- Type 8-Okay filings are the first disclosure channel for U.S. public corporations reporting materials occasions, however current SEC merchandise codes are too coarse to differentiate economically distinct occasions.
- A brand new two-stage system applies a three-tier taxonomy of 119 occasion varieties to 8-Okay filings, anchoring each tag to a verbatim quote from the supply textual content.
- Utilized to 292,984 filings from 2022 to 2026, the system generated 601,088 grounded occasion tags, now launched publicly.
- Tagging precision rises from 12% to 96% as high quality scores enhance, with unsupported tags falling to close zero on the highest thresholds.
- An occasion examine confirms the taxonomy separates economically distinct occasions — with out utilizing any language mannequin — validating the method by irregular return evaluation.
Limitations of Present SEC 8-Okay Disclosures
The Position of Type 8-Okay in U.S. Public Firm Disclosures
Type 8-Okay is the necessary reporting mechanism U.S. public corporations use when one thing materials occurs — a merger, an government departure, a chapter submitting, a big contract. These filings hit the SEC’s public database and transfer markets. Buyers, analysts, and automatic buying and selling methods all pay shut consideration to them.
The issue shouldn’t be the filings themselves. It’s the classification system sitting on high of them.
Why Present SEC Merchandise Codes Fall Quick
The SEC attaches merchandise codes to 8-Okay filings to point the kind of occasion being disclosed. In principle, this could make filtering and evaluation easy. In apply, the codes are coarse sufficient to be deceptive. A single merchandise code can concurrently cowl routine administrative modifications and the departure of a chief government — two occasions with very completely different market implications. Among the most market-moving disclosures find yourself in catch-all classes that inform analysts nearly nothing particular about what really occurred.
This isn’t a minor inconvenience. For anybody making an attempt to check how specific forms of company occasions have an effect on inventory costs — or construct automated methods that flag disclosures by occasion sort — the prevailing classification scheme creates severe noise. The granularity merely isn’t there.
Introducing a Positive-Grained Taxonomy for Occasion Tagging
The Three-Tier Taxonomy of 119 Occasion Sorts
Dolphin’s system addresses this hole by constructing a wholly new classification layer. Fairly than counting on SEC merchandise codes, it applies a fine-grained occasion taxonomy with three tiers and 119 distinct occasion varieties to 8-Okay filings. The depth of that taxonomy is what makes this method completely different from prior efforts. A 3-tier hierarchy permits each broad categorization and extremely particular labeling, relying on what a researcher or analyst wants.
The sensible problem with any fine-grained labeling system at scale is reliability. Massive language fashions can assign labels rapidly, however with out mechanisms to confirm these labels in opposition to the supply textual content, errors accumulate silently. That is the place the system’s structure makes its most vital design decisions.
Two-Stage Tagging Course of with Quote Anchoring and High quality Scoring
The tagging pipeline operates in two distinct phases. Within the first stage, the mannequin’s output is constrained to legitimate entries inside the taxonomy — stopping the system from inventing or hallucinating classes — and each assigned tag is anchored to a verbatim quote from the submitting itself, validated by fuzzy n-gram matching. This grounding step is vital: it means each label might be traced again to precise language within the supply doc, not simply to the mannequin’s interpretation of it.
The second stage re-grades every cited quote in opposition to the class definition to provide a high quality rating. That is greater than a easy confidence measure. Ablation testing exhibits the standard rating is just correctly calibrated when assigned on this devoted second move — working it as a part of the primary stage produces miscalibrated scores that don’t reliably predict accuracy.
The implication issues: the structure shouldn’t be arbitrary. The 2-stage design is functionally mandatory for the standard scoring to work as meant.
Software Outcomes and Dataset Launch
Scale of Software and Tagging Quantity
The system was utilized to 292,984 filings spanning 2022 to 2026, producing a complete of 601,088 grounded occasion tags. That could be a substantial corpus — roughly 4 years of stay SEC disclosures processed by a constant, auditable classification pipeline. The ensuing dataset has been made publicly obtainable, which implies researchers and practitioners engaged on monetary occasion evaluation, market microstructure, or NLP can construct immediately on this basis with out replicating the infrastructure.
Enhancements in Tagging Precision with High quality Scores
The headline accuracy result’s hanging. Evaluated over 5,125 stratified tags by a big language mannequin decide, tagging precision rises monotonically with the standard rating — from 12% on the low finish to 96% on the excessive finish. Unsupported tags, the place a label can’t be traced again to the supply textual content, drop from 8% to close zero as high quality thresholds enhance.
What this implies in apply: researchers and analysts utilizing the dataset can filter by high quality rating to regulate the trade-off between protection and accuracy. A high-precision subset of the dataset — filtered to solely the highest-quality tags — would carry 96% precision. A broader, lower-threshold filter would enhance protection whereas accepting some noise. That flexibility is a significant function for real-world purposes the place completely different use instances demand completely different accuracy requirements.
Public Availability of Grounded Occasion Tags Dataset
Past the technical structure, the general public launch of the grounded occasion tags dataset often is the most consequential output of this work. Monetary occasion research have traditionally trusted hand-coded samples, proprietary information, or coarse SEC classifications. A big-scale, publicly obtainable dataset of 601,088 verified occasion tags throughout 4 years of 8-Okay filings opens up new potentialities for reproducible analysis on how particular occasion varieties have an effect on market habits.
The occasion examine included within the paper reinforces this level. By analyzing unsigned irregular returns with out utilizing any language mannequin — relying purely on the taxonomy labels — the examine confirms that the fine-grained classification genuinely separates economically distinct occasions that share a single SEC merchandise code. That could be a significant empirical validation: the taxonomy is not only a labeling train, it’s capturing actual variations in how markets reply to several types of company disclosures.
For the broader discipline of monetary occasion tagging utilizing giant language fashions, this work units a methodological benchmark. The mixture of constrained output, verbatim grounding, two-stage high quality scoring, and large-scale validation produces a system the place accuracy claims are verifiable — not simply asserted. As LLMs develop into extra embedded in monetary evaluation pipelines, the query of methods to belief their outputs at scale could also be simply as vital because the outputs themselves.
FAQ
What position do Type 8-Okay filings play for U.S. public corporations?
Type 8-Okay filings are the first channel by which U.S. public corporations disclose materials occasions, corresponding to government modifications, mergers, or vital monetary developments.
Why are present SEC merchandise codes inadequate for occasion categorization?
Present SEC merchandise codes are coarse, grouping routine administrative modifications and main occasions like government departures beneath single classes, making it tough to differentiate 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 supply submitting, then regrades every quotation in opposition to class definitions in a devoted second move to assign calibrated high quality scores.
How correct is the occasion tagging system?
Tagging precision improves from 12% to 96% as high quality scores enhance, in keeping with giant language mannequin evaluations performed over 5,125 stratified tags.
Article produced with the help of synthetic intelligence and reviewed by the editorial crew.