The data

One mandatory federal filing, and everything that follows from it

Every ERISA-covered employee benefit plan in the United States is required to file a Form 5500 with the Department of Labor each year. The Department publishes the structured datasets as a public government record. That is the entire source.

This page is what a careful buyer asks for before the demo: exactly what is held, exactly what is not, and where the limits are.

Provenance

Where it comes from, and what that means for you legally and commercially.

A public US government record

The DOL EBSA Form 5500 structured datasets, downloaded from the Department’s own file service. Public record, published by the agency that collects it.

Not licensed, not scraped

Nothing here is licensed from another vendor, scraped from a competitor’s product, or bought from a data broker. There is no third-party redistribution restriction sitting behind this data, and no supplier who can withdraw it.

Reproducible in principle

Anyone can download the same files. What is not trivially reproducible is the identity resolution and the outcome definitions — which is why those are the parts documented in the most detail.

What is held on this build

2.66 GB of compressed source files across 8 datasets, form years 2009–2024. Generated from the same snapshot the product serves.

Dataset	Form years	What it carries
5500	2009–2024	The main annual return. Sponsor, plan, participant counts, plan characteristics.
5500_SF	2009–2024	Form 5500-SF, the short form for smaller plans. A separate dataset with SF-prefixed columns — not a subset of the main return, and treating it as one loses every plan that crossed the threshold.
SCH_C	2009–2024	Schedule C header. A header record only: the provider names and compensation are not in this file.
SCH_C_PART1_ITEM1	2009–2024	Schedule C Part 1 Item 1 — service provider identification.
SCH_C_PART1_ITEM2	2009–2024	Schedule C Part 1 Item 2 — where provider names and compensation actually live. This is the service-provider relationship table.
SCH_C_PART1_ITEM3	2009–2024	Schedule C Part 1 Item 3 — indirect compensation detail.
SCH_C_PART3	2009–2024	Schedule C Part 3 — terminated accountants and actuaries.
SCH_H	2009–2024	Schedule H, the large-plan financial statement. Assets, administrative expenses, and the fee series the renegotiation signal is built on.

What is deliberately not used

Stated because the absences are load-bearing. A claim this product does not make is a claim you should not assume it makes.

Attachments, including the Schedule of Assets. They sit behind a separate request and nothing here depends on them. This matters because a holdings-level claim in this market usually rests on that attachment, and Fin360AI does not make one.
Anything about individual participants. Form 5500 is a plan-level and sponsor-level return. There are no individual participant records in the source and none in the product.
Defined benefit plans, in the current signals. The panel is filtered to defined contribution plans. That filter is verified by a standing check that confirms it bites rather than that the result looks plausible — an unfiltered panel is indistinguishable from a filtered one by inspection.
Asset-manager DCIO desks, as a served audience. They sell into plans rather than being paid by them, so they never appear on a Schedule C. They cannot be counted from this data, let alone served from it.

Why the newest year is always partial

And why it is labelled rather than hidden.

Filings arrive nine to eleven months after each plan’s year end. A year is therefore still filling in long after it has ended, and the most recent year in the archive is never complete. Every ranked list states what share of a full year has been filed rather than presenting a half-filed year as a shrinking market.

The last year for which an outcome can be resolved at all is 2022 — to know whether a plan changed recordkeeper in a given year, the following year’s filing has to exist. That is why validation stops where it does, and it is a property of the filing regime rather than a choice.

Staleness is decided by the server, not by our clock. The Department folds amendments into the same download URL, so a file goes stale without its year changing. Each dataset-year is checked against its published size and last-modified date on every refresh, and only what has genuinely been superseded is re-fetched.

The central risk: identity, not modelling

Five separate defects in this project share one root cause, and it is worth understanding before you trust any number here.

Anywhere an identifier can change, a change of identifier will masquerade as a change of behaviour

Plan side. Employer-identification or plan-number renumbering looks like a termination. About four in ten apparent disappearances are plans dropping below 100 participants and moving to the short form — they are alive. The lineage resolver classifies the large majority of disappearances rather than counting them as deaths.
Provider side. A provider changing its own tax ID looks like growth of several hundred times over.
Outcome side. A provider acquisition looks like thousands of plans independently switching recordkeeper in a single year. In one year it was one firm renaming itself. Those plan-years are dropped from the outcome rather than recorded as non-events, because an acquired plan may also have switched on its own.

For a product whose entire value is detecting change, the lineage resolver is closer to the core of the build than the modelling is. The method page carries the standing checks that keep it honest, and their status on the current build.

Known limits

The ones we would rather you read before buying than discover afterwards.

Terminations and mergers are not separable

A plan that terminates and a plan that merges away look the same in the filings today. Both warrant a call; they warrant different calls. The one sub-case that is an opportunity rather than a loss is not addressable yet.

Provider relationships are presence-based

A provider counts as serving a plan when it appears on that plan’s Schedule C, not when it is the highest-compensated party. Compensation ranking turned out to be unstable enough to make an earlier proxy unusable.

Base rates drift

One signal’s base rate moved materially between the fit window and the validation window. Model performance degrades as the environment drifts away from the one the model was fitted in, and this is a market with consolidation waves and rate regimes that do exactly that.

Snapshot generated 2026-08-24. Archive last checked against the source 2026-08-24.