Form 5500 signals

Every ERISA retirement plan in the country, ranked by what it is about to do

Ranked lists of plans about to make a decision — and how well each ranking actually works on years the model never saw. Including the years it works badly.

Built on Form 5500, a mandatory annual return. That single fact is why this is one of the few commercial datasets with a complete denominator — every plan is in it, not just the ones that appeared in the trade press or answered a survey.

Form years

2009–2024

8 datasets

What a subscriber actually works from

A ranked list of named plans, each carrying the reason it ranked. The grid below is what that ordering does to a hit rate, measured on years the model was never fitted on.

Plan year 2024 · 85% filed so far · 2/4 gates

3 of 100

Called in no particular order, 1 in 33 plans changes recordkeeper in the next year.

Measured on 150,904 plan-years from 2019–2022, which the model was never fitted on. Top-decile lift 3.19x, AUC 0.695.

The top of the list, and why each plan is on it

Open the full list →

1
Snow Software 401(K) Plan
Austin, TX · 10-50M · $12.0M · 58 participants
0.973
- asset mix moved sharply last year
- large share of expense going to service providers
- high admin cost per participant vs peers
2
Yuma Regional Medical Center 401(K) Plan
Yuma, AZ · 50-500M · $180.5M · 3,237 participants
0.969
- asset mix moved sharply last year
- large plan
- high admin cost per participant vs peers
3
█████████ ████ 401(K) Retirment Plan
NC · 50-500M · 848 participants
0.965
- asset mix moved sharply last year
- high admin cost per participant vs peers
4
████ █████████ 401(K) Retirement Plan
FL · 50-500M · 0 participants
0.964
- asset mix moved sharply last year
- large plan
- high admin cost per participant vs peers
5
███████ █████████ 401(K) Plan
KY · <10M · 165 participants
0.961
- asset mix moved sharply last year
- little paid to service providers
- low admin cost per participant
6
███████ 401(K) Plan
OR · <10M · 103 participants
0.960
- asset mix moved sharply last year
- low admin cost per participant

Two rows are shown with the sponsor intact so the list can be checked against the DOL’s own published file; the rest are held back on this page, not hidden from you. Every row carries its own reason, which is the part a score alone does not give you.

150,904 plan-years scored

15,090 in decile 1

1,450 expected to change recordkeeper

A backtest on years the model never saw — not a forecast.

Why this data is different

The complete denominator is the product

Not the modelling. Anyone can fit a classifier. The reason a ranking here means something is that there is nothing outside the frame.

Mandatory, not voluntary

Every ERISA-covered plan files, every year, or it is out of compliance. There is no sampling frame to argue about and no response bias to correct for, because there was never a response.

Public record, not a licence

The source is the Department of Labor’s own published datasets. Nothing here is licensed from a competitor or scraped from another vendor’s product, so nothing here can be withdrawn from under you.

Identity resolved before anything is modelled

Anywhere an identifier can change, a change of identifier will masquerade as a change of behaviour. A plan renumbering looks like a termination; a provider changing its own tax ID looks like explosive growth. That resolver is closer to the core of this build than the models are.

From the standing checks on this build

2,121,815 panel rows from 2,121,815/4,299,671 DC (49%) of all filers; 71% of Schedule H filers

The check verifies that the plan-type filter bites, not that the panel looks plausible — an unfiltered panel is indistinguishable from a filtered one by inspection, which is exactly how the wrong universe shipped the first time.

The backtest, in full

What each signal measured, out of time

Fitted on plan-years up to 2018, scored on 2019–2022, which the fit never sees. The four gates were fixed before anything was measured. This table is generated from the same snapshot the product serves, so it cannot drift away from what a subscriber sees. How the gates work.

Signal	Base rate	Top-decile lift	Hit rate	Absence lift	AUC	Gates
Recordkeeper Change Plans most likely to change recordkeeper next year.	3.0%	3.19x	10% · 1 in 10	n/d	0.695	2/4
Plan Termination Risk Plans most likely to terminate or merge away next year.	0.5%	2.81x	1% · 1 in 76	n/d	0.660	1/4
Fee Renegotiation Plans whose administrative cost per participant is about to fall, and stay down.	11.5%	2.88x	33% · 1 in 3	2.78x	0.685	2/4

Hit rate is base rate times lift: what you experience working down the list. Absence lift scores only plan-years where nothing visible was happening — the part a competitor with the same public filings cannot copy by reading the news. n/d means the result straddles its own threshold depending on the random seed, so neither a pass nor a miss is a finding.

What the signals say about themselves

Each verdict and caveat below is pre-registered by the pipeline and rendered here verbatim. They are not edited into benefit copy on the way to this page.

Recordkeeper Change

lift passes, absence not determinable

The absence lift lands between roughly 1.78x and 2.02x across five draws, straddling its own 2.0 gate, so the moat claim can be neither made nor dismissed on this evidence. What is not in doubt is the mechanism: the model leans hardest on prior-year asset-mix movement, which is visible in the same public filings to anyone who reads them. Sell this as a better ranked list with real economics, never as something a competitor cannot build.

Plan Termination Risk

absence gate not determinable

This signal's headline result is not reproducible enough to call. Across five draws the absence lift lands between roughly 1.87x and 2.07x — it straddles its own 2.0 gate, so neither a pass nor a miss is a finding. At a 0.48% base rate the top decile contains only a few hundred positives, and ordinary floating-point variation moves enough of them to flip the verdict. Treat it as approximately at the threshold, not above it. Separately, mergers are buried inside terminations and cannot currently be separated, so the one sub-case that is an opportunity — a plan someone acquired rather than lost — is not addressable yet.

Fee Renegotiation

best economics, misses the lift gate

The lift gate misses at 2.88x, but lift is the wrong headline for this signal: at an 11.5% base rate the top decile still hits roughly one in three, far better in absolute terms than either other signal here. The naive fee-percentile sort returns 1.22x against the model's 2.88x, so the mechanical reversion this outcome was designed to survive is not what is driving it. The number to watch is the base rate, which fell 22% between the fit and validation windows — more drift than the other outcomes show, and worth re-checking on the next refresh.

The one that is withheld

Built, backtested, and then pulled. It is listed on the public site rather than quietly deleted, because a buyer who finds the defect themselves has no reason to trust anything else here.

Provider Consolidation — withdrawn — circular outcome

Backtested at AUC 1.0000, which is a leak rather than a result. The outcome is defined as top_dest_share >= 0.40 while top_dest_share is also a feature, so the model recovered one threshold on its own label and every other feature scored an importance of exactly zero. Two genuine defects were fixed on the way to finding it — a duplicated join producing destination shares above 1.0, and provider EIN consolidation read as eight separate acquisitions — and the outcome series is now plausible at 1.4-5.6% a year. The signal is nonetheless not predictive as specified and is withheld until it is rebuilt to predict year T+1 from information available at year T.

Who it is for

The same list, read four different ways

One buyer's prospect is another buyer's client about to leave. The identical ranked list is offensive intelligence to a challenger and defensive intelligence to the incumbent serving them, so the persona decides which list leads and what a row is called.

Adviser / consultant

I want plans that are reviewing what they pay.

Win fee-sensitive plans and benchmark the ones you hold.

A row is a plan in review.

Recordkeeper / TPA

I want plans that are about to change provider.

Prospect competitors' books.

A row is a prospect.

Incumbent, defending a book

I want to know which of my plans is about to leave.

Retain the clients you already serve.

A row is a client at risk.

Acquirer / investor

I want firms whose book is coming available.

Buy providers, not plans.

A row is a target.

Each one carries an honest caveat on its own home screen rather than buried in a footnote. Read all four in full.

What this is not

Stated here rather than in the small print, because these are the four assumptions a reader of a page like this usually arrives with.

Not advice. A score is a model estimate that a plan resembles the historical pattern of plans that made a particular decision. It is not investment advice, not a fiduciary opinion, and not a recommendation about any plan, sponsor or provider. Readers here are often ERISA fiduciaries; a plan-level signal must never be read as a fiduciary recommendation about that plan.
Not a forecast. Every published figure is a backtest against what already happened. It describes a historical period the model was not fitted on. It is not a promise about the next one.
Not participant data. Form 5500 is a plan-level and sponsor-level return. There are no individual participant records in it and none in this product.
Not a complete year. Filings arrive nine to eleven months after each plan’s year end. The newest year is always partial, and every ranked list states what share of it has been filed rather than presenting a half-filed year as a shrinking market.

What this is

The short version, for someone who landed here from a search result.

What does Fin360AI actually sell?: Ranked lists of U.S. retirement plans that look like they are about to make a decision — change recordkeeper, renegotiate what they pay for administration, or terminate. Each list ships with the out-of-time backtest that says how well the ranking works, including where it does not work.
What is Form 5500?: The annual return every ERISA-covered employee benefit plan is required to file with the U.S. Department of Labor. It is a mandatory filing, and the Department publishes the structured datasets as a public government record.
Why does a mandatory filing matter so much?: Because it gives the dataset a complete denominator. Every plan is in it, not only the ones that showed up in the trade press or answered a survey. A ranking is only meaningful against the whole population, and most commercial lists in this market are built on a sample nobody can characterise.
Is this an AI product?: No, and calling it one would be a smaller claim than the truth. The models are gradient-boosted classifiers over engineered features from the filings. What is hard here is not the modelling — it is resolving plan and provider identity across years so that a change of identifier does not masquerade as a change of behaviour. The method page explains why.
How often does it refresh?: Monthly. The Department folds amendments into the same download URL, so staleness is decided by the server rather than by our clock: each dataset-year is checked against its published size and last-modified date, and only what has actually been superseded is fetched.

The rest of the questions, including the awkward ones.

The honest way to evaluate this

Look at plans you already know. Sign in and check the ranking against your own judgment on a book you can verify, before anyone talks about price. If the list does not tell you something you did not know, we would rather you found that out in ten minutes than in a quarter.

Snapshot generated 2026-08-24. Last resolvable outcome year 2022.