Disclosures

1.How we make money

Subscriptions, and nothing else. Revenue comes from the tiers published on the pricing page, priced on the size of the account universe a subscriber can see and export. We also take custom engagements for warehouse-direct delivery and bulk export at the institutional tier, on the same commercial basis.

We do not take money from any of the following, and we would tell you here if we did:

• the institutions the Service scores, in any form;
• advertising, sponsorship or paid placement of any kind;
• referral, introduction or success fees on a transaction the Service surfaced;
• selling or licensing subscriber data to anyone.

2.No institution can influence its own score

Scores are computed by a pipeline from the panel, without manual adjustment for any institution. No institution can pay to be scored, to be scored differently, to be removed from a ranking, to be excluded from a competitor’s territory, or to be shown to a particular buyer. No institution sees its own score before publication or has any right of review. A subscriber’s tier determines how much of the universe they can see; it never determines what any score says.

3.Conflicts we can see, stated plainly

Two structural conflicts exist in a product like this. Neither is hypothetical:

• Both sides of a match may be customers. The buyer-seller and acquirer-target matching features pair institutions. Either side, or both, may hold a subscription. Matching is computed from profile fit alone, and a subscription confers no preference in ranking, ordering or visibility. It does mean we may be selling information about an institution to a party interested in acquiring it, which we think is defensible for public-record-derived analytics but which you are entitled to know.
• Our incentive is for the scores to look decisive. A product that says “it depends” is harder to sell than one that ranks. We manage that by publishing the weights, the decomposition and the out-of-time validation figures rather than only the headline number, and by stating the model limitations in section 5. That is a mitigation, not a cure, and you should read the numbers with the incentive in mind.

4.Data limitations

• It lags. The panel refreshes quarterly. Between refreshes, the Service shows you the last known state of an institution, not its current one.
• It can be restated. Institutions amend filings. A figure correct when we ingested it can be superseded, and our panel reflects the correction only at the next refresh.
• Normalization is interpretive. Mapping heterogeneous account codes onto a single dictionary requires judgment. Where an institution reports unusually, our normalization can be wrong in ways that propagate into every ratio built on it.
• Coverage is not uniform. Small institutions report less, so their derived ratios rest on fewer inputs and their scores are correspondingly less stable.

5.Model limitations

• Mergers are rare events. The positive class is small, which makes any merger model prone to a high false-positive rate at usable thresholds. A high score means “worth a call,” never “this will happen.”
• Validation is historical. Our published AUC and lift figures are measured out-of-time against realized events. They describe past discrimination, not future accuracy, and they will degrade as conditions drift from the fitting period.
• The validation sample is small. Those figures rest on 89 confirmed merger events in the held-out window. That puts the standard error on the published AUC near 0.04 — and that estimate is itself optimistic, because the fit pools consecutive as-of quarters, so rows within a cohort are independent and rows across cohorts are not. Read the headline as a band in the mid-0.8s rather than a figure good to the third decimal, and treat any per-segment cut of it, where the event count is smaller still, as weaker again.
• Correlation, not causation. The scorecards find institutions that resemble a historical pattern. They do not model intent, and they cannot observe a board decision, a private conversation or a leadership change until it shows in the financials.
• Calibration is approximate. A calibrated probability is calibrated in aggregate, across a population. It says very little about the specific institution in front of you.
• Vendor Signal has been graded against a real outcome, and it did not pass. Scored against publicly announced vendor adoptions over two four-quarter cohorts, it returned an AUC of 0.509 and 0.543 — the coin-flip band. Two rankings that know nothing about buying beat it in both: total assets (0.677, 0.665) and member count (0.675, 0.651). In neither cohort did one of the 25 highest-scored credit unions announce an adoption in the following year, and within the $1B–$10B tier one cohort scored 0.368, which is inverted. The label is weak — it observes announcements rather than purchases, so a zero may be a quiet buyer, measured precision is a lower bound, and announcements skew large, which flatters the very size baselines it lost to. That qualifies the result. It does not rescue it, and we would rather you read it here than discover it after paying.
• Buying-propensity scores are inferential. They rest on financial signatures associated with past purchasing, not on any observation of procurement activity at the institution.
• Liquidity Radar has been graded once, and it held. Its stress score was scored against whether a credit union’s borrowings actually rose by at least 1% of assets in the following one to two quarters — AUC 0.76 to 0.83 across three as-of quarters, and never below 0.73 across every horizon and materiality threshold tested. Two limits belong with that number. The label reads total borrowings and the quarterly filing carries no separate brokered-deposit line, so a credit union that met the same need in the brokered market is recorded as a miss: the measured recall is a floor rather than a measurement, though the measured precision is unaffected. And a single graded outcome is not a validated product — the buyer/seller classification and both dollar estimates remain ungraded, and the participation outcome that would grade them is obtainable only from a counterparty.

6.Sources and methods

The panel is built from multiple data sources, reconciled through a proprietary data pipeline. The specific sources and the pipeline itself are proprietary and are not published. What is published is the transformation: the derived ratios, the peer construction, the model weights and the per-account decomposition, so that any score can be interrogated even though its inputs are not disclosed. If that trade-off does not satisfy your compliance requirements, raise it with us before you subscribe rather than after — write to admin@infinidatum.net.

7.Use of AI

The scores are produced by explicit statistical scorecards with published weights, not by a generative model, and no answer the Service gives you about an institution is written by a language model. Where AI assistance is used in our internal pipeline — for example to extract structure from unstructured text before a human-reviewed rule set acts on it — it runs through United States providers, it is never authorized to train on our data or yours, and its output is a candidate for review rather than something published unchecked.

8.Editorial independence and corrections

Methodology decisions are made on analytical grounds and are not subject to commercial review. If a subscriber disputes a score, the remedy is the same as for anyone else: show us the input we got wrong. Corrections are handled as described in the Disclaimer, and a correction that changes a published figure is applied for every subscriber at the same refresh, not selectively.

9.Changes and contact

If any of the arrangements on this page change — if we ever take advertising, a referral fee or money from a scored institution — this page changes first, before the arrangement begins. Questions: admin@infinidatum.net.