Roughly 150 to 200 U.S. credit unions disappear into mergers every year. The population has fallen from over 10,000 in the 1990s to around 4,300 today. If you sell to credit unions, advise them, or buy them, knowing twelve months early which ones are on that path is worth real money. A vendor does not want to implement a core system at an institution that will be absorbed next spring. An acquirer wants to open the conversation before three other acquirers do.
So the question is not whether merger prediction is useful. It is whether any given model actually predicts, or merely describes.
What a merging credit union looks like before it merges
Mergers in this industry are overwhelmingly about scale. The problem shows up on the balance sheet long before it shows up in a board vote. The pattern that recurs:
- Sub-scale assets. Fixed cost — compliance, cybersecurity, a digital banking stack members now expect — does not shrink with the institution. Below roughly $50M in assets, that cost is spread across too few members to carry.
- Membership going backwards. A shrinking member count is the clearest single predictor. It removes the one thing that could eventually fix the scale problem.
- Earnings that cannot fund reinvestment. A return on assets near or below zero means the institution is not generating the capital to buy its way out.
- Capital thinning toward the regulatory floor. The 7% net worth ratio separates "well capitalized" from everything below it. It is a real cliff, and boards behave differently as they approach it.
- An aging charter with an aging field of membership. A sixty-year-old single-employer charter whose sponsor has shrunk is a specific and common story.
- Credit deteriorating faster than the institution can absorb. Charge-offs climbing on a thin capital base compress the timeline.
None of these is decisive alone. A small credit union with strong earnings and growing membership is not a merger candidate. A large one shedding members with weak earnings may well be. The signal is in the combination, and reading a combination is what a scorecard is for.
Scorecard, then calibration
There are two separate jobs here. Models that mix them together are hard to trust.
The scorecard turns those signals into one 0–100 susceptibility score: a measure of how strongly the balance sheet shows the pressures above. It scales each component to a 0–1 range against sensible bounds. Then it blends the components at published weights. Its virtue is that you can read it. Every score breaks back down into weight × feature. So you can see that an institution scored 74 because of membership decline and charter age, not because a black box said so. A scorecard is an argument you can audit.
The calibration is a second, fitted model. It learns from confirmed historical merger events and turns the scorecard's inputs into an actual one-year probability. Its virtue is that a probability means something a score does not. "18%" is a claim you can be wrong about, and you can measure how wrong.
You want both. A scorecard alone tells you the ranking but not the odds. A fitted probability alone gives you a number with no explanation attached. That is unusable in a conversation with a board.
The only validation that counts is out-of-time
This is where most merger lists fall down. It is also the question worth asking any vendor.
A model trained and tested on the same period will look excellent. It has already seen which institutions merged, so it can find the features that set them apart. Some of those features were consequences of the merger process, not causes of it. That is not prediction. That is description with extra steps.
Out-of-time validation trains the model on one period and tests it on a later period it has never seen. You score institutions as of a date, then check what actually happened in the following year. It is a harder test, and it always produces worse-looking numbers. That is exactly why it is the one to insist on.
Three numbers to ask for:
| Metric | What it answers | What "good" looks like here |
|---|---|---|
| AUC, out-of-time | Given a merged and an unmerged institution, how often does the model rank the merged one higher? | 0.75+. Below 0.65 is barely better than sorting by asset size; above 0.90, ask what leaked. |
| Top-decile capture | What share of actual mergers land in the model's riskiest 10%? | 40%+. This is the number that translates to your call list. |
| Lift over random | How much better than a random list of the same length? | 4×+. |
Two more questions matter. Without them, the first three numbers mean very little.
How many actual mergers were in the test window? An AUC of 0.90 measured on eleven events is a statement about eleven events. Mergers are rare, and held-out windows (the later period kept back for testing) are short. A vendor who will quote you three decimal places but not a count is telling you which of the two they would rather you looked at.
And what counts as a positive? In other words, what event is the model trying to predict? A model that predicts any merger is predicting a different, easier thing than one that predicts a distress merger. The reason is that most mergers are strategic combinations between two healthy charters. Widening the label (the definition of a positive) roughly quadruples the number of positives and flatters every metric in the table. Ask what the label was before you compare two AUCs.
For reference, the merger model behind this site publishes an out-of-time AUC of 0.86 against a distress label. That label counts only mergers a regulator attributes to the credit union's own condition. Here is what an AUC of 0.86 means. Take one credit union that had a distress merger and one that did not. The model ranked the merging one higher eighty-six times out of a hundred. The top-scored tenth of the list (the top decile) caught roughly 55% of those mergers. That is about 5.5× lift: 5.5× as many as a random list of the same length would catch.
Those figures rest on 89 events in the held-out window. With that many events, the standard error on the AUC is near 0.04. (The standard error is roughly how far a measured figure can sit from the true one by chance.) That estimate is itself optimistic. The fit pools consecutive as-of quarters, meaning scoring snapshots taken one quarter apart. Rows across those quarterly groups are not independent of each other. Read the AUC as mid-0.8s. It is not a number to defend to the third decimal, and we do not. The held-out window also contained no distress merger above $500M in assets. Above that size, the model is untested rather than weak. Those are backtest results on historical data, not a guarantee about any future period. The known limitations are published rather than buried.
An AUC in the mid-0.8s is genuinely useful and genuinely not clairvoyant. The ranking is strongly informative, and any individual prediction is uncertain. Anyone quoting you a 0.95 on a problem like this has done one of two things. They have leaked future information into the training set, or they are measuring something other than what you need.
Susceptibility is not the same as value
A second, subtler mistake: treating "likely to merge" as "worth acquiring."
They are close to opposites. The institutions with the highest susceptibility scores are small, shrinking and thinly capitalized. That is precisely why they merge. It is also precisely why acquiring one may absorb capital without buying much. What an acquirer actually wants is a franchise: a low-cost core deposit base, a real member relationship, clean credit, enough scale to matter.
So susceptibility and franchise value have to be scored separately. The interesting set is where the two lists overlap: institutions that are plausibly available and worth having. That set is always much smaller than either list alone. It is the only one worth a banker's time.
Four questions to ask before you buy a merger list
- Is the validation out-of-time, and what is the AUC? If the answer is a single accuracy percentage with no time split described, the model has not been tested.
- Does every score break down into its parts? You should be able to see which factors produced a given institution's number. If you cannot, you cannot defend it in a meeting, and you cannot catch it when it is wrong.
- Is susceptibility scored separately from franchise value? A single list that mixes the two is a list of small institutions.
- How often does it refresh, and what does it tell you changed? A static ranking is worked through in a quarter. The recurring value is in the changes, not the level: who moved from one score band to another this quarter, and why.
What this does not tell you
A merger is a board decision made by people. Succession, sponsor relationships, regulatory conversations and personalities drive it, and no balance sheet records them. A model reading quarterly financials is reading the pressure, not the decision. It will rank the pressure well and miss the institution whose CEO simply decided to retire.
Used as a ranking that tells you where to spend attention first, that is fine. Used as a prediction about a specific named institution, it is not. No score on this site, or anywhere else, is investment, credit or merger advice. See the disclaimer and disclosures.