Evidence
A model built to be checked.
Anyone can publish an accuracy figure. This page sets out what MTR measures, how it is validated, what it refuses to claim, and how any single number in a study can be traced back to the computation that produced it.
What is measured, and what it is measured against
MTR's validated asset is county-level entry and expansion forecasting: whether a carrier will be operating in a county it does not currently serve, and where an existing carrier will extend its footprint.
That asset is validated against held-out years rather than against the data it was fitted on. The model is trained on observed behavior up to a cut point, then asked to predict years it has not seen, and scored against what carriers actually did.
Held-out years, not in-sample fit
A model scored on the data it was fitted to will always look accurate. The validation program withholds years and scores against them.
Every county, not a sample
Scoring runs across the full national county set rather than a chosen subset, so the figure is not the product of where the model happened to do well.
13 years of observed decisions
The model is grounded in 13 years of every carrier's actual county decisions rather than in assumptions about how carriers behave, which is what makes an out-of-sample year a real test.
92%
Accuracy, 3 years ahead
This figure refers to county-level entry and expansion forecasting on held-out years, and at present to expansion within a state a carrier already operates in. It is not a claim about exit, about benefit design, about enrollment volume, or about equilibrium output, none of which it measures.
What MTR will not tell you
These are the boundary of the claim, and stating them is what makes the rest of the page worth reading.
No exit or withdrawal probabilities
The withdrawal model has not cleared our internal validation bar. Rather than publish a number we do not trust, withdrawal signals travel as an early-warning watchlist that names counties and states the limitation alongside them. A number from an unvalidated model is worse than no number.
Equilibrium is characterized, not forecast
Equilibrium output describes the structural configurations a market can settle into and how much support each carries, meaning how much of the scored field backs it once rivals respond. It is not a point prediction of what will happen, and it is not presented as one.
No financial or share projections
MTR does not produce revenue estimates, margin forecasts, or market share projections. Those quantities are not in the current output contract, and the engine does not compute them yet.
Weak support is reported as weak
Where a result rests on thin structural support, the study says so rather than smoothing it into a single confident figure. Confidence tiering travels with the result.
Any figure, back to the computation that made it
The unit underneath every figure is a scored world: one complete version of the market, every carrier's move in every county, scored. Every output is deterministic and content-addressed, so the same 4 values always recover the same result, and MTR will re-execute any delivered run on request and match it.
Run
The specific execution that produced the output, identified and immutable.
Model version
The code and formulation. Incremented whenever a change would produce different results from the same inputs.
Dataset version
The public data vintage the run consumed, pinned rather than described.
Bundle schema
The output contract itself, versioned, so a reader knows what shape the figures arrived in.
For model risk and actuarial reviewers
Checkable means three things here, and each is bounded. Any figure traces to the run, the model version, and the data vintage that produced it. Any figure in a delivered study recomputes from the scored worlds beneath it, and the world-level record for your footprint is available to your reviewers under the licence, so the math can be checked by your own resources. And because every input is public, what a study said can be scored against what CMS went on to publish.
It starts with the record
13 years of every carrier's county decisions, filed in public: who entered, who left, who held, and what they offered while they did it. Nobody answers a survey, nothing is scraped, and nothing connects to your systems. The record is long enough to know every carrier by what it did.
The moves
Every service-area decision by every carrier, county by county, from the CMS landscape and crosswalk record.
The stakes
Enrollment, plan design, and Star Ratings from the same public publications, at the same county grain.
The ground
County economics and demographics from public sources, so the board the game is played on is real.
What the engine keeps noticing
Findings from the working log, stated at the resolution they were measured.
- F-114Exit reclassification
A third of one recent year's county "exits" were reclassifications. The carriers never left. The study counts a departure only when the county actually loses the carrier.
- F-087Entry concentration
Real county entries are rare: well under 1% of the possibilities in any year. The ones that happen cluster where footprint, momentum, and county economics already point.
- F-042Withdrawal asymmetry
Entry shows up in the record before it happens. Exit mostly does not. That asymmetry is why exit signals travel as a watchlist and entry travels as a ranking.
Aggregate findings from public data. No carrier is named, by rule.
What we publish
The MTR white paper
The simulation engine, the validation program, and the evidence behind every claim on this site.
Coming soon
Validation report
The held-out-year scoring program and its results, at county resolution.
Annual, first edition to follow
Research notes
Occasional writing on Medicare Advantage market structure and competitive behavior.
Subscribe in the footer
Bring us something to check.
A live walkthrough of the study on your counties, or a conversation with the people who built the validation program. Both are available.