# Playoffs, Frozen Four probabilities and rankings

The Research Lab adds three screens to Campus Rink. Scores, team records and player profiles now use the separately sourced dataset documented in REAL-DATA-SOURCES.md. The Research Lab uses dated published ranking facts and an experimental model.

## How to use it

- **Playoffs:** expand “Choose the field & tune the model,” select a team for each seed, then simulate the bracket. Selecting an occupied team swaps the two seeds. Open any resulting matchup to inspect its probability and inputs.
- **Frozen Four:** uses the same selected field and model settings. Calculate 10,000 tournaments to see each selected team’s probability of reaching the Frozen Four, the final and the championship. A new random seed generates another sample; the same seed and inputs reproduce a result.
- **Rankings:** switch between NCAA NPI, USCHO.com, USA Hockey and My rankings. The published tables are immutable snapshots. Your ballot can use any of the three sources as its baseline. Move teams using arrows or a position selector. Ballots save locally by competition, source and chosen week; CSV export includes the comparison source date.
- “Use my ballot order” in either simulator copies the current comparison ballot into the field, then fills any remaining slots from the loaded NPI order. Editing a ballot does not silently change an existing simulation.
- A head-to-head editor shows the loaded sample or missing-data state and allows a separately labeled user scenario. Clearing an override returns to the source sample or missing data.

## Published sources and dates

Checked September 10, 2026. These are historical 2025–26 snapshots, not live 2026–27 polls. The app provides links to the publishers for newer releases.

| Source | Men | Women | Imported coverage |
| --- | --- | --- | --- |
| NCAA NPI | Page labeled through April 5, 2026 | Page labeled through April 5, 2026 | 27 men’s teams: top 25 plus Boston University and Minnesota; top 25 women’s teams |
| USCHO.com | April 13, 2026 | March 23, 2026 | Complete published top 20 men / top 15 women |
| USA Hockey / The Rink Live | April 13, 2026 | March 24, 2026 | Complete published top 20 men / top 15 women, preserving the women’s tie at sixth |

NPI is a calculated index, not a human voters’ poll. The USA Hockey poll was formerly branded USA Today/USA Hockey Magazine. Polls are not the NCAA selection committee’s ranking.

- [NCAA men’s NPI](https://www.ncaa.com/rankings/icehockey-men/d1)
- [NCAA women’s NPI](https://www.ncaa.com/rankings/icehockey-women/nc/npi)
- [USCHO men’s poll on NCAA.com](https://www.ncaa.com/rankings/icehockey-men/d1/uschocom)
- [Published men’s polls on College Hockey Inc.](https://www.collegehockeyinc.com/mens-national-polls/)
- [Published women’s polls on College Hockey Inc.](https://www.collegehockeyinc.com/womens-national-polls/)
- [USA Hockey poll history](https://www.usahockey.com/college)

NCAA’s visible page-date labels and listed records are retained as presented. They may reflect different update cutoffs. In particular, they are not assumed to be final-season records. Where the polls differ on a team record, each table preserves its own source’s record; the simulator uses only the NPI table’s record input.

[Elite Prospects’ 2025–26 NCAA page](https://www.eliteprospects.com/league/ncaa/2025-2026) supplied five men’s tournament results, dated March 29 through April 11, for a small head-to-head sample. This is not complete season history. Missing head-to-head data is omitted rather than fabricated. No women’s head-to-head games have been imported. [HockeyDB](https://www.hockeydb.com/) was requested and checked as a historical reference, but access was blocked by site verification; no HockeyDB data is represented as imported or verified.

## Tournament model

This is a **retrospective what-if**, combining April NPI snapshots and postseason polls. It contains information that would not have been available before the 2026 tournament and therefore cannot be interpreted as a pre-tournament forecast or used to claim backtested prediction accuracy.

The men’s model uses a selected 16-team field and four regions, followed by two Frozen Four semifinals and a title game. The women’s model uses 11 teams: three first-round games, five byes, four quarterfinals, two semifinals and a title game. The first five user seeds receive women’s byes. The structures reference the [2026 men’s selection announcement](https://www.ncaa.com/news/icehockey-men/article/2026-03-22/2026-ncaa-division-i-mens-hockey-championship-selections) and [2026 women’s announcement](https://www.ncaa.com/news/icehockey-women/article/2026-03-08/2026-national-collegiate-womens-ice-hockey-championship-field-announced).

Your selected field is taken as given. Automatic qualifiers, conference constraints, host placement, remaining schedules and NCAA selection decisions are not simulated. Games are treated as neutral-ice events. The model does not include injuries or current player availability.

Default relative weights:

| Input | Weight | Treatment |
| --- | --- | --- |
| Overall record | 30 | Smoothed record percentage, converted to log odds |
| NPI | 35 | Difference in NPI divided by five |
| Poll consensus | 25 | Equal blend of USCHO and USA Hockey rank scores |
| Head-to-head | 10 | Only when games are available; six neutral prior games reduce small-sample influence |

For a team, record percentage is `(wins + 0.5 × ties + 2) / (games + 4)`. Consensus is `0.1 + 0.8 × (1 − average_poll_rank / (poll_size + 1))`, then converted to log odds. A team absent from a complete published top list receives rank `poll_size + 1`. NPI does not enter this consensus score a second time.

For a matchup, head-to-head percentage is `(A wins + 0.5 × ties + 3) / (loaded games + 6)`, converted to log odds. A zero-game or missing sample is omitted. Available weights are normalized to sum to one. Matchup probability is the logistic transform of the weighted strength difference divided by the randomness setting. Larger randomness moves probabilities toward 50/50. Inputs and correlations have not been fitted or calibrated; polls and NPI overlap in their evidence.

## Meaning of the percentages

The probability shown is **P(reaching the Frozen Four | selected field, seeding, historical inputs, model settings)**. It is not an official or calibrated probability, and it is not the probability of earning an NCAA tournament berth. Teams outside the field are not evaluated; they are not assigned an artificial 0% chance in a season-wide forecast.

Each Monte Carlo trial advances exactly four teams to the Frozen Four, two to the final and one to the championship. The corresponding displayed probability columns sum to 400%, 200% and 100% before rounding. The margin column is `1.96 × sqrt(p × (1−p) / trials)` and measures approximate Monte Carlo sampling error only. It is not an uncertainty interval for the model’s realism.

## Validation completed

- Both bracket sizes and women’s first-round byes.
- Exactly 15 men’s or 10 women’s elimination games, four semifinalists, two finalists, one champion.
- Deterministic results with identical seeds and inputs.
- Monte Carlo advancement totals; equal-strength teams converge to the expected probabilities, with byes accounted for.
- Symmetric matchup probabilities, randomness sensitivity and head-to-head direction.
- Rejection of duplicate fields and zero available weights.
- Valid source/team mappings, preserved ranking ties and immutable published tables.
- Browser checks for men’s and women’s results, seed swapping, saved personal rankings and editable head-to-head assumptions.

## Next step toward actual season forecasts

Connect approved complete game results and weekly poll archives with timestamped snapshots; separate features available on each prediction date from future information. Implement current NCAA qualification rules, conference tournaments, remaining schedules and actual site/seed placement. Fit and backtest a matchup model on earlier seasons, evaluate probability calibration on held-out seasons, and expose model uncertainty separately from Monte Carlo noise. Only then label season-wide outputs as estimates of actual tournament qualification and Frozen Four chances.