athletevalue

What an NCAA men's basketball player's wins are worth to his school, and what a published roster budget would pay him. A Python package and CLI that rates players from every lineup they played, converts impact into wins and into the revenue schools attribute to basketball in federal filings, and sets that against an allocation of reported roster budgets. Every estimate carries an interval and a stated basis.

What the price side is. No public dataset records what individual college basketball players are paid. The "roster market value" this package prints is an allocation: a published conference-tier average budget, spread across a roster by role and rating. It contains no information about any player's contract and is labelled a scenario for that reason. A model that learns pay from disclosed deals is included but ships dormant, because there are no public deals to train it on.

No pay data. Every output is a research estimate built from public game data and public filings. None of them states what a player is paid, and none should be quoted as a salary. See DISCLAIMER and PRIVACY.

Install

pip install athletevalue          # or: uv add athletevalue

Python 3.12 or later. The first valuation downloads about 500 MB of public data for 16 seasons and parses 15 years of EADA spreadsheets; allow several minutes and about 5 GB of memory for a season fit (7 GB with --cv). Later runs read the cache. athletevalue cache-path prints the cache location.

Quick start

athletevalue value "Player Name" --season 2026 --team "School"

The example below is a synthetic player on a synthetic season, produced by scripts/readme_example.py; the README does not name real athletes.

A. Guard · State U (Example Conf) · 2025-26 · starter

Athletic impact            +4.6 /100   80%: +0.9 to +8.3            ▲ estimated
  offense / defense      +2.6 / +2.0
Wins above replacement           0.6   80%: 0.2 to 1.2              ○ scenario
  replacement         bench median (-0.3/100); under nba convention 0.9 / pooled 0.6
Program value                   $57k   80%: $12k to $142k           ○ scenario
  win revenue, this season          $31k   80%: $8k to $55k             ○ scenario
  bid revenue, this season           $9k   80%: $1k to $30k             ○ scenario
  tournament units              $17k   80%: $3k to $57k             ○ scenario
  with next season              $74k   80%: $16k to $177k           ○ scenario
Roster market value           $1.47M   80%: $485k to $2.54M         ○ scenario
Surplus                      -$1.40M   80%: -$2.43M to -$470k       ○ scenario
Among paid teammates  both above median   (relative to team medians)

Drivers
+ on-court impact ranks 1 of 72 rated players
~ box-score prior +3.4 per 100; own possessions moved the rating +1.1
+ on the floor for 58% of team possessions
~ team went 10-4, net rating +12.1
+ program reports $12.00M basketball revenue (D1 median $7.00M)
+ Example Conf roster budget tier: power (published average, collectives included); the House cap for direct revenue sharing across all sports is $20.50M

Price basis: allocation. Status: ● reported ◆ derived ▲ estimated ○ scenario
As of 2026-09-17 · games through 2025-12-29 · model mbb-v0.5.0
No contract data is used. The market value spreads a published conference-tier budget across the roster by role and rating.

from athletevalue import api

valuation = api.value_player("Player Name", season=2026, team="School")
print(valuation.summary())
valuation.war.value, valuation.war.lower, valuation.war.upper

table = api.value_team("School", season=2026)  # polars DataFrame, one row per player

Other commands:

Seasons are keyed by ending year: 2026 is the 2025-26 season.

What each number means

Every estimate carries an 80% interval and a status:

• ● reported: stated by a primary source.
• ◆ derived: computed from reported data by a fixed procedure.
• ▲ estimated: depends on a statistical fit or a published third-party figure.
• ○ scenario: depends on an assumption a user should question, such as which replacement level to use or how a conference splits tournament money.

A result takes the status of its weakest input. Because the replacement level is a choice, wins and every dollar figure are scenarios; the summary prints wins under all three definitions.

Validation

Each season fit is checked against the SportsDataverse league-wide RAPM, which is fit on the same possessions and shrinks toward zero (a reimplementation check), and against Bart Torvik's team ratings, which are not (an external check).

athletevalue validate --season 2026 reproduces this table. Torvik's site refuses some cloud address ranges; when that happens the Torvik rows are skipped and reported.

The layers built on the ratings have no published reference, so athletevalue validate --season 2026 --extended checks them against the data's own outcomes:

Every number in these tables comes from docs/_static/data/snapshot.json, built by scripts/build_snapshot.py; a test fails if the README and the snapshot disagree.

What the box-score prior buys. Held-out squared error per 100 possessions, games held out, box totals and the team adjustment rebuilt from training games:

Most of the gain is the team adjustment restoring the scale of team strength; the box model itself adds about 0.1%. On 1,735 players who appear in both 2025 and 2026, the 2025 rating with the prior (trained on 2021–2024) correlates 0.52 with the 2026 no-prior rating, against 0.40 without it. The extended validation runs that check as a gate.

Data and licensing

Code is MIT. Curated data in src/athletevalue/data (the team crosswalk and the deal registry) is CC BY 4.0; see LICENSE-DATA. Data downloaded at runtime stays under its own terms and is never redistributed. A commercial user should do their own diligence on the SportsDataverse chain.

KenPom and the Knight-Newhouse College Athletics Database restrict redistribution. The package has a reader for a user's own KenPom export, for local comparison, and an import-linter contract that stops fitting code from importing it.

Limitations

• Program value measures an association. It does not estimate what one more win would cause. The win coefficient is identified from a school winning more or less than its own norm, and schools that spend more win more and earn more in the same years, so the coefficient can be biased upward. It counts only revenue schools attribute to men's basketball in EADA, which omits conference media money, donations and brand effects, a bias downward. The net direction is unknown. Sixty-one percent of school-seasons report revenue equal to expense and are excluded from the fit but kept in each school's revenue base, so for most programs the dollar figure is an elasticity from revenue-tracking schools applied to a budget number.

  

• The replacement level is a choice. The NBA convention is −2.0 per 100. In these data the pooled low-minute players rate about −12, and the players ranked 9th–12th on each team about −1.2. The default uses the bench median; the summary shows all three.

  

• Market value is an allocation for 2025-26 only. It spreads reported average roster budgets by role and rating and sums to the budget in every draw. Every conference outside the report's named top and middle tiers gets the bottom tier, including some programs whose real budgets are far higher.

• Ratings are noisy and their intervals condition on the prior. A starter's net rating has a posterior SD near 2.6 points per 100 with the prior and near 4 without. The prior mean is treated as known. On synthetic seasons with known ratings, the 80% intervals without a prior cover 93% of true values: conservative, not overconfident.

  

• Players are rated within a season. A person_id links the same player across seasons for validation; ratings do not yet borrow strength from earlier seasons.

• Results agree to about 1e-15 relative between runs, because BLAS is multithreaded. ATHLETEVALUE_DETERMINISTIC=1 makes repeats on one machine bit-identical. Upstream files are pinned by SHA-256, so a re-cut release is detected rather than silently used.

METHODOLOGY.md gives the equations, fitted values and caveats for each layer. The documentation site renders it with the full validation report, every assumption with its source, and the API reference.

Roadmap

• 0.6.0: allocation dispersion and per-school budgets; a finer status taxonomy; a sport seam, so a second sport is a package rather than a fork.
• Later: multi-season ratings with recency weights, in-season updates from CollegeBasketballData, then women's basketball.

Contributing

See CONTRIBUTING. The deal registry is not accepting rows until its correction and takedown process is in place; see PRIVACY.

Citation

If you use athletevalue in a publication, cite the software:

@software{hodge_athletevalue_2026,
  author  = {Hodge, John},
  title   = {athletevalue: open models of college basketball player impact,
             program value and roster-market pay},
  year    = {2026},
  version = {0.5.0},
  doi     = {10.5281/zenodo.22818303},
  url     = {https://github.com/jman4162/athletevalue},
  note    = {Python package, MIT license; data files CC BY 4.0}
}

10.5281/zenodo.22818303 resolves to whichever version is current. Each release also gets its own DOI, which Zenodo mints when the GitHub Release is published: 10.5281/zenodo.22818304 is 0.4.0. A release cannot name its own DOI, because the DOI does not exist until the release does, so cite the version-independent one above unless you need to pin an exact version.

CITATION.cff carries the same metadata in a machine-readable form. Estimates are research outputs; read DISCLAIMER before using them for decisions.