Prop Commander
Evaluation complete
A WNBA player-prop research system and the evaluation harness that proved its own model has no edge over the market. Graded on 1,010 props across 13 game dates, then retired, with the result and the dataset published.
View the source on GitHub
Benchmark chartDevelopment details
- Source code
- Public
Built with
Screenshots
The market coefficient is positive on both samples. The model’s 95% interval includes zero on both, so it adds nothing the price did not already hold.
Stated probability against what happened, on the 513 frozen decisions. Neither curve tracks the diagonal at this sample size, and the market’s Brier score beats the model’s.
What it does
Prop Commander projects the outcome of WNBA player props, such as whether a player goes over 14.5 points, and compares each projection with the sportsbook’s no-vig price. A FastAPI and SQLite backend runs a daily pipeline over The Odds API and ESPN’s public data, with rate limiting, failure handling, and injury-report parsing. A React and TypeScript front end shows the board, the watchlist, and graded results.
The question it existed to answer
Does the model know anything the market does not? Every prediction is frozen at capture time with its probability, the market’s probability, and a full context snapshot, then graded automatically after the game. Features are computed as of a cutoff before tip, so no sample can see the result it is predicting.
The answer was no
Across 1,010 graded props on 13 game dates, a logistic regression of the outcome on both the market price and the model’s probability gives the model a coefficient indistinguishable from zero, and adding the model to the market makes out-of-sample predictions slightly worse. Recalibration, recentring the projected mean, a better minutes model, subgroup niches, line movement, and cross-book dispersion were each tested as ways to close the gap, and none did. The market’s own Brier score is 0.2464 against 0.2500 for a coin flip, so the total predictable signal in these markets is about 1.6 percent, and the market already holds it.
What survived
The projection model was retired in August 2026 and paid odds capture stopped. The evaluation harness, the free data pipeline, and the graded dataset stay in the repository so the negative result can be reproduced. The audit also fixed defects that had looked healthy from the outside: duplicated player identities, grading that had silently stopped for nine days, and an unattended capture loop that paid for the same game seven times because SQLite connections were never closed.