Scikit-learn models for NFL, NBA, PBA, tennis, and football. Trained on earlier results.
Loading your schedule...
Choose a matchup to predict the winner. Add ArenaPlus odds when you want to compare the price.
Python trains logistic regression and gradient boosting, chooses on earlier validation data, and checks predictions on a later holdout.
Track the decision, the price, and what happened next.
Recording a bet only saves it to this device. Settle after checking the sportsbook's result and retirement rules.
A chronological holdout, with every result kept on its side of time.
Separate NFL, ATP, and WTA classifiers learn from pre-match Elo, recent form, scoring margins, rest, and experience. Tennis adds surface-specific features. Scikit-learn logistic regression and gradient boosting compete on an earlier validation period. The selected model is calibrated on that period, then evaluated on a later holdout. Injury, lineup, weather, and live-match inputs are not included.
Evaluation uses chronological 60% training, 20% calibration, and 20% testing partitions. The production model then refits using 80% for training and 20% for calibration. Ratings only update after a result's availability time. Tennis tournament-date imports wait 21 days before updating, because a tournament start date is not a match completion timestamp.
A lower Brier score means better probability accuracy. A score below the binary 0.25 baseline (0.333 for football) does not establish a profitable edge against bookmaker prices. No historical ArenaPlus odds or return backtest is included.
Import results, keep their source, and check how recent they are.
Final scores and schedules from ESPN public scoreboards. Preseason, All-Star games, and live scores are excluded. Results use a conservative two-day delay.
The predictor accepts verified historical PBA results in normalized CSV format. No verified complete PBA feed is connected yet.
Download CSV header template ↗Schedules, team scores, neutral venues. Historical team codes are mapped to current franchises.
Source repository ↗Jeff Sackmann supplies the 2022-2024 archives. TennisCourtLog maintains 2025 onward from tennis-data.co.uk. Results are normalized to match winners; unknown surfaces, retirements, walkovers, and incomplete scores are excluded. Recent match dates use a two-day availability delay; older tournament dates use a 21-day delay.
Historical ATP mirror ↗ Historical WTA mirror ↗ Current ATP and WTA source ↗ Dataset license: CC BY-NC-SA 4.0 ↗Enter every side of the same winner market and the time you checked them. This workspace does not have a verified ArenaPlus odds feed.
Open ArenaPlus ↗