Most of this describes the NBA model, the one I've developed furthest.
The other sports use the same approach at an earlier stage.
Game-outcome model
A stacked blend of gradient-boosted trees and logistic regression (~65/35), with the mix chosen to minimize Brier score (how far predicted probabilities sit from what actually happened, lower is better). It learns from the modern era, 2013-14 onward, where the pace and three-point era resemble today's game. Training uses expanding-window cross-validation, meaning it is only ever tested on seasons later than the ones it learned from. Test AUC ≈ 0.755, which is a measure of how well it separates winners from losers, where 0.5 is a coin flip. Features are auto-pruned from 563 engineered columns down to the 140 that actually carry signal.
Feature engineering
Every sport starts with Elo ratings, a running team-strength number that rises and falls with results, in a standard form and a faster-moving one that weights blowouts and resets toward the mean each season. On top of that: recent-form trends, opponent-adjusted efficiency, and rest and schedule context. The NBA model adds basketball-specific inputs like RAPM (a player's net impact per 100 possessions, separated from his teammates and opponents with ridge regression), WOWY (how a team does with a player on the floor versus off it), and Dean Oliver's Four Factors (shooting efficiency, turnovers, offensive rebounding, and free-throw rate). Every rolling stat is lagged by one game so the model never sees the game it is trying to predict.
Probability calibration
Raw model output is calibrated, meaning it is adjusted so that a stated 65% actually wins about 65% of the time. It picks the best of three standard methods (Platt, isotonic, or temperature scaling) on a season held out from training, judged by the same Brier score.
Margin, totals, uncertainty
A separate ridge regression predicts the final point margin. Its error has a spread of about 12.5 points, and that spread is what turns a predicted margin into a probability of winning by a given amount: the wider it is, the less certain the same predicted margin becomes. Player stat projections come with conformal prediction intervals, a range the real number lands inside roughly 90% of the time, rather than a single guess.
The other sports
NFL blends a win-probability model with a margin-based one; soccer is a three-way model that accounts for draws; NHL uses an Elo-based stack with a goalie adjustment. Market lines from Pinnacle feed the models as an input signal, devigged, which strips out the sportsbook's built-in cut to leave a fair probability the models can learn from.
