🐻⬇️🏀

Classic Elo fit on scored games with home-court handling. Good general win model baseline and already inherently time-aware because ratings update game by game.

Win/loss model of the form logit(P(team i beats team j)) = eta_i - eta_j + beta_home. This is the clean static paired-comparison baseline.

Same paired-comparison form as Bradley-Terry, but newer games receive more weight. This is the simplest bridge between static BT and fully dynamic team models.

Margin model of the form margin_ij = eta_i - eta_j + beta_home. It is the linear paired-comparison baseline for scoring margin targets.

Same linear margin model as Margin, but trained with exponentially decayed sample weights so recent games matter more.

Classic points-scored versus points-allowed baseline. Good sanity check for how much simple season scoring rates already explain wins.

Transforms raw scoring into points-per-possession style efficiency before building win-strength estimates. This is the pace-adjusted baseline before opponent correction.

Best current in-repo adjusted-efficiency baseline: adjusted offensive and defensive efficiency, opponent correction, and pythagorean-style conversion back to win strength.

Uses a log-style adjustment so blowouts do not dominate as hard as they do in raw margin style models. This is the closest current proxy for the log(margin) family discussed in planning.

Model of the form points_i = intercept + off_i + def_j + beta_home. This is the two-parameter team baseline without pace adjustment.

Same two-parameter points model as Points Off/Def, but with recency weighting. Useful for testing whether newer offensive and defensive form should move faster than season-long estimates.

Simple hybrid that averages log-odds from the strongest current team baselines. This is the clean 'blend the good baselines' model rather than a new estimator family.

Current strongest team hybrid in the strict bakeoff. It keeps the blend structure from Core Ensemble but swaps in the recency-weighted points off/def component.

AP Poll comparison baseline for D1 NCAA only. It is treated as an ordinal ranking signal, not a calibrated probability model. On validation pages it only appears in the ranked-games scope, where every model is re-evaluated on games involving at least one ranked team.

Intentional dumb baseline. Useful only to ensure richer models beat a naive home-court rule.

Training-only average margin baseline. Another cheap sanity check for whether a richer team model adds signal beyond raw per-game margin.

Glicko-style paired-comparison model with team uncertainty that expands during idle stretches and tightens with new games. This is the first explicit temporal paired-comparison model beyond static BT plus recency weighting.

Context-heavy second-stage win model that blends our strongest component ratings with rest and venue context. This remains an exploratory team baseline: useful in the bakeoff and model pages, but not yet clean enough for the main public team-output surfaces.

Compatibility alias for the historical player_model_outputs surface. The richer first-class rows are RAPM Overall and RAPM Off/Def.

Classic player RAPM with one overall impact coefficient per player. This is the clean scalar adjusted-impact baseline.

Two-parameter player RAPM variant with offense and defense estimated together. This is the first explicit expansion beyond a single player eta.

Compatibility alias for the historical player_model_outputs surface. The cleaner browseable scoring version is Plus/Minus Rate.

Simple player baseline: no teammate or opponent adjustment, just observed plus/minus scaled to possession volume.

Same raw per-100 plus/minus baseline as Plus/Minus Rate, but weighted toward recent games. Useful as a cheap temporal player baseline.

Single 2D vector per player. Five-player lineup vectors are summed, passed through a tanh nonlinearity, then interacted against the opponent lineup vector to predict segment margin. The displayed player rating is that player's neutral-context projection onto the learned scoring head, with the vector coordinates stored in metadata.

Sequential Elo-style update model on team points per 100 possessions. Positive values mean a player's minutes tend to lift their own team's scoring rate.

Defensive side of the points event model. Positive values mean a player's minutes tend to reduce the opponent's points per 100 possessions.

Neutral-context sum of Points For Elo and Points Prevention Elo. This is the cleanest interpretable 2D replacement for 'offense points created / defense points prevented' before moving to richer vector models.

Sequential event model on team rebounds per 100 possessions. Positive values mean a player's minutes tend to raise their own team's rebound rate.

Defensive side of the rebound event model. Positive values mean a player's minutes tend to reduce the opponent's rebounds per 100 possessions.

Sequential event model on team offensive rebounds per 100 possessions. Positive values mean a player's minutes tend to create more second-chance opportunities for their own team.

Opponent-suppression side of the offensive-rebound event model. Positive values mean a player's minutes tend to deny the opponent second-chance possessions.

Neutral-context sum of Offensive Rebounds For Elo and Offensive Rebound Prevention Elo.

Sequential event model on team defensive rebounds per 100 possessions. Positive values mean a player's minutes tend to end possessions by cleaning the defensive glass.

Counterpart to Defensive Rebounds For Elo. Positive values mean a player's minutes tend to reduce the opponent's defensive rebounds per 100 possessions.

Neutral-context sum of Defensive Rebounds For Elo and Defensive Rebound Prevention Elo.

Neutral-context sum of Rebounds For Elo and Rebound Prevention Elo. Good quick read on which players most tilt the glass overall.

Sequential event model on team assists per 100 possessions. Positive values mean a player's minutes tend to lift assisted shot creation for their team.

Defensive side of the assist event model. Positive values mean a player's minutes tend to reduce how cleanly the opponent generates assisted offense.

Neutral-context sum of Assists For Elo and Assist Prevention Elo. Useful for separating orchestrators from players who mainly score without driving team assist volume.

Sequential event model on made threes per 100 possessions. Positive values mean a player's minutes tend to raise their team's made-three volume.

Defensive side of the three-point event model. Positive values mean a player's minutes tend to reduce the opponent's made-three rate.

Neutral-context sum of Threes For Elo and Three Prevention Elo. Useful for spotting spacing gravity plus perimeter containment in one scalar.

Sequential event model on team stocks per 100 possessions, where stocks = steals + blocks. Positive values mean a player's minutes tend to create more disruptive defensive events for their team.

Opponent-suppression side of the stocks event model. Positive values mean a player's minutes tend to reduce the opponent's steals-plus-blocks rate.

Neutral-context sum of Stocks For Elo and Stocks Prevention Elo.

Sequential event model on negative turnovers per 100 possessions. Positive values mean a player's minutes tend to improve team ball security.

Opponent-suppression side of the turnover event model. Positive values mean a player's minutes tend to make the other team cough the ball up more often.

Neutral-context sum of Ball Security Elo and Turnover Force Elo. This is the clean turnover-margin read from the online event family.

Online segment-margin model with one scalar per player, updated stint by stint from lineup scoring outcomes. This is the simplest streaming analogue to batch RAPM.

Planned temporal generalization of the team grid. Lets strength move across days or weeks and supports uncertainty-aware backtesting in rolling windows.

Planned team model that builds team ratings from player-level estimates, projected usage, lineup mixes, and possibly injuries/availability.

Planned probabilistic analogue of the current team grid. Most useful once we want uncertainty-calibrated rolling bakeoffs and better budget-aware hyperparameter search.

Planned alternative to pure off/def RAPM: one shared overall quality parameter plus a second axis for how that value leans offense versus defense.