2015-2016 NCAAMD3 Model Performance
Season
Every ranked model predicted the same 2779 games (2015-12-06 to 2016-03-19, 15 weekly windows) out of sample: each week's games were predicted by models fit only on the season's earlier games. 30 of 2809 games are left out because a model had no rating for a team yet.
Model Performance Leaderboard
Ranked by full-season AUC. The final week is the season's last 7 played days (3 games, 2016-03-13 to 2016-03-19), predicted from every earlier game: a small sample that shows the latest form.
| # | Model | AUC | Acc | Brier | LogLoss | AUC final wk | Acc final wk |
|---|---|---|---|---|---|---|---|
| 1 | Margin Margin Linear team-strength model fit on point differential instead of binary wins. More → | 0.773 | 69.6% | 0.195 | 0.574 | - | 66.7% |
| 2 | Margin Recency Margin Recency Margin regression with exponential recency weights on newer games. More → | 0.772 | 69.6% | 0.195 | 0.574 | - | 66.7% |
| 3 | Points Off/Def Recency Points Off/Def Recency Off/def points regression with exponential recency weights. More → | 0.770 | 69.3% | 0.197 | 0.577 | - | 66.7% |
| 4 | Points Off/Def Points Off/Def Raw points regression with separate offensive and defensive team parameters. More → | 0.769 | 69.2% | 0.197 | 0.577 | - | 66.7% |
| 5 | Core Ensemble Core Ensemble Equal-logit blend of Elo, recency BT, recency margin, log-adjusted pyth, and points off/def. More → | 0.766 | 69.4% | 0.197 | 0.576 | - | 33.3% |
| 6 | Recency Ensemble Recency Ensemble Equal-logit blend of Elo, recency BT, recency margin, log-adjusted pyth, and recency points off/def. More → | 0.766 | 69.5% | 0.197 | 0.576 | - | 33.3% |
| 7 | Adjusted Context Blend Adjusted Context Blend Chronologically trained component blend with rest and signed venue context. More → | 0.758 | 69.0% | 0.215 | 0.646 | - | 66.7% |
| 8 | Adjusted Efficiency Adjusted Efficiency Opponent-adjusted efficiency model with separate offensive and defensive components. More → | 0.757 | 68.4% | 0.201 | 0.584 | - | 33.3% |
| 9 | Log Adjusted Log Adjusted Log-scale adjusted efficiency model that downweights blowout leverage. More → | 0.756 | 68.4% | 0.201 | 0.584 | - | 33.3% |
| 10 | Bradley-Terry Bradley-Terry Static logistic paired-comparison model with one team strength parameter. More → | 0.741 | 66.9% | 0.205 | 0.595 | - | 33.3% |
| 11 | Bradley-Terry Recency Bradley-Terry Recency Static Bradley-Terry with exponential recency weights on newer games. More → | 0.740 | 67.0% | 0.206 | 0.595 | - | 33.3% |
| 12 | Pythagorean Pythagorean Pythagorean win expectation from raw points scored and allowed. More → | 0.739 | 67.7% | 0.210 | 0.609 | - | 33.3% |
| 13 | Avg Margin Baseline Avg Margin Baseline Predict from simple average scoring margin in the training window. More → | 0.739 | 67.6% | 0.207 | 0.599 | - | 33.3% |
| 14 | Dynamic Bradley-Terry Dynamic Bradley-Terry Time-evolving paired-comparison model with latent team strength drift. More → | 0.737 | 66.7% | 0.208 | 0.601 | - | 66.7% |
| 15 | Elo Elo Streaming paired-comparison rating with recency baked into sequential updates. More → | 0.701 | 63.9% | 0.225 | 0.641 | - | 66.7% |
| 16 | Home Team Baseline Home Team Baseline Always favor the home team with a fixed prior. More → | 0.557 | 55.7% | 0.249 | 0.691 | - | 66.7% |
Methodology
Out-of-sample windows
- The season is split into 7-day windows ending on its last played day
- Each window is predicted by models refit on all earlier games of the season
- The first window starts after 21 days of games
- Every model is scored on the same games; a model that cannot predict most of them is listed as not compared
Elo / Bradley-Terry
- Elo: the published league Elo (preseason priors and carryover where configured), frozen at the window start
- Bradley-Terry: static logistic fit on the earlier games
- Recency variants weight games with a 120-day half-life in NCAA leagues (keeping the non-conference games that link conferences) and a 30-day half-life in pro leagues
Pythagorean and margin models
- Raw, adjusted and log-adjusted Pythagorean from scores; efficiency from box-score possessions
- Margin: ridge regression on point margin with a learned home advantage
- Ensembles average the log-odds of five families
Baselines
- Home Team: the home team wins 60% (50% at neutral sites)
- Avg Margin: the higher average margin so far wins
- Models should beat these to add value