Dynamic Bradley-Terry
Time-evolving paired-comparison model with latent team strength drift.
| Status | implemented | Short Name | DynBT |
|---|---|---|---|
| Rating Style | elo | Where Shown | teams rankings, team pages, model performance |
Model Card
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.
Validation on this page uses stored season-level full-fit rows plus strict weekly rolling holdouts. Rolling ranks are based on average weekly log loss in the selected league/season. Current scope: All Games. All scored games in the selected league and season. AP Poll is excluded here.
Selected League / Season
Use the selected scope to inspect this model against peers and across rolling weekly holdouts.
| Full Fit AUC | - · 0 games |
|---|---|
| 7-Day Holdout Log Loss | 0.506 (#8) · 290 games |
| Rolling Avg Log Loss | 0.567 (#6) · 14 windows |
| Window Wins | 2 · 72.6% rolling accuracy |
Full-fit AUC is only available where the older stored-rating batch produced that model for the selected scope. Recency and ensemble families can still have valid 7-day and rolling holdout results even when the full-fit columns are blank.
Rolling Holdout Curves
Selected model plus comparison peers for NCAAWD2 2025-2026. The focus line is Dynamic Bradley-Terry. Click legend items to hide/show peers.
Weekly strict holdout log loss. Lower is better. Showing 6 models across 14 windows.
Recent Windows
| Holdout | Log Loss | Rank | Winner |
|---|---|---|---|
| Feb 4 - Feb 8 | 0.513 | #7/16 | Recency Ensemble (0.495) |
| Jan 28 - Feb 3 | 0.533 | #8/16 | Margin Recency (0.477) |
| Jan 21 - Jan 27 | 0.496 | #8/16 | Margin Recency (0.476) |
| Jan 14 - Jan 20 | 0.509 | #8/16 | Recency Ensemble (0.487) |
| Jan 7 - Jan 13 | 0.514 | #11/16 | Core Ensemble (0.462) |
| Dec 31 - Jan 6 | 0.574 | #1 | Dynamic Bradley-Terry |
| Dec 24 - Dec 30 | 0.444 | #12/16 | Log Adjusted (0.055) |
| Dec 17 - Dec 23 | 0.584 | #12/16 | Recency Ensemble (0.514) |
Selected Scope Peer Table
Comparison snapshot for NCAAWD2 2025-2026. Rows are ordered by rolling log-loss rank; full-fit AUC may be blank for models that do not have legacy full-season rows stored for this scope.
| Model | Full AUC | AUC Rank | 7d LogLoss | 7d Rank | Rolling LL | Rolling Rank | Windows |
|---|---|---|---|---|---|---|---|
| Recency Ensemble | - | - | 0.484 | #1 | 0.530 | #1 | 14 |
| Core Ensemble | - | - | 0.484 | #2 | 0.530 | #2 | 14 |
| Margin Recency | - | - | 0.487 | #4 | 0.532 | #3 | 14 |
| Margin | 0.849 | #4 | 0.492 | #5 | 0.538 | #4 | 14 |
| Avg Margin Baseline | 0.851 | #2 | 0.535 | #12 | 0.562 | #5 | 14 |
| Dynamic Bradley-Terry | - | - | 0.506 | #8 | 0.567 | #6 | 14 |
| Bradley-Terry | 0.860 | #1 | 0.518 | #10 | 0.573 | #7 | 14 |
| Bradley-Terry Recency | - | - | 0.527 | #11 | 0.581 | #8 | 14 |
League × Season Summary
Cross-scope validation history for Dynamic Bradley-Terry. Rolling ranks are by average weekly log loss.
| League | Season | Full AUC | AUC Rank | 7d LL | 7d Rank | Rolling LL | Rolling Rank | Windows | Wins |
|---|---|---|---|---|---|---|---|---|---|
| WNBA | 2026 | - | - | 0.582 | #12 | 0.622 | #3 | 3 | 1 |
| WNBA | 2025 | - | - | 0.549 | #6 | 0.628 | #5 | 21 | 1 |
| NCAAWD2 | 2025-2026 | - | - | 0.506 | #8 | 0.567 | #6 | 14 | 2 |
| NCAAW | 2025-2026 | - | - | 0.603 | #9 | 0.555 | #8 | 22 | 0 |
| NCAAMD2 | 2025-2026 | - | - | 0.564 | #9 | 0.595 | #10 | 13 | 0 |
| NCAAM | 2025-2026 | - | - | 0.688 | #14 | 0.595 | #10 | 22 | 0 |
| NBA | 2025-2026 | - | - | 0.764 | #11 | 0.635 | #11 | 33 | 1 |
| GLEAGUE | 2025-2026 | - | - | 0.860 | #14 | 0.691 | #13 | 22 | 3 |
Output Coverage
Stored site-output rows for this model.
| League | Season | Rows |
|---|---|---|
| WNBA | 2025 | 17 |
| NCAAWD3 | 2025-2026 | 585 |
| NCAAWD2 | 2025-2026 | 522 |
| NCAAW | 2025-2026 | 592 |
| NCAAMD3 | 2025-2026 | 646 |
| NCAAMD2 | 2025-2026 | 521 |
| NCAAM | 2025-2026 | 704 |
| NBA | 2025-2026 | 37 |
| GLEAGUE | 2025-2026 | 31 |