Track Record
Every prediction, every year. No cherry-picking, no revisions.
| Year | Our Pick | Actual | Score | +/- | 64 | 32 | 16 | E8 | FF | CH | |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 2013 | Louisville | Louisville | ✓ | 1610 | +700 | 27/32 | 13/16 | 5/8 | 3/4 | 2/2 | 1/1 |
| 2014 | Michigan St. | UConn | ✗ | 660 | -250 | 26/32 | 10/16 | 3/8 | 1/4 | 0/2 | 0/1 |
| 2015 | Kentucky | Duke | ✗ | 1240 | +330 | 28/32 | 12/16 | 6/8 | 4/4 | 1/2 | 0/1 |
| 2016 | Villanova | Villanova | ✓ | 1510 | +600 | 29/32 | 15/16 | 7/8 | 2/4 | 1/2 | 1/1 |
| 2017 | North Carolina | North Carolina | ✓ | 1630 | +720 | 29/32 | 15/16 | 6/8 | 2/4 | 2/2 | 1/1 |
| 2018 | Villanova | Villanova | ✓ | 1190 | +280 | 25/32 | 9/16 | 3/8 | 2/4 | 1/2 | 1/1 |
| 2019 | Virginia | Virginia | ✓ | 1440 | +530 | 22/32 | 13/16 | 8/8 | 2/4 | 1/2 | 1/1 |
| 2021 | Baylor | Baylor | ✓ | 1540 | +630 | 26/32 | 12/16 | 6/8 | 2/4 | 2/2 | 1/1 |
| 2022 | Kansas | Kansas | ✓ | 1250 | +340 | 31/32 | 11/16 | 4/8 | 1/4 | 1/2 | 1/1 |
| 2023 | UConn | UConn | ✓ | 1200 | +290 | 26/32 | 11/16 | 4/8 | 1/4 | 1/2 | 1/1 |
| 2024 | UConn | UConn | ✓ | 1620 | +710 | 30/32 | 14/16 | 4/8 | 3/4 | 2/2 | 1/1 |
| 2025 | Florida | Florida | ✓ | 1730 | +820 | 27/32 | 13/16 | 6/8 | 4/4 | 2/2 | 1/1 |
| 2026 | Michigan | Michigan | ✓ | 1400 | +490 | 32/32 | 12/16 | 5/8 | 2/4 | 1/2 | 1/1 |
| TOTAL | 11/13 | 1386 avg | 358/416 86.1% | 160/208 76.9% | 67/104 64.4% | 29/52 55.8% | 17/26 65.4% | 11/13 84.6% |
Methodology
Every prediction above was produced by a model that did not see the season it was predicting. We use leave-one-out cross-validation: to generate the 2013 pick we train on every other year, to generate the 2026 pick we train on every year except 2026, and so on. No result was tuned after the tournament was played.
The engine is a multi-model ensemble — XGBoost for the champion classifier and most rounds, Random Forest and LightGBM for key first-round matchup types — with matchup-specific routing for historically unstable bracket spots and a market-consensus cross-check against closing lines. Core features are drawn from KenPom efficiency metrics, which is why the backtest starts in 2013 (KenPom ratings only go back to 2010, and the model needs multiple training years before its first prediction).
The 2020 tournament is excluded because it was cancelled due to COVID-19. All other years in the KenPom era are included, with no cherry-picking.
Click any year to see the full bracket prediction