Fig. 3 — enlarged from sheet 01
Can you predict the NCAA tournament better than picking the higher seed every time? Two models trained on six tournaments estimate each matchup, then the whole 2026 bracket gets simulated 50,000 times to turn those estimates into title odds for all 64 teams.
I wanted to answer what sounds like a simple question: can you predict the NCAA tournament better than picking the higher seed every time? That sounds easy to beat and it is a brutally strong strategy. Decades of data say the 1-seed beats the 16-seed 99% of the time, the 5 beats the 12 about 65% of the time, and 8 versus 9 is a coin flip. Those numbers are already well calibrated. They are the market, and anything I built had to justify itself against a baseline refined by thousands of games.
Instead of seed number alone, the model sees how teams play: shooting efficiency, ball security, rebounding, free throw rate, tempo, strength of schedule, and recent momentum. Two teams with the same efficiency rating can play very different styles, and a team that lives on offensive rebounds plays a different game against a great defensive rebounding team than its headline number suggests.
There’s one more ingredient. I built a network of every team’s expected win probability against every other team and ran PageRank and betweenness centrality on it. A team favored against many teams who are themselves favored against many teams gets a boost. It measures how central you are to the hierarchy of the field, not how you compare head to head.
Two models make the predictions. A linear model looks at the differences between the two teams. A small neural network can pick up interactions the linear model can’t express, like a slow defensive team facing a fast offense in the second weekend. The two are blended by their Brier scores, so whichever has been more accurate gets more weight. Both get the seed-based historical win rate as an input, the way you’d hand a forecaster the historical average temperature before the satellite data. With only six years of training data the model shouldn’t waste its capacity rediscovering that 1-seeds usually win; it starts from that prior and learns when and how much to adjust. It also knows which round it’s predicting, because the 12 over 5 upset is a first round staple and the Elite Eight tends to go chalk.
Then the whole tournament, all 63 games, gets simulated 50,000 times. When a team shows 4% title odds, it won 2,000 of those tournaments, including the ones where it faced a 12-seed in the Sweet 16 instead of the expected 4. An upset dial pulls every probability toward or away from 50/50 and re-runs the simulations without retraining, so you can look at a wild year next to a chalk year.
Validation was leave-one-year-out: train on five tournaments, predict the sixth, rotate through each year as the holdout. The ensemble beat both individual models and the seed baseline on bracket scoring. The gains were modest, because the baseline is that strong, and they came mostly in the 5-12 and 6-11 games, where the seed history is closest to a coin flip.
The main bottleneck was data. Only 2025 had measured KenPom team stats; the other years use seed-level archetypes with synthetic noise, which washes out the within-seed variance that is the whole reason to build a model. More years of measured stats would be the single biggest improvement.