Predicting
March Madness 2026

Seeds are hard to beat but not impossible

It turns out the simplest March Madness strategy of just picking the higher seed is very good. A 1 beats a 16. A 2 beats a 15.

Picking the higher seed gets 70% of games right. Our model of 15 features (from 110+ candidates) over 23 seasons of training data only improves accuracy by 2.3 percentage points.

15 features from 110+ candidates

We started with over 110 candidate features to predict win probability between two teams, and ended up keeping only 16.

The most important feature by far is AdjEM (Adjusted Efficiency Margin): the difference between a team's points scored and points allowed per 100 possessions, adjusted for opponent strength. A team that beats weak opponents by 20 looks less impressive than one that beats strong opponents by 10. AdjEM accounts for more importance than all other features combined.

Try it yourself:

Select two teams to see how the model predicts win probability.

Over 90 features were evaluated and rejected. Raw counting stats (points, rebounds, steals, blocks) were too noisy without opponent adjustment. Ranking systems (KenPom, Sagarin, Massey, and 6 others) were highly correlated with each other and with AdjEM. Coach experience (tournament wins, appearances, Final Four trips) was statistically redundant with program history and seed. Clutch performance (close-game win rate, overtime record) suffered from small sample sizes. Shooting variance and conference tournament results also added no value in cross-validation.

Games where the model predicts an underdog win

• 10-Santa Clara over 7-Kentucky (model: 70.5%). Santa Clara climbed 70 spots in the rankings this season while Kentucky dropped 24. The Broncos' efficiency margin improved by +13.3 points; Kentucky's fell by −7.5. Their recent win rate (.700 vs .400) confirms the trend. By March the gap in actual quality is far smaller than the seed gap suggests.
• 9-Iowa over 8-Clemson (model: 70.7%). Iowa is simply the better team by AdjEM: 28.1 vs 24.5. They're also healthier (injury rank 23 vs 37) and improved +21 ranking spots while Clemson slipped −4. The 8/9 seed line is where committee judgment is weakest.
• 12-Northern Iowa over 5-St John's (model: 50.3%). St John's has the higher AdjEM (30.9 vs 13.1), but their worst loss was a 32-point blowout compared to Northern Iowa's −15. Northern Iowa is also winning recent games by a wider margin (10.8 vs 9.8) and opponents rely more on three-point shooting (0.311 vs 0.270) — a volatile style in March.
• 12-Akron over 5-Texas Tech (model: 50.1%). The classic 12-over-5. Akron won their last 10 games by an average of 13.8 points while Texas Tech went 6–4 with a +5.2 margin and a fading win trend (−0.10 vs +0.30). Historically, 12-seeds beat 5-seeds 36% of the time. This one is a coin flip.

Probabilities compounding

We use a Monte Carlo simulation to simulate the entire bracket 50,000 times. Each simulation draws random outcomes weighted by our model's probabilities. In a bracket, Team A's path to the Final Four depends on who else advances. Simulation captures this path dependency that a pairwise model alone cannot.

Who wins March Madness 2026?

Advancement probabilities from 50,000 simulated tournaments. The green and red deltas show where our model diverges from the seed-only baseline: the places where all those extra features actually shift the prediction.

Season-by-season accuracy:

The model works poorly for upset-heavy years like 2011 and 2014.