How It Works

Every prediction starts with KenPom efficiency metrics. If you follow college basketball analytics, you already know KenPom. It measures every Division I team across four dimensions:

• ▸Offensive efficiency — points scored per 100 possessions, adjusted for opponent strength
• ▸Defensive efficiency — points allowed per 100 possessions, adjusted for opponent strength
• ▸Tempo— pace of play, which affects matchup dynamics
• ▸Strength of schedule — how tough a team’s path has been

In total, the model builds 650+ matchup features per game, including decomposed components of each rating, conference adjustments, consistency measures, late-season trajectory data, and injury/availability signals.

• ▸Late-season trajectory — rank changes and defensive shifts in the final 30 days, sourced from Barttorvik and KenPom recency data
• ▸Injury monitoring — pre-tournament player availability, with injury interactions built into the Elite Eight model

BracketProof uses a multi-model ensemble — XGBoost for the champion classifier and most rounds, with Random Forest and LightGBM for key first-round matchup types where they outperform. Each model type has different inductive biases, and the ensemble exploits that diversity.

The key design choice: BracketProof trains separate models for each round. A 12-seed beating a 5-seed is a different problem than an Elite Eight game between two top-10 teams, and the model treats them differently.

The system tests 5 million feature and hyperparameter configurations across all rounds to find the optimal model settings — tree depth, learning rate, regularization, feature selection, and more. Over 65 million individual model evaluations across 13 years of leave-one-out validation.

The championship pick is worth 320 ESPN points— more than the entire first round combined. Getting it right changes everything. That is why BracketProof has a dedicated championship model that looks at what makes tournament winners different from regular-season contenders.

Champions tend to share specific traits:

• ▸Elite defensive efficiency — you don’t win six straight games without being able to get stops
• ▸Top-tier overall efficiency margin — the gap between offense and defense
• ▸High strength of schedule — a team that went through the Big 12 or SEC gauntlet is more proven than one that dominated a weak conference
• ▸Consistency — low variance in performance across the season

The champion model then cascades through the bracket, ensuring every game supports the champion’s path to the title.

The model doesn’t just pick the better team — it identifies specific matchup vulnerabilities that generic rankings miss.

• ▸Round-specific feature selection — each round uses different features, and first-round games go further with per-matchup models tuned to each seed pairing. A 5-vs-12 upset depends on different signals than a 1-vs-16 blowout. Elite Eight games weight turnover differential, offensive efficiency, perimeter defense, and injuries. What matters changes as the field shrinks.
• ▸Injury signal — the injury flag is a top feature in the Elite Eight model, where a single missing player can swing a tight matchup. The system also generates injury-cautious bracket variants that flip borderline calls on injured teams.
• ▸Late-season trajectory — a team’s final 30-day trend — rank changes, defensive shifts, and overall form — provides context the system surfaces alongside each pick. Peaking teams and declining teams look different by March, and the analysis flags these trends when they matter to the matchup.

College basketball changed dramatically starting in 2021. The NIL era and transfer portal reshaped how rosters are built. Teams now turn over 40–60% of their players year to year. Coaching tenure matters less when Kentucky’s roster is 80% new faces.

Models that rely on historical program identity or static recruiting rankings break down in this environment.BracketProof’s approach — training on measurable efficiency data that updates every season — adapts naturally. The model doesn’t care who recruited a player or where they transferred from. It measures what the team does on the court right now.

The results back this up. Since the NIL era began in 2021,BracketProof has called 6 of 6 champions and averaged 1,457 ESPN points per year — including a true forward test in 2026 where the model correctly predicted Michigan before the tournament started.

Every result on BracketProof uses leave-one-out cross-validation. Here is what that means in plain English:

The results you see are genuinely out-of-sample. The model could not have cheated.

The 2026 tournament goes one step further: it was a true forward test. The model was locked before Selection Sunday and the predictions were published before any games were played.