About March Madness Arena

March Madness Arena simulates the entire NCAA Tournament bracket using an AI model that analyzes each matchup in real-time. The simulation streams results as each game is decided, allowing you to watch the bracket unfold game by game.

The simulation follows the actual tournament structure: First Four play-in games, then the Round of 64, Round of 32, Sweet 16, Elite Eight, Final Four, and Championship.

Each game is simulated using Google's Gemini 3 Flash model with structured output. The AI acts as an "elite March Madness analyst" filling out a bracket to win a pool, not just picking favorites but weighing matchup-specific factors like defensive efficiency, tempo control, tournament DNA, and luck regression.

For every matchup, the model returns a structured JSON response with the predicted winner and reasoning, ensuring consistent and parse-able results.

The model uses a temperature of 0.7 to introduce controlled variance: running the same bracket multiple times will produce different outcomes while keeping the reasoning analytically grounded. The real variety comes from the prompt design: qualitative edge framing and tournament-specific analysis give the AI genuine reasons to pick upsets, rather than relying on raw randomness.

The prompting strategy is designed to produce realistic tournament outcomes with appropriate upset rates. Each game prompt includes:

• Team profiles: name, seed, conference, program tier (blueblood, power conference, mid-major), and Cinderella status
• Statistical analysis: KenPom rankings, adjusted efficiency margins, offensive/defensive ratings, tempo, and strength of schedule, presented with qualitative edge framing rather than explicit percentages, so the AI analyzes the matchup instead of anchoring on a number
• Tournament-specific KenPom factors: the AI is told which KenPom qualities matter most in March: elite defense (AdjD top 25) that grinds out close games, tempo control that neutralizes more talented opponents, style clashes (elite offense vs elite defense treated as coin flips), and luck regression for overperforming teams
• Historical seed matchup data: upset rates from 1985-2025 (e.g., 12-seeds beat 5-seeds 36% of the time)
• Tournament context: running totals of upsets vs chalk picks, with calibration nudges when the sim is running too chalky or too chaotic compared to historical norms
• Upset indicators: KenPom rank vs seed divergence, luck regression signals, strength of schedule gaps, unlucky teams that are better than their record, and elite defensive profiles
• Site & travel: each game includes the 2026 neutral venue (First Four in Dayton; R64/R32 pod sites; regionals; Final Four in Indianapolis). The model nudges tight matchups when one team’s footprint is closer to the arena or has a stronger traveling fan base, without overriding KenPom.

The ensemble win probability combines three different models to provide a baseline prediction:

The KenPom logistic model uses adjusted efficiency margin differences with the formula: 1 / (1 + 10^(-marginDiff/11))

Rather than showing the AI an explicit percentage (which would anchor it into always picking the favorite), the ensemble probability is translated into a qualitative edge description (toss-up, slight edge, favored, or clear favorite) so the AI genuinely evaluates matchup-specific factors instead of deferring to a number.

The system is calibrated to produce historically accurate upset rates. The AI is given explicit guidance on expected upset frequencies:

• Round of 64: ~10-12 upsets expected (lower seed wins)
• Round of 32: ~4-6 upsets expected
• 5 vs 12 matchups: 12-seeds win 36% of the time
• 8 vs 9 matchups: Essentially a coin flip (47% vs 53%)
• 1 vs 16 matchups: Only 2 upsets in tournament history (1%)

The tournament context tracker monitors actual vs expected upsets and provides dynamic calibration:

• Too chalky: If upsets are well below expected, the AI is told to pick the underdog when any legitimate indicators exist
• Too chaotic: If upsets significantly exceed expectations, the AI leans toward higher seeds
• On track: No calibration. The AI picks based purely on the matchup data and its own analysis

Combined with qualitative rather than numeric probability framing, this produces brackets with genuine variety: different champions, different Cinderellas, and different upset patterns across simulations.