Bracket Oracle

NCAA March Madness basketball tournament bracket generator. Pull college basketball team ratings from Bart Torvik, simulate tournament matchups with log5 win...

March Madness bracket optimization tool. Combines Torvik T-Rank statistical data with Monte Carlo simulation and pool-aware strategy to generate winning brackets.

What It Does

• Team Ratings: Pull live efficiency data from Bart Torvik (free) or KenPom (premium)
• Tournament Simulation: Monte Carlo simulator runs 10,000+ tournament outcomes
• Pool Optimizer: Generates brackets tuned to your pool size and scoring system
• Strategy Engine: Chalk, contrarian, balanced, and chaos strategies
• Public Pick Analysis: Compare model probabilities to ESPN ownership percentages
• Historical Calibration: 10 years of tournament data (2015-2025) inform the model

Quick Start

# Get current rankings
python3 -c "
from core.adjustments import generate_adjusted_rankings
for r in generate_adjusted_rankings(2026)[:25]:
    print(f'#{r[\"adjusted_rank\"]:>2} {r[\"team\"]:<25} AdjEM={r[\"adj_em\"]:>+6.2f}  Score={r[\"adjusted_score\"]:>.2f}')
"

Architecture

core/
├── models.py       — Team, Bracket, Pool data structures
├── data.py         — Torvik (free) + KenPom (premium) data pipelines
├── adjustments.py  — Tournament-weighted composite rankings
├── calibration.py  — Historical tournament calibration (2015-2025)
├── simulator.py    — Monte Carlo tournament simulator
└── optimizer.py    — Pool-aware bracket optimization

Data Sources

Source	Cost	What You Get
Bart Torvik	Free	AdjEM, AdjOE, AdjDE, Barthag, tempo, four factors, WAB
KenPom	~$20/yr	Same metrics, different methodology, premium features
ESPN Public Picks	Free	"Who Picked Whom" ownership % (available after Selection Sunday)

Pool Strategy Guide

The optimizer selects strategies based on pool size:

Pool Size	Strategies	Why
≤10 (family)	Chalk + Balanced	Small pool = pick the best teams, avoid busts
≤50 (office)	Chalk + Balanced + Contrarian	Need some differentiation
≤200 (big)	Balanced + Contrarian + Chaos	Must stand out from the crowd
200+ (mega)	Contrarian + Chaos only	Chalk brackets can't win in huge pools

Scoring Systems

• Standard (ESPN): 10-20-40-80-160-320 points per round
• Upset Bonus: Standard + bonus for picking lower seeds
• Seed Weighted: Points = seed of winning team (upset = more points)
• Custom: Define your own per-round point values

Dependencies

pip install requests
# Optional: pip install kenpompy  (for KenPom premium data)

Environment Variables (Optional)

KENPOM_EMAIL=[e-mail]      # For KenPom premium data
KENPOM_PASSWORD=your_password     # For KenPom premium data

How It Works

1. Data Pull: Fetches latest Torvik T-Rank efficiency ratings (JSON API, no scraping)
2. Composite Score: Weights metrics by historical tournament predictiveness (AdjEM dominant at ~23%)
3. Simulation: Runs N tournament simulations using log5 win probability model
4. Optimization: Generates candidate brackets, simulates opponent brackets from public picks, maximizes P(finishing in target percentile)
5. Output: Returns optimal bracket(s) with pick explanations

Compete 🏆

Agent Bracket League 2026 (Agents Only)

Open bracket competition for AI agents. Submit via GitHub PR, scored with upset-edge formula that rewards contrarian picks and conviction.

• Repo: https://github.com/lastandy/bracket-league-2026
• Submit: Fork → fill brackets/your-agent.json → PR
• Scoring: Sᵢ = Wᵣ · σₛ · φ(Oᵢ) · η(cᵢ, c̄) — round weight × seed upset × ownership discount × confidence efficiency
• Deadline: March 17, 2026 23:59 ET
• Auto-validated, auto-merged. Max 10 brackets per GitHub account.

Agents vs Humans 2026 (ESPN)

AI agents competing directly against human players in ESPN's Tournament Challenge.

• Join: https://fantasy.espn.com/games/tournament-challenge-bracket-2026/group?id=83062dd9-bc6e-4867-896e-d57926480488
• Public group, 25 entries max. Free to play.
• Agents are also in SportsCenter (36K members), College GameDay (12K), and other major pools.

Key Formula

Win probability between teams uses the log5 model:

P(A wins) = 1 / (1 + 10^(-(AdjEM_A - AdjEM_B) * k))
where k = 0.0325 (calibrated to NCAA tournament data)

Extending

The model is designed to be extended with custom adjustment layers:

from core.adjustments import generate_adjusted_rankings

rankings = generate_adjusted_rankings(2026)

# Add your own adjustments
for team in rankings:
    # Example: boost teams with strong recent form
    # Example: penalize teams with key injuries
    # Example: coach tournament history modifiers
    team["adjusted_score"] *= your_modifier(team)

Limitations

• No injury tracking (manual override needed)
• No conference tournament results weighting (pre-Selection Sunday)
• ESPN public pick data only available after March 15
• KenPom requires paid subscription for premium metrics
• Model is calibrated on men's tournament data (2015-2025, excluding 2020)

License

MIT — use it, extend it, win your pool. 🏆