Skip to content

About

A composite analytical pipeline that predicts NCAA Tournament outcomes

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

March Madness 2026: Live Bracket Analysis & Prediction Engine

A live tournament prediction pipeline that generates a full 63-game bracket, tracks accuracy in real-time, and refits models as results come in. Built on KenPom efficiency metrics, NCAA NET rankings, committee evaluation metrics (SOR, WAB, KPI, BPI), quantitative models (GARCH volatility, HMM regime detection, Kalman momentum filtering, historical seed priors), Monte Carlo simulation, and EV-optimized bracket construction with contrarian leverage overlay.

Tournament Accuracy Tracker

Last updated: 2026-03-31T23:37:06Z

Metric Value
Games tracked 30
Correct picks 24/30
Accuracy 80.0%
Upsets caught 2/8

Accuracy Dashboard

Game-by-Game Results

Game Matchup Score Winner Our Pick Confidence Result
E1 (1) Duke vs (16) Siena 71-65 Duke Duke 98.4% Correct
E8 (2) UConn vs (15) Furman 82-71 UConn UConn 95.9% Correct
E6 (3) Michigan State vs (14) North Dakota State 92-67 Michigan State Michigan State 91.9% Correct
E4 (4) Kansas vs (13) Cal Baptist 68-60 Kansas Kansas 87.2% Correct
E3 (5) St. John's vs (12) Northern Iowa 79-53 St. John's St. John's 81.1% Correct
E5 (6) Louisville vs (11) South Florida 83-79 Louisville Louisville 73.0% Correct
E7 (7) UCLA vs (10) UCF 75-71 UCLA UCLA 67.0% Correct
E2 (8) Ohio State vs (9) TCU 64-66 TCU Ohio State 61.7% Wrong
M1 (1) Michigan vs (16) Howard 101-80 Michigan Michigan 92.9% Correct
M8 (2) Iowa State vs (15) Tennessee State 108-74 Iowa State Iowa State 97.3% Correct
M6 (3) Virginia vs (14) Wright State 82-73 Virginia Virginia 92.9% Correct
M4 (4) Alabama vs (13) Hofstra 90-70 Alabama Alabama 84.7% Correct
M3 (12) Akron vs (5) Texas Tech 3-6 Texas Tech Texas Tech 78.3% Correct
M2 (8) Georgia vs (9) Saint Louis 77-102 Saint Louis Georgia 56.4% Wrong
M7 (10) Santa Clara vs (7) Kentucky 18-23 Kentucky Kentucky 55.0% Correct
S8 (2) Houston vs (15) Idaho 78-47 Houston Houston 96.8% Correct
S6 (3) Illinois vs (14) Penn 105-70 Illinois Illinois 96.1% Correct
S4 (4) Nebraska vs (13) Troy 76-47 Nebraska Nebraska 93.1% Correct
S3 (5) Vanderbilt vs (12) McNeese 78-68 Vanderbilt Vanderbilt 82.8% Correct
S7 (7) Saint Mary's vs (10) Texas A&M 50-63 Texas A&M Saint Mary's 60.7% Wrong
S5 (6) North Carolina vs (11) VCU 78-82 VCU North Carolina 59.7% Wrong
S2 (8) Clemson vs (9) Iowa 61-67 Iowa Iowa 57.9% Correct
W1 (1) Arizona vs (16) Long Island University 92-58 Arizona Arizona 98.6% Correct
W8 (2) Purdue vs (15) Queens (NC) 104-71 Purdue Purdue 96.9% Correct
W6 (3) Gonzaga vs (14) Kennesaw State 73-64 Gonzaga Gonzaga 94.8% Correct
W4 (4) Arkansas vs (13) Hawaii 97-78 Arkansas Arkansas 89.3% Correct
W3 (5) Wisconsin vs (12) High Point 82-83 High Point Wisconsin 82.9% Wrong
W7 (7) Miami (FL) vs (10) Missouri 80-66 Miami (FL) Miami (FL) 64.0% Correct
W5 (6) BYU vs (11) Texas 71-79 Texas BYU 50.9% Wrong
W2 (8) Villanova vs (9) Utah State 76-86 Utah State Utah State 51.3% Correct

Record: 24/30 (80.0%) | Upsets caught: 2/8 | All misses were upsets

How It Works

  1. Composite Scoring -- Each team gets a weighted score from up to 11 normalized (0-100) metrics
  2. Injury Adjustment -- Real-time injury data adjusts team NetRtg values using player impact penalties
  3. Quantitative Enhancements -- GARCH volatility replaces fixed win-probability divisor, HMM detects team regime states, Kalman filter tracks momentum, historical seed priors provide Bayesian blending
  4. Monte Carlo Simulation -- 10,000 full-bracket simulations (63 games each) with both baseline and enhanced models
  5. EV-Optimized Bracket -- Per-slot expected value optimization with contrarian leverage adjustment

Data Sources

Data File Source
KenPom efficiency ratings kenpom.csv kenpom.com
NCAA NET rankings + quad records ncaa-net-rankings.csv ncaa.com
Team sheet metrics (SOR, WAB, KPI, BPI) scraped_data/tournament_teams.csv warrennolan.com
Game-by-game results scraped_data/tournament_games.csv warrennolan.com
Tournament bracket bracket.csv Manual entry
Injury reports scraped_data/injuries.csv boydsbets.com
Player stats (MPG, PPG) scraped_data/player_stats.csv espn.com
Injury adjustments scraped_data/injury_adjustments.csv Computed from injuries + player stats + KenPom
Historical seed win rates scraped_data/historical_seed_rates.csv Wikipedia (1985+)

Quickstart

pip install -r requirements.txt
jupyter notebook final_four_analysis.ipynb

Refreshing Data

Warren Nolan team sheets:

python scripts/scrape_net_teamsheets.py
python scripts/filter_tournament_teams.py

Injury data (full pipeline):

# Step 1: Scrape injury reports (boydsbets.com -- not in sandbox allowlist)
python scripts/scrape_injuries.py --refresh

# Step 2: Scrape player stats from ESPN + compute injury adjustments
# Takes ~1 min (68 teams, ~0.7s courtesy delay between requests)
python scripts/scrape_player_stats.py --refresh

# Step 3: Re-run the notebook with injury-adjusted data
jupyter nbconvert --execute final_four_analysis.ipynb --to notebook

Both scrapers cache results for 6 hours. Use --refresh to force re-scrape.

Historical seed data:

python scripts/scrape_historical_brackets.py

Falls back to hardcoded rates if Wikipedia is unavailable.

Without injury/quant data: The notebook gracefully degrades -- it runs the original 9-feature model with unadjusted NetRtg values. No scrapers or quant dependencies required.

Composite Model Weights

With all enhancements, the full scheme uses 11 features. The model degrades gracefully when components are unavailable.

Metric Enhanced Baseline Source
KenPom Net Rating 0.15 0.175 kenpom.csv
KenPom Offensive Rating 0.10 0.10 kenpom.csv
KenPom Defensive Rating (inv) 0.10 0.10 kenpom.csv
NCAA NET Rank (inv) 0.10 0.125 ncaa-net-rankings.csv
Strength of Record (inv) 0.10 0.10 tournament_teams.csv
Wins Above Bubble 0.10 0.10 tournament_teams.csv
NET Strength of Schedule (inv) 0.05 0.05 tournament_teams.csv
Q1 Win % 0.10 0.10 ncaa-net-rankings.csv
Q1+Q2 Win % 0.10 0.10 ncaa-net-rankings.csv
Injury Health 0.05 0.05 injury_adjustments.csv
Kalman Momentum 0.05 -- quant_models.py

Additional fallback weight schemes activate when team sheet or quad data is unavailable.

Monte Carlo Simulation

Baseline model uses a logistic function on injury-adjusted net efficiency differentials with a fixed divisor:

P(A wins) = 1 / (1 + 10^(-(NetRtg_A - NetRtg_B) / 22))

Enhanced model replaces the fixed divisor with per-matchup GARCH volatility, applies HMM regime adjustments, and blends with historical seed priors:

combined_vol = sqrt(GARCH_vol_A^2 + GARCH_vol_B^2)    # replaces fixed 22
P_model = 1 / (1 + 10^(-margin / combined_vol))        # volatility-adjusted
P_final = bayesian_blend(P_model, historical_prior)     # seed-aware blending

Each model runs 10,000 full-bracket simulations (63 games: R64 through Championship). The notebook outputs a side-by-side comparison.

Results

Top 20 by Composite Score

# Team Region Seed Composite NetRtg NET# SOR WAB
1 Arizona West 1 86.6 37.62 3 1 3
2 Duke East 1 85.2 38.90 1 3 2
3 Michigan Midwest 1 84.7 37.58 2 2 1
4 Florida South 1 79.4 33.78 4 5 6
5 Houston South 2 78.7 33.39 5 6 5
6 Iowa State Midwest 2 76.3 32.38 6 14 9
7 Purdue West 2 75.9 31.19 9 7 4
8 UConn East 2 75.7 27.85 10 4 7
9 Illinois South 3 74.5 32.09 8 17 18
10 Gonzaga West 3 74.2 28.10 7 11 17
11 Virginia Midwest 3 73.5 26.71 12 9 8
12 Michigan State East 3 73.2 28.30 11 13 12
13 Vanderbilt South 5 72.9 27.50 13 12 10
14 Nebraska South 4 72.0 26.15 14 10 13
15 St. John's East 5 71.3 25.89 16 16 14
16 Arkansas West 4 70.9 26.04 15 8 11
17 Alabama Midwest 4 70.4 25.70 18 15 15
18 Kansas East 4 69.2 24.41 21 18 16
19 Louisville East 6 68.4 25.42 17 26 24
20 Tennessee Midwest 6 67.8 26.02 20 22 23

Baseline Monte Carlo (10,000 sims per region)

East South West Midwest
(1) Duke 51.8% (1) Florida 30.5% (1) Arizona 49.7% (1) Michigan 47.5%
(2) UConn 13.7% (2) Houston 25.9% (2) Purdue 22.8% (2) Iowa State 24.4%
(3) Michigan St 12.1% (3) Illinois 23.1% (3) Gonzaga 10.2% (3) Virginia 8.4%
(5) St. John's 6.4% (5) Vanderbilt 8.1% (4) Arkansas 7.5% (4) Alabama 6.6%
(4) Kansas 5.5% (4) Nebraska 6.9% (5) Wisconsin 3.7% (6) Tennessee 4.8%

Enhanced vs Baseline Final Four Probabilities (Top 16)

The enhanced model (GARCH + HMM + Kalman + historical priors) redistributes probability mass away from top seeds and toward mid-seeds:

Team Baseline FF% Enhanced FF% Diff
Duke 51.8% 31.1% -20.7%
Arizona 49.7% 30.4% -19.3%
Michigan 47.5% 29.0% -18.5%
Florida 30.5% 28.7% -1.8%
Iowa State 24.5% 19.0% -5.5%
Houston 25.9% 18.9% -7.0%
Purdue 22.8% 18.7% -4.1%
UConn 13.7% 17.6% +3.9%
Illinois 23.1% 13.5% -9.6%
Gonzaga 10.2% 13.1% +2.9%
Virginia 8.4% 13.1% +4.8%
Michigan State 12.1% 12.6% +0.5%
Arkansas 7.5% 9.9% +2.4%
Alabama 6.6% 9.7% +3.1%
Nebraska 6.9% 9.6% +2.7%
Kansas 5.5% 9.3% +3.8%

EV-Optimized Bracket

ESPN fantasy scoring: R64=10, R32=20, S16=40, E8=80, FF=160, Championship=320.

Metric Value
Total EV (pure) 722.8 pts
Total EV (leverage-adjusted) 826.2 pts
Predicted Champion Duke

Final Four:

Pick Region Seed Composite
Arizona West 1 86.6
Duke East 1 85.2
Illinois South 3 74.5
Michigan Midwest 1 84.7

Leverage Sensitivity Analysis

Leverage Weight Total EV Leverage EV Champion Upsets
0.0 722.8 722.8 Duke 3
0.5 722.8 774.0 Duke 3
1.0 722.8 826.2 Duke 6
1.5 722.8 879.2 Duke 7
2.0 722.8 932.8 Duke 8

Visualizations

Offensive vs Defensive Efficiency

Efficiency Scatter

Top 25 by Composite Score

Composite Bars

Regional Contender Comparison

Radar Charts

Quantitative Model Analysis

Quant Analysis

Injury Impact Dashboard

Injury Dashboard

EV-Optimized Bracket Analysis

EV Bracket Analysis

Quantitative Models

Model Class Purpose
Hierarchical GARCH(1,1) HierarchicalGARCH Per-team performance volatility from game-by-game margins; replaces fixed win-probability divisor
Gaussian HMM TeamHMM Detects hot/cold team regimes from quad-adjusted margins; BIC selects 2-4 states
Kalman Filter KalmanMomentum Tracks late-season momentum as a random walk on margin residuals (Q=2, R=10)
Historical Prior HistoricalPrior Bayesian blending with 1985+ seed matchup win rates (e.g., 1v16: 99.4%, 8v9: 51.5%)
Enhanced Simulator QuantEnhancedSimulator Full 63-game MC combining all models with graceful degradation
EV Optimizer EVOptimizedSimulator Per-slot EV maximization with contrarian leverage overlay
Public Ownership PublicOwnership Seed-based public pick % estimates for contrarian leverage scoring

All models are in scripts/quant_models.py. Each is optional -- the notebook and simulator degrade gracefully if any model fails or its dependencies (arch, hmmlearn, filterpy) are missing.

Project Structure

final_four_analysis.ipynb         # Main notebook (run all cells)
requirements.txt                  # Python dependencies
scripts/
  quant_models.py                 # GARCH, HMM, Kalman, Prior, EV Optimizer, Enhanced Simulator
  scrape_historical_brackets.py   # Scrapes Wikipedia for historical seed win rates
  scrape_net_teamsheets.py        # Scrapes Warren Nolan team sheets
  filter_tournament_teams.py      # Filters scraped data to tournament teams
  scrape_injuries.py              # Scrapes injury reports from boydsbets.com
  scrape_player_stats.py          # Scrapes player stats, computes injury adjustments
scraped_data/
  tournament_teams.csv            # Team-level metrics (68 teams)
  tournament_games.csv            # Game-by-game results (~2200 games)
  ncaa-net-rankings.csv           # NCAA NET rankings (365 D1 teams)
  kenpom.csv                      # KenPom efficiency data (68 tournament teams)
  bracket.csv                     # Tournament bracket (32 first-round games)
  historical_seed_rates.csv       # Historical seed matchup win rates (1985+)
  injuries.csv                    # Injury reports (tournament teams only)
  player_stats.csv                # Per-player MPG/PPG for tournament teams
  injury_adjustments.csv          # Injury penalties + adjusted NetRtg per team
images/
  efficiency_scatter.png          # ORtg vs DRtg scatter plot
  composite_bars.png              # Top 25 composite score bar chart
  radar_charts.png                # Regional contender radar comparisons
  quant_analysis.png              # Baseline vs enhanced FF probs + GARCH volatility
  injury_dashboard.png            # Injury impact visualization
  ev_bracket_analysis.png         # EV by round + top 10 highest-EV picks

Key Metrics Glossary

Metric Source Description
NetRtg KenPom Adjusted efficiency margin (points per 100 possessions)
ORtg / DRtg KenPom Adjusted offensive / defensive efficiency
NET Rank NCAA Official NCAA evaluation tool used by the selection committee
SOR Warren Nolan Strength of Record -- resume quality ranking
WAB Warren Nolan Wins Above Bubble -- wins beyond the tournament bubble threshold
KPI Warren Nolan Kevin Pauga Index -- predictive metric used by the committee
BPI ESPN Basketball Power Index
Q1 / Q2 NCAA NET Quadrant records (Q1: vs top 75, Q2: vs 76-150)
GARCH Vol quant_models.py Per-team performance volatility from GARCH(1,1) on game margins
HMM State quant_models.py Current regime (hot/cold) from Gaussian Hidden Markov Model
Kalman Momentum quant_models.py Late-season trend from Kalman-filtered margin residuals (0-100)
Leverage EV quant_models.py EV weighted by inverse public ownership for contrarian edge

About

A composite analytical pipeline that predicts NCAA Tournament outcomes

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages