🐻⬇️🏀

2025-2026 NCAAWD2 Model Performance Analysis

Comparing prediction accuracy across 3444 games using multiple rating models.

Model Performance Leaderboard

Models ranked by AUC on 2025-2026 games. Hover over column headers for explanations.

# Model AUC Acc Brier LogLoss n AUC 7d Acc 7d Brier 7d n 7d
1 Bradley-Terry 0.861 76.8% 0.156 0.479 2985 0.828 77.6% 0.171 290
2 Avg Margin (Baseline) 0.852 76.6% 0.160 0.486 2985 0.806 74.1% 0.180 290
3 Pure ELO 0.850 76.2% 0.161 0.492 2985 0.831 75.5% 0.169 290
4 Margin Regression 0.850 76.9% 0.161 0.489 2985 0.851 76.9% 0.161 290
5 Points O/D 0.799 72.5% 0.184 0.548 2985 0.840 76.9% 0.165 290
6 Pythagorean Log 0.799 72.7% 0.195 0.616 2985 0.834 75.5% 0.175 290
7 Pythagorean Adjusted 0.799 72.7% 0.195 0.620 2985 0.836 75.5% 0.174 290
8 Pythagorean Raw 0.790 71.9% 0.189 0.559 2985 0.730 70.7% 0.223 290
9 Home Team (Baseline) 0.570 57.0% 0.246 0.685 2985 0.561 56.2% 0.248 290
10 Bradley-Terry Recency - - - - 0 0.828 74.5% 0.174 290
11 Pythagorean Efficiency - - - - 0 - - - 0
12 Margin Recency - - - - 0 0.856 76.6% 0.159 290
13 Points O/D Recency - - - - 0 0.848 76.2% 0.163 290
14 Core Ensemble - - - - 0 0.851 76.9% 0.158 290
15 Recency Ensemble - - - - 0 0.851 76.9% 0.158 290
16 Adjusted Context Blend - - - - 0 0.853 75.5% 0.159 290
17 Dynamic Bradley-Terry - - - - 0 0.836 75.9% 0.167 290

Methodology

ELO / Bradley-Terry

  • ELO: Iterative updates, K=64, HCA=100
  • BT: Static logistic regression on all games
  • Both model win probability, not margin
  • ELO updates after each game; BT fits once

Pythagorean Models

  • Raw: Bill James formula (pts scored/allowed)
  • Efficiency: Pace-adjusted (pts per possession)
  • Adjusted: KenPom-style opponent adjustment
  • Log: Log-linear multiplicative scale

Margin Regression

  • Team-level ridge regression on point margin
  • Linear Bradley-Terry (margin target)
  • Alpha=0.05 (CV-tuned)
  • Learns home advantage from data (~6 pts)

Baselines

  • Home Team: Always predict home wins (60%)
  • Avg Margin: Higher average margin wins
  • Models should beat these to add value