About the Model

2026 MLB Season Win Predictions

Model Architecture

Ensemble Approach

Three machine learning models combined via weighted average:

XGBoost (50%) - Gradient boosting with 300 trees, depth=5, lr=0.01 (optimized)
Random Forest (40%) - 300 decision trees, max features=sqrt(n) (optimized)
Ridge Regression (10%) - Linear model with L2 regularization

Weights optimized via autoresearch (43 experiments, 1.5% MAE improvement)

Training Data

359 team-seasons from 2014-2025 (12 years, all 30 MLB teams)

Validated via walk-forward cross-validation on 2021-2024 test years.

43 Predictive Features

Historical Performance (12)

Wins (previous year)
Pythagorean wins
OPS, ERA, WHIP
Runs scored/allowed
Home runs, batting average
3-year rolling averages

Advanced Metrics (19)

WAR: Batting, Pitching, Total, Bullpen, Rotation
Park Factors: Runs/HR adjustments for ballpark
Home/Road Splits: Win% at home vs away
Age: Roster age (batting/pitching)
Pythagorean Luck: 3-year mean reversion
Schedule Strength: Division opponent quality
Manager: Win% and tenure

2026 Projections (7) + Historical Projections (3)

Historical Projections (ZiPS/Steamer 2014-2025): proj_wins, proj_over_under, proj_win_pct (47.6% importance!)
2026 FanGraphs: OPS, ERA, runs, WHIP projections

Contextual (3)

Pythagorean luck (mean reversion)
Over/under previous year
Win percentage trends

Top 10 Most Important Features

1. pythag_wins_prev

20.6%

2. wins_prev_year

6.0%

3. home_win_pct

5.1%

4. era_prev_year

5.0%

5. whip_prev_year

5.0%

6. runs_allowed

3.9%

7. schedule_strength

3.5%

8. total_war

3.3%

9. park_adj_runs

3.1%

10. bat_war

2.9%

Features ranked by XGBoost importance scores. Historical projections (proj_wins, proj_over_under, proj_win_pct) dominate with 47.6% combined importance.

Data Sources

Historical Statistics

Baseball Reference via pybaseball (2014-2025)

Historical Projections (2014-2025)

ZiPS/Steamer archives from FanGraphs blog posts (360 team-years, 0.706 correlation with actuals)

2026 Projections

FanGraphs Depth Charts (scraped via Firecrawl)

Transactions

Baseball Reference 2026 transaction log (525 moves, WAR-matched)

Injuries

FanGraphs Injury List (real-time, 122 players on IL)

Home/Road Splits

MLB Stats API (2014-2024 historical splits)

Manager Data

Baseball Almanac, Baseball Reference (2026 roster)

Model Performance

3.69

Mean Absolute Error (wins)

Walk-forward CV (2021-2024)

0.877

R² Score

88% variance explained

81.8

Mean Predicted Wins

Expected: ~81 wins