Regression Outcome Accuracy
55%
from predicted goals -> win/draw/loss
Headline Results
Classification Ensemble Accuracy
42%
direct multiclass prediction
Regression Recall (Wins / Losses / Draws)
71% / 76% / 4%
draw prediction remained hardest
Classification Recall (Wins / Losses / Draws)
63% / 29% / 20%
class imbalance impacts draws + losses
Pipeline
- Scrape: team match + shooting/passing/creation/defense stats from FBref.
- Clean: normalize teams, parse scores, add time + venue features.
- Engineer: 10-match rolling stats and opponent-aware features.
- Train: time-aware CV with Random Forest and XGBoost.
- Evaluate: confusion matrices, classification reports, feature importances.
Why The Results Are Hard
Your data already shows structural difficulty: draws are only 24.8% overall (1,564 of 6,312 rows), while home-perspective outcomes are skewed (3,096 wins vs 1,652 losses). In this setup, direct classification tends to overpredict wins and underperform on draws/losses.
MLS has also shifted quickly during this window (2018-2024): team count in your data rises from 23 to 29, roster mechanisms are unique and evolving (DP + GAM/TAM + U22), and competition formats have changed. Those shifts create non-stationarity that makes historical patterns less stable for ML models.
Interactive Team Explorer
Pick a team and season to inspect per-match performance and compare it to that season's league average.
Visual Results
