MLB Game Prediction Model

Video URL: https://www.loom.com/share/3918571438644e149fb047c1da675311
Kaggle Dataset (mined by me): https://www.kaggle.com/datasets/parkerjackson17/mlb-2015-thru-2025

Project Overview

A machine learning project that predicts MLB game outcomes using historical data from 2015-2025. The model uses differential features between home and away teams to predict game winners.

Dataset

• Time Period: 2015-2025 MLB seasons
• Features: Team statistics (batting, pitching, recent performance), weather conditions, park factors
• Target: Binary classification (home win vs away win)
• Split: Training (2015-2023), Testing (2024-2025)

Key Features

• diff_run_diff: Difference in run differentials between teams
• diff_ops: Difference in On-Base Plus Slugging
• diff_whip: Difference in Walks + Hits per Inning Pitched
• diff_wins_last_10: Recent performance differential
• park_factor: Venue-specific scoring environment
• Weather features: temperature, wind speed

Models Implemented

Model	Test Accuracy	Test ROC-AUC	Notes
Logistic Regression	~0.55	~0.57	Baseline model with L1/L2 regularization
Decision Tree	~0.55	~0.56	GridSearchCV for hyperparameter tuning
Random Forest	~0.56	~0.57	RandomizedSearchCV with 20 iterations
XGBoost	~0.55	~0.57	Gradient boosting with advanced regularization; RandomizedSearchCV with 30 iterations.

Key Findings

• Most Important Feature: diff_run_diff (run differential) consistently dominates across all models
• Baseline Comparison: All models outperform naive "always predict home win" baseline (~53%)
• Home Team Bias: Models predict home wins more accurately than away wins
• Feature Insights: WHIP differential and OPS differential show significant predictive power

Technologies Used

• Python Libraries: pandas, numpy, scikit-learn, xgboost, seaborn, matplotlib
• ML Techniques: GridSearchCV, RandomizedSearchCV, cross-validation
• Feature Engineering: Differential features, standardization, park factors