Machine Learning-Based MLB Baseball Game Prediction System: End-to-End Practice from Data Scraping to Intelligent Forecasting

This project is a production-grade MLB game prediction system developed by Roman Esquibel. By integrating multi-source data such as real-time Statcast data, historical team/player performance, pitcher trends, and recent team momentum, it builds an end-to-end automated machine learning pipeline to generate daily game win probability predictions. The system covers the entire process from data scraping, feature engineering, model training to prediction output, and can be applied to scenarios like sports betting decision-making, team analysis, and education. It features modularity, no data leakage, and full automation.

In the field of professional sports, data-driven prediction has become an important tool for team management, media analysis, and the sports betting industry. As one of the most data-rich sports leagues, MLB's Statcast system records hundreds of thousands of data points per second, providing a foundation for machine learning models. This project aims to build a scalable, intelligent production-grade prediction system to forecast daily MLB game outcomes with high accuracy, supporting scenarios like sports betting decision-making, team performance analysis, and baseball education.

The system is an end-to-end automated machine learning pipeline with core capabilities including:

• Data Acquisition Layer: Scrape game schedules from MLB official website, obtain past 30 days of pitch-by-pitch data via Statcast API, recent team records, and historical results from Baseball Reference;
• Feature Engineering Layer: Convert raw data into features for pitchers (average pitch speed, spin rate, strikeout count, etc.), batters (average exit velocity, home run count, etc.), and team status (recent wins/losses, score difference, etc.);
• Model Layer: Use a random forest classifier to predict home team win probability, output CSV results containing win probability and betting recommendations.

Technical modules include:

• Data Scraping: scrape_matchups.py (daily game schedules), scrape_statcast.py (30 days of pitch-by-pitch data), scrape_team_form_mlb.py (team status), scrape_game_results.py (historical results);
• Feature Construction: build_pitcher_stat_features.py (pitcher features), build_batter_stat_features.py (batter features), map_batter_ids.py (ID matching);
• Model Training: historical_main_features.py constructs a training set without data leakage, train_model.py trains the random forest model. Evaluation metrics: accuracy 92.4%, MAE 0.076, MSE 0.076, MAPE 7.58%;
• Automation: run_daily_pipeline.py automatically completes the entire process from data scraping to prediction output.

The output CSV includes game date, home/away teams, win probability, and recommendation results (recommend home team if home team probability >0.5). In actual tests, the model's accuracy is about 64%, which is better than simple benchmarks (53-55%) and public systems like ESPN Elo model (58-62%) and FiveThirtyEight model (58-62%), placing it in the upper tier.

System application scenarios include:

• Sports betting decision support: Provide data-driven betting recommendations;
• Game simulation and prediction: Assist media analysis and preview;
• Player performance tracking: Identify status trends;
• Team strength visualization: Generate comparison charts;
• Education and research: Serve as a reference for complete machine learning projects.

Project improvement plans include:

• Data Expansion: Integrate weather, stadium factors, injury reports, lineup changes, etc.;
• Model Upgrade: Try deep learning (LSTM, XGBoost), ensemble learning, add confidence intervals;
• Deployment and Interaction: Develop real-time dashboard (Streamlit), automatic alerts, API services.