Full version history and evolution of the MLB prediction engine. Currently running V5.1.
Full machine learning rebuild. XGBoost and Logistic Regression ensemble with 58 features including live weather forecasts, ballpark factors, and pitcher matchup intelligence. 75% market weight ensures the model respects sharp closing lines while finding genuine edges.
Initial XGBoost + LR ensemble trained on 7,179 games (2023 to 2025). 54 features. Walk-forward backtest: 55.9% win rate, +7.30u, +5.5% ROI across 418 qualifying games.
Added pitcher matchup scoring layer. K/9, BB/9, HR/9 scored against opponent strikeout rate, walk rate, and home run per game. 2026 blend: 16% weight at 20 games played, scaling to 100% at 100 games.
Lowered minimum edge to 8% based on backtest showing the 8 to 10% bucket had the best ROI (+3.1%). Added dynamic regression that fades from 0.80 at season start to 0.20 at 100 games played. MAX_FAV_ML capped at -195.
Major rebuild adding Over/Under totals market, bullpen ERA layer, rolling team stats framework, and ESPN data enrichment. First version to post picks publicly to the Telegram channel.
First version to include totals (O/U) predictions. Tempo multiplied by efficiency to project run scoring. Added bullpen quality layer.
The original MLB moneyline model. Basic team stats, pitcher data, and market consensus. Served as the foundation for everything that followed.
Every model upgrade is tested, validated, and deployed automatically. Picks post to Telegram every morning by 8 AM PST.