TELEBASEBALL

ABOUT TELEBASEBALL

DATA INFRASTRUCTURE

HISTORICAL DATA

Retrosheet data spanning 1960-present—over 60 years of play-by-play, box score, and game state records.

FEATURES

Advanced sabermetrics (wRC+, wRAA, BABIP, FIP, xFIP), rolling averages (7/14/30/60 game windows), park factors, weather normalization, rest days, pitcher-batter matchups.

INGESTION

Daily team and player performance metrics, automated validation, temporal feature engineering with one-game lag to prevent lookahead bias.

MODEL ARCHITECTURE

MONEYLINE MODEL

LightGBM gradient boosting. Bayesian hyperparameter optimization. Temporal cross-validation. Outputs calibrated probability distributions. Primary drivers: pitcher quality, recent form, situational context.

TOTALS MODEL

NGBoost probabilistic regression. Generates full probability distributions over run totals. Uncertainty quantification critical for over/under. Features: park factors, weather, umpire tendencies, pitcher quality, offensive form.

EDGE IDENTIFICATION

MARKET ANALYSIS

Odds aggregated from multiple sportsbooks. Consensus closing lines constructed. Model implied probability vs. market implied probability comparison.

BACKTESTING

Historical validation against 2010-present data. Walk-forward analysis for threshold optimization. Edge requirements calibrated to Sharpe ratios and maximum drawdowns.

BET SIZING

Kelly Criterion based on estimated edge magnitude. Bankroll management constraints enforced. Minimum edge thresholds prevent over-betting.

PRODUCTION PIPELINE

ETL PROCESS

Automated data ingestion, validation, and storage. Previous day's performance metrics processed. Schedule parsing extracts lineups and starting pitchers.

FEATURE ENGINEERING

Rolling averages recalculated with temporal shifts. Fresh odds data scraped and normalized. Feature vectors constructed for model inference.

PUBLICATION

Predictions published ~60 minutes before first pitch. Full audit trail maintained for reproducibility. Performance attribution tracked end-to-end.

QUALITY ASSURANCE

• Temporal data integrity: rolling averages shifted back one game

• Cross-validation uses time-series splits, not random splits

• Performance monitoring via out-of-sample metrics (log loss, Brier score, AUC-ROC)

• Drift detection alerts trigger model retraining when performance degrades