20 posts tagged “machine learning”.
See how to turn calibrated ML probabilities into disciplined bets: position sizing, risk controls, and the honest limits of the math.
Step-by-step guide to building a sports prediction API in Python, from live data collection and feature engineering to model calibration and FastAPI...
Our NHL model replayed the 2025 Stanley Cup playoffs on pre-playoff data: 52 of 86 games correct (60.5%), with the Cup Final correctly read as a coin flip.
PSG, Arsenal, Atlético Madrid, and Bayern Munich are the 2025-26 UCL semifinalists. Our production club-soccer model (trained on 3,000+ matches, 75%...
Most 'AI picks' sites are ChatGPT wrappers around sportsbook lines. An honest 2026 guide to telling real ML predictors from narrative generators.
Most sports prediction models are accurate but not calibrated. Learn what calibration means, how to measure it, and how ZenHodl applies it across sports.
Our MLB model replayed the entire 2025 postseason on pre-playoff data: 28 of 47 correct (59.6%). Full breakdown, including every missed ALCS pick.
Our NFL model replayed the 2025-26 playoffs on pre-playoff data only: 9 of 13 correct (69.2%), including Seattle over New England in Super Bowl LX.
After the 2025 NBA Finals wrapped with OKC beating Indiana in seven games, we ran our NBA win-probability model on all 84 postseason games using only...
Our first college football model backtested at -2.1c/trade over 728 simulated positions. The missing feature, the diagnosis, and the corrected numbers.
Our college basketball model backtested on all 5,345 non-tournament games of 2025-26: 68.2% accuracy, Brier 0.2064, ECE 4.39%, full reliability table.
Our NCAAMB model replayed all 67 games of March Madness 2026 on pre-tournament data: 48 correct (71.6%), 3-for-3 on the Final Four. Every miss listed.
How win probability models actually work — from Elo ratings through logistic regression to XGBoost. Includes calibration, feature engineering, and the...
Inside ZenHodl's NBA prediction engine: XGBoost models, team stats, real-time injury overlays, and isotonic calibration. How our ML system processes...
Train and evaluate a calibrated NBA win probability model in Python using play-by-play game states, walk-forward validation, XGBoost and isotonic regression.
Our NBA bot had 65% accuracy and was losing money — a calibration bug left it confidently wrong. How we found it, and why calibration beats accuracy.
A 6-feature model outperforms a 50-feature model at making money — even with a worse accuracy score. Here's why, and what it means for how you should...
A 65%-accurate model can be a money loser. A 53%-accurate well-calibrated one prints money. Here's how to measure calibration with Brier score and ECE...
A complete step-by-step tutorial to implement an Elo rating system in Python for sports prediction. Covers the math, K-factor tuning, season resets, and...
How to implement Elo ratings from scratch in Python for sports prediction. Covers home advantage, K-factor tuning, and season resets.