About

Every morning, a machine-learning model predicts each MLB game: who wins, by roughly what score, and how many runs get scored. A second model works one plate appearance at a time and turns into hitter outlooks — the chance each starter gets a hit or homers tonight. Every prediction is logged before first pitch and graded against the box score after. The record is published whole: hits, misses, and everything.

The models train on eight seasons of play-by-play from the official MLB feed and nearly six million individual pitches of Statcast tracking data — exit velocity, pitch types, quality of contact. Team form is distilled into rolling features: recent scoring, starting-pitcher quality over the last ten starts, bullpen workload, and the actual posted lineup — who is in tonight, who is missing, and how much that lineup usually produces. Park effects and game-time weather forecasts feed the run-total side.

The game model is a gradient-boosted system that predicts each team's runs directly — the winner, the score call, and the game total all fall out of the same prediction, so they can never disagree with each other. The hitter model classifies every plate appearance into eight outcomes (strikeout, walk, single, homer, and so on) from the batter's shrunken skill profile crossed with the opposing pitcher's arsenal, then aggregates the night's expected line analytically.

A deliberately simple rating system runs alongside as the in-house benchmark. Every change to the models must beat what came before on thousands of paired games — with significance tests, not vibes — before it ships. Most candidate ideas fail that bar, and the failures are logged too.

A scheduled job runs each morning: it ingests yesterday's games, re-grades every outstanding prediction, refreshes team and player form, retrains on a weekly cadence, and posts the day's picks. A second run in the late afternoon picks up the actual starting lineups once teams post them, sharpening the hitter outlooks and the missing-regulars signal before first pitch.

Accuracy alone can flatter, so the record page shows more: how far the score calls miss, whether a "60% chance" wins 60% of the time, and the model's picks graded next to the betting market's closing expectations — the strongest public benchmark there is. The historical record is a strict walk-forward backtest: for every past game the model was trained only on games that came before it, under the same rules it plays by every night now. An automated test guarantees no prediction ever peeks at the future.

Betting advice. The models and their outputs are published for curiosity, learning, and record-keeping only.

Python and LightGBM for the models; PostgreSQL on AWS for eight seasons of games, plate appearances, and pitches; the pipeline runs as scheduled batch jobs on AWS Fargate; this site is a static Next.js frontend reading a small FastAPI service. The whole system — ingestion to models to this page — is a successor to hoopmodel.com, rebuilt cloud-native for baseball.