Skip to main content
⚾MoneyBallScore

MLB Prediction Methodology

The full process behind our MLB game predictions, published without omission.

Core Principles

MLB predictions run on a pure quantitative model. Unlike KBO, which combines sabermetric factors with an AI judge-agent debate layer, MLB skips the LLM debate layer entirely — at 30 teams / 435 possible matchups the scale doesn't fit a daily per-game debate — and derives win probability purely from a weighted sum of 14 quantitative factors (single scoring_rule='mlb_v0.1' version, no agent re-judging confidence day to day).

Every factor comes from public data sources only. Full weights and rationale are published on the 14-factor model page.

Data Sources

  • MLB Stats API (statsapi.mlb.com) — schedule, results, recent form, head-to-head. Official live data.
  • Baseball Savant (Statcast) — xwOBA, Barrel%, starter xwOBA-against, wOBA standard deviation. Batted-ball quality from launch angle and exit velocity — an MLB-only layer with no KBO equivalent.
  • FanGraphs MLB — starter/bullpen FIP·xFIP, lineup wOBA, WAR, defensive SFR.

Quantitative Model

The 14-factor model is the 10 KBO-equivalent factors (FIP, xFIP, wOBA, bullpen FIP, recent form, WAR, head-to-head, park factor, Elo, defensive SFR) plus 4 Statcast-only factors. Team Elo ratings aren't sourced externally like KBO's — we run our own K-factor update loop (K=4, K=6in the postseason, values cited from FiveThirtyEight's published MLB Elo model) after every game. Home teams get the same empirically measured home-field bonus as KBO (+1.5pp).

Full weight table, definitions, and sources are on /en/mlb/factors.

Verification

Every prediction is auto-scored after the game ends and published live on /en/mlb/accuracy. We publish Brier score (a calibration measure) and a calibration chart alongside raw accuracy, so overconfidence shows up transparently rather than getting hidden behind a single win-rate number.

Limits & Disclaimer

  • The MLB model has run for a shorter period than KBO, so its sample size is smaller.
  • Without an LLM debate layer, qualitative context like injuries or trades isn't factored in.
  • These predictions are informational content only, not betting or investment advice.