The rating system
Every team carries an Elo rating — the same family of rating system used in chess and popularized for sports by FiveThirtyEight. Everyone starts the season at 1500. After every game, the winner takes rating points from the loser: beat a stronger team and you gain more; lose a game you were supposed to win and you fall harder.
Three adjustments make it sharper than a win-loss table:
- Home advantage. Home teams win more in every league. The model adds a fixed rating bump to the home side, sized per league from historical home-win rates.
- Margin of victory. A blowout says more than a walk-off. Rating changes scale with the log of the margin, damped when the favorite wins big (which teaches nothing).
- League-tuned volatility. A 162-game baseball season moves in small steps; a 34-match MLS season moves faster. The update size is tuned to each league's schedule length.
From ratings to a pick
The gap between two teams' ratings (plus the home bump) converts directly to a win probability. We label picks honestly: a lean is under 58% — a coin flip with a breeze; an edge is 58–66%; strong is 66%+ — and even those lose about a third of the time. Anyone selling you certainty on a single game is selling you something.
The receipts
The model updates its numbers until 15 minutes before gametime, then the prediction locks — and every change along the way is logged publicly, so there are no silent revisions. After the final, the page flips to a recap graded hit or miss, and the running record on the predictions board counts everything (currently 38–28).
The factors — what's in the number vs what's context
The model also watches the weird stuff: game-time weather (temperature, humidity, dew point, pressure, wind and gusts, cloud cover, rain) at every open-air park, rest days, back-to-backs, travel distance and time zones crossed, day vs night, roof type, divisional vs interleague structure, and probable starting pitchers with their FIP. A factor only moves the probability after backtesting against this season's full results proves it improves accuracy — as of July 11 that's the rest edge, back-to-backs, wind and rain variance, orientation-aware wind-blowing-out, and the starting-pitcher FIP gap (validated point-in-time across 1,400+ games, no lookahead). Everything else — travel distance, time zones, altitude, day-game body clocks, matchup structure, player availability — appears on game pages as context, because the backtest says it doesn't help or can't yet be tested honestly. The engine is league-agnostic: baseball's pitcher signal has league counterparts (goalies, quarterbacks, star availability) queued as their seasons and data arrive. MLB data via the MLB Stats API; weather via Open-Meteo.
What the model can't see
Ratings know results, and now schedules, skies, and starters — but not clubhouse mood, bullpen availability, or a manager's lineup card an hour before first pitch. Real games move on things no rating system fully captures, which is exactly why the probabilities stay humble. Treat the picks as a sharp starting point, not betting advice. We don't take bets and we don't sell picks; we help fans get to games — that's the whole business.
MODEL LAST UPDATED 2026-07-17 · RATINGS REBUILD DAILY FROM FULL SEASON RESULTS.