How a projection is made.

1. Every hitter and pitcher, as rates

For each hitter and each pitcher the engine estimates how often a plate appearance ends each of eight ways: strikeout, walk, out, single, double, triple, home run, or reaching on an error.

• Expected outcomes count three times as much as actual ones. A ball's exit velocity and launch angle say more about the next month than whether it happened to fall in.
• Small samples are pulled toward a starting point. That starting point is the player's previous season where he has one, and the league average where he doesn't. Strikeout rates are trusted quickly (halfway after about 60 plate appearances); singles and triples take hundreds.
• The window is the last 90 days, each day counted the same. Recent games get no extra weight, because in our tests they didn't predict better.

2. The matchup, including handedness

Hitter and pitcher rates are combined with an odds-ratio formula, then adjusted for the platoon matchup. The adjustment is the league-wide gap for that pairing of hands (left-handed hitters, right-handed hitters and switch hitters, against each throwing hand), not each player's own split. A player's own split is mostly noise until thousands of plate appearances, and using it made predictions worse. Bullpens are priced as a mix of left- and right-handed relievers, weighted by how much each pen actually uses its lefties.

3. The fitted correction

A gradient-boosted model starts from those odds and learns a correction from 210 pre-game features: each hitter's and pitcher's results and swing decisions, the pitcher's velocity, movement, release and our pitch-quality scores, how the hitter has done against pitches shaped like tonight's starter's, and the pitcher's role. Trained on 2022 onward and tested walk-forward through 2026, it improved every market it touches in both test windows.

4. Park and pitch mix

Park effects come from real geometry (outfield distances and fence heights) and the air: density from the game-hour forecast of temperature, dew point and pressure, plus wind. The park moves home runs fully, doubles and triples about half as much, and singles about a quarter as much. A batted-ball layer then turns each hitter's contact, built from balls in play only, into singles, doubles, triples and outs for tonight's park.

The pitch mix enters through the model in step 3: its reads of how the hitter has done against each pitch type and each pitch shape the starter throws, weighted by how often he throws them. On a matchup page, "Pitch mix" is what those reads add, measured by scoring the same matchup again with each read set to what the hitter's overall line against that hand would predict; "Starter" is the rest of the starter's effect.

5. The game

Run totals and win probabilities come from an exact walk through every base and out state, inning by inning, rather than a random simulation. Player props use closed-form distributions over the plate appearances each player is expected to get. A starter's expected length comes from his starts only, never his relief outings.

6. Calibration

Where a market has earned one, a calibration curve maps the engine's number to the rate that actually happened on earlier dates. A curve ships only if it beats the engine's own number on later dates it never saw, and every curve is thrown away when the engine version changes.

The record

Every prediction is written before first pitch and frozen. Later builds can add the result and the closing price, never change the prediction. Every market carries a status computed from that evidence; see the model cards and the record.

Our number beside the book's

BTB publishes projections and matchup reports, not plays. Where a book posts a line, our number sits beside it for context. How much to read into a gap:

• Every market carries a status computed from its frozen record; see the model cards. A Research only market is context and nothing more.
• A gap smaller than 3 points, or than the market's own measured calibration error, is inside the noise.
• For strikeouts, the model also has to beat a coin flip at the book's own line on the live record, under the engine serving. It doesn't yet. The current evidence.

No stake is ever suggested, and no BTB model has been shown to beat the book's price.

What the engine does not use

• Weather in run totals. Weather shapes park air for batted balls, but the separate run adjustment has never been switched on; it needs its own held-out test first.
• Recent form. His last 12 balls in play, his last 10 games and his Statcast lines over his last 5, 15 and 30 games, shown for context; not weighted (recent results did not predict the next two weeks better than the season).
• Team defense and bullpen fatigue. Both tested; neither improved totals. The study.
• Umpires, apart from one small strikeout nudge. The strikeout projection reads the home-plate umpire: for a short, fixed list of umpires with a known zone (a hand-set table from the 2024 season), each batter's strikeout chance moves by about one to two points, up for a wide zone and down for a tight one. Every other umpire, or one not yet posted, is neutral, and no other market reads the umpire. When we switched the nudge off in a replay, nothing changed measurably; it touched fewer than one game in ten.