PropEngine is a from-scratch baseball model, not a repackaged odds feed. Below is a plain-English tour of how projections are built, how every pick is graded from official data, and the metrics we hold ourselves to. No black boxes, no cherry-picked records.
Every batter–pitcher matchup is simulated at the plate-appearance level. Each hitter's expected rates for the relevant outcome — home run, hit, total bases, RBI, or a pitcher's strikeouts — start from their own recent and season performance, then get adjusted for the specific things that actually move the number: the opposing pitcher's arsenal and handedness, the ballpark, weather and wind, the umpire's tendencies, bullpen quality and fatigue, lineup slot, and recent form. Drivers that don't carry signal for a given market are deliberately left out, so each projection shows only the stats that genuinely drive it.
The same engine rolls those matchups up to the game level to produce a win probability for each team (the moneyline) and a projected total of combined runs. These feed the Games page and are graded on their own track record — separate from the player props — because a probability model should be judged on how calibrated it is, not on a single night's box score.
When games go final, the engine reads the official MLB StatsAPI box scores and grades each pick automatically — win, loss, or push — against the line. Props grade off the batter's or pitcher's actual line; moneylines grade off who won; totals grade off the real combined runs. Grading is server-side, so the record updates whether or not anyone has the site open.
Win rate alone is noisy. The honest scoreboard is calibration:
The public track record reflects the picks the engine actually recommends — positive expected value at the price available, or a clear model edge — not every prop it prices. The full board is still graded behind the scenes to keep the model calibrated across the entire probability range, but the headline record is the plays we'd stand behind.
Records are descriptive performance history, not a promise of future results. Markets — especially game moneylines — are efficient, and no model beats them every night. Where live prices aren't available, some markets are graded against the model's own projected line and clearly labeled as such until real book prices return. We'd rather show a modest, accurate record than a flattering one.