Win probability

Each game is simulated 10,000 times, pitch-sequence by plate appearance, with that day's actual lineups. The win probability is simply the share of those simulations each team won. 58% doesn't mean "will win" — it means that in a hundred parallel tonights, we expect that team to win about 58 of them. Baseball is the highest-variance major sport: even 65% favorites lose constantly, by design, and a good model says so.

Fair line vs. book line

The fair line is our win probability converted into American odds with no house margin — what the moneyline "should" be if our model is right. The book line is the sportsbook's actual price (we benchmark against the sharpest widely-quoted market). Comparing them is the entire point of the site: when our fair line is meaningfully better than the price on offer, that's a disagreement worth understanding.

Edge

Edge = our win probability minus the probability implied by the book's price after removing the house margin. A +4% edge on a team means we think they win 4 points more often than the market is pricing. Two honest warnings that we'd rather over-repeat: small edges (under ~2%) are mostly noise — ours and the market's estimates wobbling past each other. And very large edges deserve suspicion, not excitement: when we disagree with the market by 8+ points, sometimes we've found value, and sometimes the market knows something the model doesn't (a late scratch, a bullpen day). The Edges board ranks disagreements; it does not certify them.

Projected total & over/under %

The projected total is the average combined runs across all 10,000 simulations. The over/under percentages are the share of simulations that landed over or under the book's posted line — not a gut lean, a count. Because run totals are lumpy (there's no 8.37th run), a projected total of 8.4 against a line of 8.5 can still produce a near-coin-flip over/under split. That's real, not a bug.

The score distribution chart

The histogram on each game page is the model's full opinion: how often each team scored exactly 0, 1, 2… runs across every simulation. The dashed lines mark the averages. What to look for: width. A narrow distribution means a predictable game; a wide one means volatility. Two games with the same projected total can have completely different shapes — one lives between 7 and 10 runs, the other swings from 4 to 13. Point estimates hide this; the distribution is why we built SaberSlam.

Projected player lines

Per-game expected stats for each starter in the lineup, averaged across all simulations: PA (plate appearances — leadoff hitters get more, that's why the column declines down the order), expected singles, doubles, triples, homers, walks, strikeouts, and the classic rate stats (AVG, OBP, SLG, OPS). Read them as per-game expectations, not predictions: an expected 0.21 home runs doesn't mean "no homer tonight," it means roughly one every five games from this matchup. Useful for DFS lineups and prop context. Both probable starters get a line too — innings, strikeouts, hits, walks and runs allowed while they were still in the game, plus how often they reach 4, 6 or 8 strikeouts. The runs column is labelled R rather than ER because the simulation models no errors, so every run it charges a pitcher is earned by construction.

Game statuses

Projection pending: the game is on the schedule with market odds, but the starting lineup hasn't been announced yet — we refuse to project a lineup we're guessing at. Projected: lineups are in, the simulation has run, numbers are live. Live: first pitch has happened; the pre-game projection stays frozen (that's the honest version — we never retroactively "update" a prediction). Final: the result is in and the game is graded on the Track Record page automatically.

What moves the numbers

Every simulation accounts for: the actual starting lineups and probable pitchers; batter-vs-pitcher-handedness (platoon) splits; each player's multi-season track record blended with industry rest-of-season projections (so a hot week doesn't fool the model, and rookies carry their minor-league signal); ballpark run environments; game-time temperature (domes fixed at 72°F); starter fatigue as the lineup sees him a second and third time, and a performance-aware hook — traffic and runs burn a starter's leash faster than clean innings do, the change usually comes between innings rather than mid-frame, and a starter cruising through a quiet game is left in past his usual workload; realistic bullpen chains with the good arms in close games; team defense; and home-field advantage. When a lineup posts with a star sitting, the projection reflects it within the next sweep — which is often before the market fully adjusts.

Track record metrics

Brier score: the average squared error of probabilities — the standard accuracy score for forecasters. Lower is sharper; 0.25 is what always-guessing-50% scores, so anything below that is real signal. We publish ours next to the market's so you can see who was closer, on the same games. Closing-line value: how the price we logged compares to the market's final (closing) price — the metric professional bettors respect most, because it can't be faked by a lucky week. We report it three ways, because they answer different questions and it is easy to pass one off as another. Mean closing-line value is how far the market moved toward our pick, in implied-probability points, and it is the one we hold ourselves to; the house margin is roughly 2 points, so a figure below that is a real signal that still doesn't cover the toll. Beat the closing line is the plain share of picks the market moved toward at all, whatever the size. Still +EV at the close is the hardest of the three: the share that would remain profitable at the closing price once the house margin is removed. Model lean (totals): whichever side of the posted line the model gave more than 50%, graded against the actual final score. Pushes (landing exactly on the line) count for neither side.

How to actually use this

SaberSlam is a second opinion engine, not an oracle. The workflow we'd suggest: check the Edges board for where the model and market disagree; open those game pages and look at the distribution shape and the data notes (small-sample warnings appear when a projection leans on thin data); check the Track Record so your confidence in the model is calibrated by evidence, not vibes. And the boring sentence that matters: this is analytics for entertainment and research. It is not betting advice, no edge is a guarantee, and no model output is a reason to wager money you care about. 21+. Gambling problem? 1-800-GAMBLER.

Season projections

Once a week, the model simulates every remaining regular-season game 2,000 times at the plate-appearance level — projected rotations, current lineups and bullpens, park factors — then replays the rest of the season 100,000 times through the full playoff bracket, with a team-strength uncertainty term so the win ranges reflect what we genuinely don't know about true talent. Division, wild card, pennant and World Series odds are the share of those replays where each thing happened. Postseason series are pitched by each team's projected October staff — its top four starters, rated individually with the real bullpen behind them — slotted game by game the way a real series is (game one is ace against ace, rotations reset each round) with venue-aware home advantage. Rosters are as of the run date, with one refinement: players on the 10- or 15-day injured list re-enter the simulation at a projected return date estimated from two seasons of real IL stints (rehab assignments sharpen the estimate when they exist). Trades and 60-day IL moves show up at the next weekly update, and the page banner always tells you when that was and when the next one lands. These are projections for analysis and entertainment — not predictions of specific outcomes, and never guarantees.

The same weekly run produces the awards and leader boards below the standings. Those two things are not the same kind of number. A leader board is simulation output: the model draws every remaining plate appearance 20,000 times, adds the result to what a player has already done, and counts how often he finishes on top — so it only covers categories the engine genuinely models, which is why you will not find stolen bases, saves, RBI or pitcher wins there. MVP and Cy Young odds are softer science: the simulator projects the stat lines, but a separate scoring model then guesses how voters would rank them, using weights we chose to reflect how voting has historically gone rather than weights fitted to real ballots. It has no idea about defense, position, narrative or a September collapse, and you should read it as a rough map of the race rather than a forecast of the vote.