How It Works

Dafty model

MLB. Predicts each team's runs, turns that into a win probability and sizes a bet when the price is right. Behind Picks, Results and the model pick on each game.

MLB sim

MLB. Plays each game out plate appearance by plate appearance, 20,000 times. Behind the Slate, every game page, player pages and Props.

NFL sim

NFL. Plays each game out snap by snap, 10,000 times. Behind the NFL slate, Game Script, player pages and the Ledger.

MLB is off for the winter and returns on 2027 Opening Day. Until then the MLB pages show the last slate.

Baseball. Every pitch from Baseball Savant since 2015: velocity, movement, whiffs, how hard and where the ball was hit. Plus fielding, framing and sprint speed; MLB's feed for starters, lineups, umpires and scores; Retrosheet for history.

Football. nflverse play-by-play, rosters, depth charts, injury reports, snap counts and Next Gen Stats, plus game-day actives, the referee crew and the weather forecast.

Prices. US sportsbook lines, captured through the day.

A gradient-boosted model (XGBoost) predicts how many runs each team scores. It reads 47 inputs per team, each measured going into the game, with recent seasons counting more than old ones.

Starting pitching

Strikeout and walk rates, contact allowed, ground balls, Stuff+ and location grades, workload, rest and velocity trend. Both starters.

Bullpen

Contact allowed, whiffs, barrels, velocity and how hard the pen has worked the last three days. Both pens.

Defense

Team Outs Above Average and catcher framing. Both sides.

Lineup

Lineup quality and strikeout rate, adjusted for the starter's hand, plus how the offense has hit over 7 and 14 days.

Situation

Home field, park, the plate umpire's run and strikeout tendencies, day or night, rest days and travel miles.

The predicted run difference becomes a win probability through a curve fit on 12,148 games the model never trained on. Openers lean the projection toward the bullpen. Weather isn't an input.

A moneyline pick's tier comes from the model's win probability for that side:

The size comes from the edge and the price, not the tier. One-eighth Kelly (1u = 1% of bankroll), quarter units, 5u cap. Lean and Moderate dogs get half. Under 0.1u is no bet.

The win probability is set against the best price across the books, vig included. A run line only qualifies when the favorite is at least 62% to win, shorter than −150 on the moneyline, and the −1.5 is priced at −135 or better.

Totals return for the 2027 season. Until then they show the model's number for reference, with no stake.

A game locks 45 minutes before first pitch. The price and size from the last run before lock are what Results grades, win or lose.

Five tests. Each trained on every season before the test year, then predicted a full season it had never seen. Directional accuracy is how often the model picked the team that won.

2025 also helped tune the model's settings, so it isn't a fully blind year. The live 2026 record is on Results.

The sim plays tonight's game 20,000 times, one plate appearance at a time. Each plate appearance ends in one of seven results: strikeout, walk, single, double, triple, home run or an out in play.

Who's batting

The posted lineup. Before it posts, the team's last lineup against a starter of the same hand, marked Projected.

Each matchup

Batter and pitcher skill, pulled toward league average when the sample is small and split by handedness. Plus contact quality, the pitcher's arsenal, the plate umpire, the park, temperature, wind and the fielders behind him.

The bullpen

The starter's exit follows how deep he has worked lately and how the game is going. Relievers come in by their odds of pitching tonight. An arm that worked both of the last two days sits.

On the bases

Runners move the way real runners have from each base-out spot, faster or slower by sprint speed. Extra innings start with a runner on second.

Every number comes from those games: who wins, the run line, totals, the first five innings, the first inning and player props. Recalibrated daily on the last 35 days; reruns as lineups post.

The NFL sim plays each game 10,000 times, snap by snap: the call, the throw or the run, the yards, the clock, fourth downs, kicks and overtime under the current rules.

Team strength

Pass and run efficiency on both sides of the ball (EPA per play), adjusted for opponents. Garbage time counts for less.

The quarterback

Tonight's starter: his efficiency, completions over expected, depth of throw, scrambles and sacks.

Who gets the ball

Target, carry and red-zone shares. When a player is out, his work goes to the teammates who pick it up.

Context

Home field, rest, travel, altitude, dome, wind and rain, and the referee crew. Missing linemen and edge rushers are priced in.

The sim gives win, cover and over chances, plus passing, rushing and receiving yards, passing TDs and touchdown scorers. A second run against league-average defenses powers Game Script. The Ledger is a paper record; nothing is placed.

The market number is each book's price with the hold (the book's cut) taken out, combined across books. Edge is the sim's chance minus the market's, in percentage points.

Far from the market, the sim is more likely missing something. Gaps over 12 points in MLB, 6 on NFL game lines or 18 on NFL props are marked Outside model range: Full Access shows them with a warning, free views hold them.

One game every slate is free in full, even signed out, usually the first game on national TV. A free account sees every game: the projected score, win chance and total, the likely ranges, the market prices and the scouting.

Full Access adds what the model would bet: its picks and sizes, the sim against every line, the props board, trends, signals, NFL lines and the Ledger. See plans

Every MLB pick is tracked on Results: every outcome logged, every unit accounted for. The terms you see on the site are in the Glossary.

Questions? @DaftyDimes on X