Methodology How the models work

MLB

The model plays out each plate appearance. It estimates every batter's and pitcher's skill from several seasons of data, then adjusts for platoon matchups, ballpark, weather, umpire, catcher framing and defense, and simulates each game 1,000 times using posted or projected lineups, probable starters and bullpens. It runs four times a day: a full rebuild the evening before, then three game-day refreshes as starters, lineups and weather post. One limitation: until a lineup is posted, the model uses one projected from the team's recent lineups, so early numbers can move.

NFL

Each team has offensive and defensive ratings built from play-by-play efficiency (expected points added per play), with the margin adjusted when a team changes starting quarterback or loses a key cornerback. The official injury report, game-day inactives and each team's depth chart decide who plays. A drive-by-drive simulation runs each game 5,000 times, and player projections split team volume by recent usage. The model re-runs several times a day, timed to land after inactives for each kickoff window. One limitation: early in the season, ratings lean mostly on last season.

NBA

Each team has pace, offensive and defensive ratings, rebuilt from every game on each run and pulled toward league average when evidence is thin. The ratings combine into projected scores with a home-court factor, and the margin is adjusted for player availability (the league's official injury report) and back-to-backs. Player lines come from minutes and per-minute rate models tied back to team totals. It runs three times a day: the night before, then twice on game day. One limitation: the model is trained on regular-season games only.

NHL

Team strength comes from expected goals, split by even strength, power play and penalty kill; it updates game by game and is blended with last season's form early on. Each side's expected goals are adjusted for the likely starting goalie and for away teams on a back-to-back, then turned into a score distribution with a separate overtime and shootout step. The model runs the night before and then hourly through the afternoon and evening of each game day, so it picks up goalie confirmations and late scratches. It prices the starting goalie DailyFaceoff lists as confirmed or likely; until one is posted, it falls back to a pick from each team's recent workload. Skater injuries reach it only through the projected lineups.

College Basketball

Each team has offensive and defensive ratings per 100 possessions, adjusted for opponent strength, plus a pace rating and a home-court edge; every run refits them from all of this season's games. Before a team has played, it starts from last season's rating pulled partway back toward average. Win probabilities use a slightly heavy-tailed distribution, which fits college scores better than a bell curve. The model runs twice a day, the night before and the morning of each game. One limitation: it works at team level only, with no player injuries.

College Football

Offensive and defensive ratings come from play-by-play efficiency (expected points added per play, success rate and explosiveness), adjusted for opponent and updated after every game. Between seasons, ratings carry over based on how much production each team returns. Projected totals also use pace and the kickoff wind and temperature forecast, and win probabilities come from the spread of outcomes around the projected margin. Forecasts refresh once each morning. One limitation: the model has no injury or starting-quarterback information, and a team in its first FBS season starts at league average, so its games are left out of the record.

Tennis (ATP and WTA)

Each player has serve and return ratings that update after every match, blended with surface-specific form and a small age adjustment; the two tours are modeled separately. An exact point, game, set and match calculation turns the ratings into win chances and game counts, respecting best-of-three and best-of-five formats. The model runs every four hours. One limitation: forecasts for players with little match history are much less reliable, and mid-match retirements aren't predicted.

How grading works

The record

Every completed game is graded against its final score, using the last forecast published before it started. The record counts how often the model's favorite won outright, with the average miss on the projected margin and total beside it.

Calibration

Games are grouped by the favorite's win probability (50–60%, 60–70% and so on) and compared with how often those favorites actually won.

Not counted

Shown with their results but left out of every record figure: preseason games, games whose only forecast was published after they started, college football games against FCS opponents or a team in its first FBS season, and tennis walkovers.

Against the line

Since Oct 10, 2026, every game is also settled against the first market line captured after its forecast was published, DraftKings first: the spread in football and basketball, the moneyline and run line in baseball, the moneyline and puck line in hockey, and the games spread and moneyline in tennis. The model's side is the one it rates above the line. Pushes return the stake; postponed and cancelled games, walkovers and retirements are no action. It is a simulated 1 unit a game, not a recommendation, and a few weeks of games are far too few to tell skill from luck.

See every sport's record →

What this is not

These are statistical predictions published for information. They are not betting advice, and nothing here is a recommendation to wager.