Each sport has its own model, retrained automatically every few hours with the latest results.

Football (NFL & college)

Starts from the betting market's view of each game (with the bookmaker's margin removed), then adjusts it with opponent-adjusted team ratings, play-by-play efficiency, quarterback form, rest and travel (NFL), and recruiting talent and returning players (college). Before a line is posted, a stats-only model is used.

Basketball, baseball & hockey

Every build, three ways of picking each game are re-tested on past seasons the models never trained on, using the same games for all three, and the most accurate one makes the picks:

• Our stats model — an opponent-adjusted team rating (beating good teams counts more; big wins count more than narrow ones), rest and back-to-backs, recent results and scoring margin, record, and MLB probable pitchers / NHL probable goalies.
• The betting line — the sportsbooks' closing moneyline with their margin removed.
• A blend — the betting line adjusted by our stats, where each stat only moves the line as much as it earned in testing.

Right now the blend wins in college basketball (92.8% correct on its most confident quarter of picks). In the NBA, MLB and NHL the betting line itself tested slightly better than anything we built on top of it, so those picks follow its no-vig probability, and the card says so. We'd rather show you the more accurate number than our own. Before a line is posted, the stats model is used.

What goes into a pick

Each game card's Why this pick lists only the biggest few reasons for that game, but every pick weighs all of a model's inputs at once, and each input is itself built from every past game:

• NFL — the betting line (spread, moneyline and total) plus 37 stats: opponent-adjusted ratings, play-by-play efficiency (EPA, success rate, explosiveness, third downs), quarterback form and changes, rest, travel and more.
• College football — the betting line plus 19 stats: ratings, recruiting talent, returning production, last season's SP+, form and rest.
• College basketball — the betting line plus 8 matchup inputs: opponent-adjusted strength (built from every result and margin since 4 seasons back), rest, back-to-backs, recent results, scoring margin and record.
• NBA, MLB & NHL — the pick follows the no-vig betting line (it tested most accurate); the same matchup inputs, plus the probable starting pitcher or goalie, explain the game and are re-tested against the line every build.

In football the betting line is the anchor, and the stats shown under "Why this pick" are how the model adjusts it. Factors that barely move a game aren't listed, so a card with two reasons isn't a model that only looks at two things.

How predictable is each sport?

Tested season by season on games the models never trained on: college basketball 72.5%, college football 73.4%, NBA 68.8%, NFL 66.4%, MLB 57.1%, NHL 55.9%. Every one of those matches the betting market on the same games, which is the best anyone reliably does. Baseball and hockey are close to coin flips by nature. The most confident picks are right far more often than average: 93% in college basketball, 86% in the NBA, about 66% in baseball and hockey.

Data sources (refreshed automatically)

• ESPN — schedules, live scores, probable starters and current sportsbook odds for every sport.
• The Odds API — prices from every major US sportsbook in the live game view.
• nflverse / nflfastR — NFL lines and play-by-play.
• CollegeFootballData.com — college football lines, talent and ratings.

Player props (NFL)

Every time picks are made, the model scores a fixed list of player props for every regular player in every game: passing yards and touchdowns, rushing yards, receiving yards, receptions and anytime touchdowns, at the standard "X+" lines used for player props. Most of them are never shown. A prop appears only if all of these hold:

• The model is at least 75% sure, after adjusting its raw estimate to how often estimates like it actually came true in earlier seasons.
• The line is worth a look — at least 200 passing yards, 50 rushing or receiving yards, or 4 catches. Near-automatic lines (a starter's 150 yards, a receiver's 2 catches) are skipped.
• Props like it have held up: in a season-by-season test on games the model never trained on, that type of prop at that confidence hit at least as often as claimed.
• The player is cleared: the game's official injury report is out and he isn't listed Out, Doubtful or Questionable.
• Its live record keeps up: if a prop type's graded results fall more than 7 points behind what the model claimed, that type switches off automatically.

Props are pregame only: posted before kickoff, frozen once the game starts, and graded afterwards on the track record. Most games will have none, and that's by design. A 78% prop still misses about one time in five, and the betting market prices likely outcomes accordingly, so a high percentage is not "easy money".

Disclaimer

For information and entertainment only. This site does not take bets and is not betting advice. Predictions are probabilities, not certainties: a 70% pick should lose about 3 times in 10. "Why this pick" shows the top few reasons only; the model doesn't rely on just those — it combines every input listed under What goes into a pick, and things it can't see (late injuries, weather changes, motivation) can still decide a game. Player props are probabilities too; none is a sure thing. Odds shown are informational. If you choose to gamble elsewhere, you must be of legal age in your jurisdiction. Gambling problem? Call 1-800-GAMBLER (US).