Skip to content

How it works

Scores, rosters and box scores come from the providers listed on Sources. Everything built on top of them (team and player ratings, game simulations, season odds) is our own, and this page explains how. Every forecast is also scored after the game, and the results are public on Model accuracy.

Team ratings

Each team has two ratings, updated after every game:

  • Elo starts every team at 1500 and moves it after each result. A win counts for more against a strong opponent and on the road, and a big margin counts for a little more than a narrow one. Between seasons, ratings drift back toward average.
  • Adjusted efficiency splits a team into an offense rating (how much more than average it scores) and a defense rating (how much it holds opponents under), corrected for who it played and where. Recent games count more, blowouts are capped, and early-season ratings lean on last season until there are enough games to go on.

Both are in the league's own units (points, runs or goals), so a net rating of +4 means a team would beat an average opponent by about four on a neutral field. The simulator blends the two half and half.

Player ratings

A player's impact rating is their value per game compared with an average player at the same position, in the same units. Each sport uses box-score measures suited to it: weighted on-base average and fielding-independent pitching in baseball, yards per dropback, carry and target in football, Game Score per minute in basketball, goals created per 60 minutes in hockey, and expected goals and assists per 90 minutes in soccer.

Small samples are pulled toward average, so a player with a hot first week doesn't jump to the top. Offensive linemen and others with no box-score stats are treated as average.

Injuries and lineups

Before each game, each player's chance of playing comes from the injury report: out 0%, doubtful 10%, questionable 50%, probable 90%. A team's expected strength drops by what its missing players are worth over a replacement, weighted by how important the position is, so losing a starting quarterback, goalie or pitcher costs more than losing a depth player. Players who have already been out for a while count for less, because the team's recent results already reflect their absence.

Weather, venue, rest and travel

Outdoor games use the forecast for the start hour, refreshed more often as the game gets closer. Domes ignore weather; retractable roofs are open or closed with a probability that depends on the conditions. The effects are deliberately simple and published:

  • Baseball: heat, altitude and wind blowing out to centre field add carry to fly balls and more home runs.
  • Football: wind, cold and rain lower passing efficiency and field-goal range and raise fumbles.
  • Soccer: heat, strong wind and rain lower scoring, and altitude helps the home side.

Back-to-backs, extra rest days, distance travelled and time zones crossed adjust each team's expected score, by amounts set for each sport.

Game simulations

Each game in the next three days is played 5,000 times, and simulated again every six hours or so as ratings, injury reports and forecasts change. In every run the simulator first decides who plays (from each player's chance of playing) and whether a retractable roof is open, then plays the game with a model built for the sport: possessions in basketball, plays and downs in football, plate appearances in baseball, power plays and goalies in hockey, and a goals model with draw correction in soccer. The results give win probabilities, the likely score and its range, the spread and total, and projections for each player's stat lines.

The factor breakdown re-runs the game with one factor switched off at a time (team strength, home field, injuries, weather, venue, rest and travel) on the same random numbers, so you can see how much each one moved the odds. Pro members can change any of these on the Simulator.

Season odds

The rest of each schedule is simulated 10,000 times with a faster model that allows for teams being better or worse than their rating. Every run fills the standings, applies the league's playoff format and records where each team finished, which gives playoff, division and top-seed chances and the range of final records. Season odds refresh after each ratings update, at most every six hours.

Checking the forecasts

Every finished game is forecast again using only information from before it, then scored with the Brier score (0.250 is a coin flip; lower is better), log loss, accuracy and calibration: when the model says 60%, does the home team win about 60% of the time? The simulator is compared with Elo alone, with the league's home-win rate, and with the forecasts the site actually published before each game. All of it is on Model accuracy, league by league.

Limits

  • Probable starting pitchers and goalies aren't in the data, so the simulator picks them by each player's share of starts.
  • Player ratings come from box scores, so defense in baseball and hockey, offensive-line play and on/off effects aren't captured.
  • Leagues without an injury feed are assumed to be at full strength unless a scenario says otherwise.
  • Playoff tiebreakers are random rather than the leagues' full rules.

For informational and entertainment purposes only. Not betting or financial advice. Simulations are probabilistic estimates.

Cookies

Cookies keep you signed in. Optional ones remember your settings. No ads or tracking. Cookie policy