How DGB Ratings works

Every team gets an offense and a defense rating, measured against an average team in that league. Offense is how many more points (goals in the NHL) a team scores than average; defense is how many fewer it allows. Net rating is the two added together: how many points better than an average team the team would be on a neutral field.

Each result adjusts a team's ratings for who it played and where. Beating a strong opponent on the road moves a rating more than beating a weak one at home. Ratings update game by game, with older results gradually counting less, so a team that has improved shows it within a few weeks.

Before a season has much data, ratings start from a preseason estimate: betting win totals for college football, the NFL and the NBA, preseason point totals for the NHL, and preseason efficiency ratings for college basketball. That starting point fades as real games are played.

Odds come from simulating the rest of the season thousands of times. Each simulated season draws team strengths from the range the ratings allow, plays every remaining game with a realistic amount of randomness, and then applies that league's real rules: division and conference standings, tiebreakers, seeding, byes and play-in spots. A team's playoff odds are the share of simulated seasons in which it makes it.

The same simulations power the Path Simulator, Wins Needed pages and the Scenario Builder: picking a result keeps only the simulated seasons where that result happened and recounts the odds.

For college basketball, the NCAA Tournament field is chosen the way the committee does it: conference tournament champions get automatic bids, and the at-large spots go to the strongest résumés, judged mainly by wins above bubble (how many more wins a team has than a bubble-quality team would against the same schedule).

The playoff field isn't picked by power rating; it's picked by a committee. So the CFP odds model the committee, not the best team. Résumés are scored by wins above bubble, with a penalty for each loss that grows past two, and the formula is tuned against the committee's own weekly rankings. Some committee judgment that a formula can't capture is added as random noise, so close calls stay uncertain.

The field follows the current 12-team rule: the ACC, Big Ten, Big 12 and SEC champions plus the best Group of Five champion get automatic bids, Notre Dame is in if it ranks in the top 12, and the remaining spots go to the best remaining teams. Conference championship games are simulated too. Losing one doesn't count against a team the way a regular-season loss does, since the committee hasn't treated it that way.

Betting markets are hard to beat, and a ratings model alone trails them. So every published pick blends the two: 80% the betting line, 20% the model, across all five sports. When no line is posted yet, the pick is the model alone. Once a game kicks off, its pick and win probability are locked and stop moving.

Watchability ranks games by how worth watching they are, from three parts:

• Quality (45%): how good the two teams are, by net rating.
• Competitiveness (25%): how close the game is expected to be; a coin flip scores highest.
• Stakes (30%): how much the result moves each team's playoff or tournament odds, measured as the difference between winning and losing this game.

The weights are a judgment call: there's no right answer for “watchable” to fit against. Scores are locked at kickoff, so a game's rating reflects what it looked like going in.

Models re-run several times a week, and more often on game days. The time of the latest run is shown under each sport's menu. Scores, schedules and TV listings come from ESPN and CollegeFootballData.com; betting lines come from the same feeds.

Ratings only know results. They don't see injuries, suspensions, weather or a coaching change until those show up on the scoreboard, so a team missing its starting quarterback can look better than it will play. Probabilities are probabilities: a 20% chance should come through about one time in five, and it will.