How the scores work

One method for all four leagues. The only thing that differs is where the win-probability data comes from.

The idea. A game is exciting when the odds of winning keep changing, especially late and suddenly. So we take the home team's win probability on every play and measure how much it moved.

Game score blends four measurements, each turned into a percentile against every game in that league this season:

Stakes uses only things known before the game: both teams' records, a bonus when both are at or above .500, division or conference games, rivalries (NFL), and poll rankings (college). It does not know about playoff races yet; that's coming.

Excitement index = 0.8 × Game + 0.2 × Stakes, on a 0–100 scale. Because Game is a percentile, 50 is an average game that season and the top few each week usually land above 80. Season rank is the game's place among every scored game that season.

Where the win probability comes from. NFL, college football and NBA: ESPN's in-game win probability model, read play by play from the game summary as soon as the game is final. NHL has no published per-play model, so we built a simple one: each team's remaining goals follow a Poisson process whose rates come from the betting line, which gives a win probability for any score and clock; overtime and shootouts are treated as near coin flips. Because every league's Game score is a percentile within that league, the leagues are comparable to each other even though the underlying models differ.

Spoilers. The site never shows scores, winners, margins, or whether a game went to overtime. Highlight links go only to the league's official channel, matched by team names and date, and any title that mentions an outcome is rejected automatically. Rankings do leak a little by nature: a lopsided matchup ranked #1 was probably close. That's the point.

Updates. The site rebuilds about every 20 minutes. A game appears once ESPN marks it final and its play-by-play is complete, usually within 20 to 40 minutes of the final whistle.

Source code: github.com/perkinskit/wortha-watch.