Every Goal, Replayed
The last second before release for every goal in the tracked archive: all ten skaters, the defending goalie and the puck, sampled ten times a second. Search any goal, scrub through it, and see where it sits among 10,372 others.
Goal map
Similar goals sit close together · click a dot to replayReplay
Attacking left to right · final second before release| Date | Team | Vs | Scorer | Goalie | Archetype | Per · Clock | xG |
|---|---|---|---|---|---|---|---|
| Loading 10,372 goals… | |||||||
What the goals look like
Each goal was described by fifteen tracked measurements: release distance and angle, puck movement, scorer speed, defenders in the shooting lane, bodies at the net front, goalie movement. Clustering those measurements gives six recurring shapes. The borders are soft (silhouette 0.18), so read these as neighbourhoods, not categories. Each card loops the goal closest to its group's centre. The black tick on each bar is the league-wide rate.
Team fingerprints
Each cell shows how far a team's share of goals in that archetype sits from the league share, in percentage points. A boxed cell is outside its 95% interval. With 192 cells, about ten would be boxed by chance alone, so treat single boxes as leads. Click a team to load its goals above.
Findings
What this archive can't tell you
- Goals only. There's no matching set of saved or missed shots, so nothing here is a scoring rate or a defensive skill. "More cross-seam goals against" can mean more cross-seam chances or worse finishing luck; this data can't separate them.
- Two-thirds of goals. 10,372 of 15,678 regulation 5-on-5 goals survived the checks (stable shift-confirmed lineup, all twelve players tracked, a complete one-second window). Close and short-sequence goals are under-represented by 2–3 points each season.
- Release time is inferred, and 1,013 goals carry a location error over 10 ft (flagged in the detail panel). The puck's path after release isn't in these windows; the dashed line to the net is drawn, not tracked.
- Archetypes come from one clustering (k-means, k = 6, stable across seeds, adjusted Rand 0.999). They describe this sample and haven't been tested as forecasting inputs.
- Season-to-season checks use 64 team pairs (two transitions); the permutation p-values treat them as independent, which slightly overstates certainty.