How the PuckDelta NHL prediction model works: methodology and testing
Expected goals, player and goalie ratings, game forecasts and a season simulator, each tested on games it had not seen: 13,512 past games forecast with only earlier seasons, every live forecast sealed before puck drop.
The steps
- Official NHL data: every play, shift and lineup sheet.
- Expected goals: each unblocked shot gets its chance of scoring.
- Player, goalie and team ratings.
- Game forecasts: regulation, overtime and shootout chances.
- The rest of the season simulated many times, with official tiebreaks and the playoff bracket.
How it is tested
| Past games forecast with only earlier seasons | 13,512, 2015–26 |
|---|
| Favourite under 60% | 61% |
|---|
| Favourite won in October / April | 57.4% / 62.8% |
|---|
How games end: forecast and actual
| Ending | Forecast | Actual |
|---|
| Regulation | 77.4% | 77.5% |
| Overtime | 14.9% | 14.9% |
| Shootout | 7.7% | 7.6% |
Season projections
| When | Typical miss (points) | Final inside the 90% range |
|---|
| Opening night | 13.0 | 92% |
| Two weeks in | 12.2 | 93% |
| Six weeks in | 9.9 | 88% |
| Three months in | 6.7 | 94% |