How our predictions work
Every hockey prediction on Hockeyora comes from a rating model, checked against the betting market and tested on past seasons before it goes live.
The model
Every team gets two ratings per league: how many goals it scores and how many it allows in regulation compared with an average team, adjusted for the opponents it has faced. The ratings come from a weighted regression over this season’s finished games, plus a home-ice edge measured for each league. A game decided in overtime or a shootout counts as the tie it was after 60 minutes.
Recent games count most, but slowly: a single hockey game says little, so a team’s rating reflects a long stretch of its season. Early in a season last season’s ratings act as a starting point and fade as new games come in.
The two ratings give each team an expected number of goals, and a Poisson distribution turns those into the chance of every regulation score. Hockey ends level after 60 minutes more often than independent scoring would suggest (trailing teams pull the goalie, leading teams sit back), so tied scores are scaled up by the share measured in each league. Every tie is then split into an overtime or shootout win, leaning slightly to the stronger team.
That one table of final scores prices every market: the moneyline (overtime and shootout included), the three-way regulation result, the puck line, total goals and both teams to score. When bookmakers price a game, the model is blended with their de-vigged consensus (70% market for the moneyline and the regulation result, 60% for the puck line and the total). A pick against the market favourite has to clear a large edge the model itself backs; otherwise it is pulled back towards the market.
What the labels mean
Win probability of 65% or more. Even a 75% pick misses one time in four.
57% to 64%: a real but modest edge.
Below 57%. The model has a side, but the game is close to a coin flip.
Tested on a full season
| Competition · 2023–24 | Games | Winners called | Always home team | Better record |
|---|---|---|---|---|
| NHL | 1,400 | 59.4% | 53.7% | 57.4% |
| KHL | 864 | 59.8% | 55.1% | 59.9% |
| SHL | 411 | 60.3% | 56.7% | 55.7% |
| Liiga | 495 | 61.6% | 54.9% | 56.1% |
| Czech Extraliga | 422 | 64.9% | 54.5% | 63.4% |
| National League | 408 | 60.3% | 58.1% | 58.8% |
| DEL | 404 | 58.7% | 56.7% | 58.6% |
Walk-forward: every game predicted using only the games played before it, the model alone (no bookmaker prices). “Always home team” and “better record” are the simple rules the model has to beat.
Hockey is the hardest of the big sports to call: one bounce or one hot goalie decides many games, and even the betting market picks the NHL winner only about 60% of the time. The model beats always backing the home team in every league. Against the team with the better points record it is level in the KHL (59.8% against 59.9%) and the DEL (58.7% against 58.6%) and ahead everywhere else, and its probabilities were closer to what happened in all seven.
What it cannot see
- The model does not know who starts in goal, who is injured or who is resting; the market blend is how that news reaches the probabilities.
- Early in a season, and for teams new to a league, ratings lean on last season and are less certain.
- A 65% pick still loses about one time in three. Probabilities describe many games, not a single outcome.
Every pick in public
Every prediction is published before the face-off and settled after the final buzzer. The accuracy page keeps the running record by competition and by confidence, misses included.
Predictions are statistical estimates, not guarantees. If you bet, set a budget first; our responsible gambling page lists free, confidential help.