NHL analytics: understanding hockey prediction models

Hockey sits near baseball on the predictability spectrum. Games are decided by a handful of goals, pucks deflect off skates and sticks, and one hot goaltender can steal a game his team was outplayed in. For a prediction model, that means modest probabilities and a constant fight against noise.

Why hockey is hard to predict

A typical NHL game features only a few goals per team. When scoring events are that rare, each one carries enormous weight, and a single bounce can decide the outcome. The better team frequently loses — not because it was secretly worse, but because five or six scoring events is too few for skill to reliably show through.

Overtime and shootouts add another layer. A game tied after regulation may be decided by a short three-on-three period or a skills competition, both of which are closer to coin flips than regulation play.

What the numbers look like

Realistic NHL win probabilities for regular-season games rarely stray far from the 40–65% range. EdgeIQ's engine uses a starting home advantage of 33 Elo points for the NHL — about a 55% chance for the home side between equal teams — and a small rating update per game, reflecting how little any single result reveals.

Inputs EdgeIQ uses for hockey

• Team rating difference, updated after each result and weighted by margin.
• Goals for and against per game, plus average margin.
• Recent form over the last five and ten games.
• Home/away splits, rest days and games in the past week.
• Strength of schedule and recent head-to-head results.

What analysts often add

Public hockey analytics has developed shot-based measures — shot attempts, shot quality, and expected goals — because goals alone are too rare to measure team quality quickly. These can be more predictive than goal differential over short spans. EdgeIQ's current engine works from game-level results rather than shot data, which is a real limitation and a reason its hockey projections should be read as broad team-strength estimates.

Goaltending

Which goalie starts can matter a great deal, particularly on the second night of a back-to-back when teams often use a backup. EdgeIQ does not ingest confirmed starting goaltenders, so checking the announced starter is a sensible step before relying on any NHL projection.

In hockey, a 58% favorite is a solid favorite. Upsets are frequent by the nature of the sport, not a sign that a model is broken.

How to judge an NHL model

• Use probability-quality measures like Brier score and calibration rather than pick accuracy alone; winner accuracy will always look modest in hockey.
• Compare against simple baselines — home team, or a basic rating — on the same games.
• Expect long streaks of results in both directions; hundreds of games are needed to judge fairly.