High-danger chances are a location model, not a danger meter
High-danger chances have become the go-to shorthand in modern hockey analysis. A team out-chances its opponent in the high-danger column and the verdict writes itself: they created the better looks and deserved more. Analysts quote high-danger Corsi-for percentage the way an earlier generation quoted shots on goal, as if the phrase “high danger” were a direct readout of how threatening a team’s offense was. But high-danger chances are not measured, in the sense a goal or a shot is measured. They are classified — produced by a model that takes each shot’s location and sorts it into a danger bucket. The label describes where a puck was shot from, and it infers the danger. Those are not the same thing.
Where the label comes from
Public high-danger definitions start from the recorded x-y coordinate of each shot attempt and draw zones on the ice. Shots from the inner slot are tagged high danger; shots from the perimeter are low. Some models add a handful of adjustments — rebounds, rush chances, whether the shot came off a quick play — but the backbone is geography. A shot from a given patch of ice gets the high-danger stamp regardless of what was actually happening on the play. The model is a reasonable generalization, because shots from the slot do go in more often than shots from the point. But it is a generalization about a location, applied to an individual event that the model never actually watched.
Two shots from the same spot are not the same shot
This is where the label frays. A puck one-timed across the slot after a cross-ice pass, arriving on a goalie who is still moving post-to-post, is one of the deadliest plays in hockey. A puck shot from the identical coordinate, straight on, into a set and square goaltender with a clear sightline, is nearly a free save. The location model files both as one high-danger chance, because both left from the same square of ice. Everything that actually makes a chance dangerous — whether the goalie had to move, whether a pass forced a lateral push, how much time the shooter had — happens in the seconds the coordinate cannot see. The stat counts the address, not the play that arrived at it.
Screens, deflections, and traffic vanish
A point shot through a triple screen with a tip in front is genuinely dangerous, and the location model will usually call it low or medium, because it fired from distance. Meanwhile an unscreened wrister from the top of the circle into a goalie looking right at it may sneak into a higher bucket than it deserves. Traffic, screens, and deflections are enormous drivers of whether a shot beats a goaltender, and they are exactly the variables a coordinate-based model handles worst. So the danger label systematically undercredits the greasy, in-tight offense that wins playoff series and overcredits clean looks from the “right” zip code. It is not that the model is careless; it is that location alone cannot carry the weight the name puts on it.
The recording itself is noisy
Underneath the model sits shot-location data logged by human trackers, and that data has known biases. Rink-to-rink scorer tendencies have been documented for years—some buildings systematically log shots closer to the net than others—which means the same shot can land in different danger buckets depending on where the game was played. Because high-danger counts are built on those coordinates, they inherit every quirk of the people entering them. A metric that changes based on which arena’s scorekeeper recorded the night is measuring the recording as much as the hockey. This is the same lesson the shot-quality debate taught the sport a decade ago, and the danger label did not escape it.
Why it is still worth having
None of this makes high-danger chances useless, and dismissing them would be its own overcorrection. Over a full season, at the team level, high-danger differentials do correlate with goals and with winning, precisely because the location generalization is broadly true—the slot really is more dangerous than the point. As a descriptor of territory and shot mix across a large sample, the metric earns its place, and it is a real improvement over raw shot totals that treat a point shot and a slot chance as equal. The trouble starts when the season-long tendency gets applied to a single game or a single shift, where the law of large numbers has not had time to rescue the model from its blind spots.
How to read it honestly
Use high-danger chances as a large-sample territorial stat, not as a play-by-play danger meter. Over sixty games, a team consistently winning the high-danger battle is genuinely generating better shot locations, and that is worth knowing. Over one game, a 12-8 edge in high-danger chances tells you little about who created the superior looks, because the label cannot see the passes, screens, and goalie movement that separated them. Pair it with the eye test and with pass-and-movement data where it exists, and treat the phrase “high danger” as the modeled estimate it is. StatLine’s team and goaltending pages carry the goal and save outcomes that let you check whether a chance edge actually translated into the net.
The playoff blind spot
The model’s weakness is most costly in exactly the games that matter most. Postseason hockey compresses time and space: defenses collapse, goaltenders face a wall of bodies, and goals are increasingly manufactured through screens, deflections, net-front scrambles, and rebounds off initial saves. That is precisely the offense a coordinate-based danger model handles worst, because so much of its value is created in front of the goalie rather than at the release point of the shot. A team can be badly out-chanced on the location sheet while winning the only battle that decides playoff series—the greasy, in-tight one—and the high-danger column will quietly tell the wrong story. In April and May, the gap between where a shot came from and how dangerous it actually was tends to widen, and that is the gap the label cannot measure.
The honest read
High-danger chances are a real and useful abstraction — a model’s best guess at shot quality, built mostly from where the puck was shot — and over long stretches they describe offense better than the raw shot counts they replaced. But the word “danger” promises more than a coordinate can deliver. The stat cannot see the pass that moved the goalie, the screen that blinded him, or the scorekeeper who logged the shot a few feet off. Read it as a location model with a confident name, strongest in bulk and weakest in the moment. The most dangerous chance of the night is often the one the model filed as ordinary.