Shot Quality in the NHL

Shot quality is where expected goals started. "Not all shots on goal are created equal," Alan Ryder wrote in the first line of Shot Quality, a January 2004 paper at HockeyAnalytics.com, and he defined a shot's quality as its chance of becoming a goal. Working from the 2002-03 season (62,351 shots, 5,810 of them goals), he priced each one by its distance, its type and whether it was a rebound, then built a team measure he called shot quality against.

Why did he care about the quality of shots a team allowed? A team that allows dangerous shots makes its goaltender look worse than he is, so Ryder's measure set out to split defense from goaltending. New Jersey allowed the least dangerous shots in his data, and he put the spread between the league's best and worst teams at roughly 19 goals either way for an average club.

The argument that followed was over how much of it teams control, and Dawson Sprigings and Asmae Toumi were still calling shot quality the subject of spirited debate when they published their expected-goals model in 2015. Evolving-Hockey notes that the earliest expected-goals models were often described as shot-quality models. Today the per-shot average falls out of any of them, and Evolving-Hockey publishes it for players as xFSh%, expected goals divided by unblocked attempts.

Price every unblocked attempt with an expected-goals model, add the prices up, and divide by the number of attempts. The model does the real work, and MoneyPuck's prices each shot on these inputs:

• Distance to the net and the angle of the shot
• The shot type
• The last event before the shot, where it happened and how many seconds earlier, which is how rebounds get picked up
• How many skaters each side had, and whether the net was empty
• Whether the shooter was on his off wing

Here's a worked example, with invented round numbers. Team A takes 50 unblocked attempts in a game at 0.06 expected goals each, for 3.0 expected goals. Team B takes 30 at 0.10 each and gets the same 3.0. They've built the same expected goals out of opposite shapes, and the shape tells you whether a team is living at the point or at the crease.

MoneyPuck sorts attempts into danger tiers by the same probabilities, with low danger under 8%, medium between 8% and 20%, and high at 20% or more. Last season 73% of the league's unblocked attempts were low danger and 7.9% were high danger, which is why the league average sits so close to the bottom of the scale.

Shot quality measures danger per attempt, and the league average sits low because most shots aren't dangerous. Last season the average unblocked attempt carried 0.073 expected goals in all situations and 0.061 at five-on-five. Dallas led the league at five-on-five at 0.068 (Seattle and Los Angeles were last at 0.058), and the whole league fit inside a band about a hundredth of an expected goal wide.

Quality and quantity didn't travel together. Across the league at five-on-five, a team's attempt rate and its expected goals per attempt had a correlation of minus-0.10, which is to say none worth mentioning. Volume did most of the work. Attempt rate correlated at 0.90 with a team's expected goals per 60, and quality at only 0.33, so Colorado led the league in five-on-five expected goals per hour on the most attempts in the league at league-average quality. Dallas went the other way, with the best quality and the fewest attempts, and finished 22nd. Quality alone didn't carry it.

Detroit's problem was the quality. The Wings took 40.5 unblocked attempts per 60 at five-on-five, right at the league median, but their 0.059 expected goals per attempt ranked 28th. They allowed 0.061 per attempt, 15th, so the defense was average and the offense's attempts were among the least dangerous in the league. Why does that matter? Because a team taking low-danger shots needs a lot more of them to make up the difference.

Grading defense apart from goaltending

Ryder's original use still works. A team's expected goals allowed per attempt says how dangerous its opponents' shots were before the goalie touched them. Minnesota allowed the least dangerous attempts at five-on-five last season (0.056) and Carolina the most (0.068), even while Carolina finished second in five-on-five expected-goal share on sheer volume. You'd grade Carolina's goalies on a harder workload than its shot counts suggest.

Explaining a shooting percentage

A player's shooting percentage depends partly on where he shoots from, so analysts check quality before they call a hot or cold year luck. Andrew Copp's attempts carried 0.107 expected goals each last season, 31st among the 238 forwards with 150 or more, so his 9 goals were a finishing problem with good looks behind it.

Reading a shooter's role

Per-shot quality tells you where a player does his shooting. Zach Hyman's attempts carried 0.144 expected goals each, the highest rate among regular forwards, against a median of 0.087. Alex DeBrincat sat below the median at 0.083 and made it up on volume. Detroit's highest-quality shooters were James van Riemsdyk (0.117, seventh), Emmitt Finnie and Copp, and none of the three took more than 185 unblocked attempts, which left much of Detroit's volume to lower-danger shooters like DeBrincat, Lucas Raymond and Patrick Kane.

It's only as good as the model

Shot quality inherits every blind spot of the expected-goals model underneath it, including the passes the play-by-play doesn't record and the shot locations scorers get wrong. Ryder flagged the same problem in 2004, when his data had wraparounds logged from 60 feet.

Blocked shots are left out

The average is built on unblocked attempts, so a team whose shots get blocked a lot can post the same number as one that never shoots through traffic. You'd need the blocked-shot counts to tell them apart.

Averages hide the mix

A team at 0.060 could be taking steady mid-range shots or mixing point shots with a few great chances. The danger tiers show you the mix, and the single average can't.

It repeats less than volume

Team shot quality carried over between the last two seasons at a correlation of 0.48 for and 0.54 against at five-on-five, against 0.62 for attempt rates, so one season of it is a weaker signal than the shot counts.

Shooter talent isn't in it

MoneyPuck's base model doesn't factor in shooting skill, so a sniper's shot from the slot is priced exactly like a grinder's from the same spot. That's deliberate. Finishing gets measured separately, as goals above expected.