EXPECTED PRIMARY POINTS VETTER THAN POINTS, SHOTS
Primary assists aren’t a very useful metric.
The weights in the Expected Primary Points/60 (PrP/60) model are based on the likelihood of certain passing sequences resulting in a goal. They were broken down as follows: Odd Man, Point, Royal Road, Behind the Net, Center Lane, & Outer Lane.
The rate at which players contribute to Danger Zone events (Royal Road and Behind the Net shots and shots assists) also outperforms Primary Points/60,
Primary assists has little value as a tool to measure performance. Looking at basic shot assists is more than twice as better, expected primary assists three times as better, and isolating just the danger zone shot assists four times as better. Using assists to evaluate a player’s playmaking ability is ridiculous.
Defensemen are typically more valued for facilitating the transition from defense to offense, which is why looking at point totals can obscure a tremendous amount of worth defenseman possess.
One of the big things in recording this data on defensemen is Entry Assists – or how often are they assisting on/generating a controlled entry. After all, a defenseman’s first job should be springing attacks in transition.
We can isolate how repeatable a defenseman’s entry contributions are (any pass leading to a controlled entry or the defenseman entering the zone themselves, provided there is a shot recorded at the end of the sequence). These all have about the same predictive power as a defenseman’s primary points themselves, but are more repeatable and also identify a role vital to sustaining good team offense: that of the puck-moving defenseman
If you’re a team struggling to generate offense, a defenseman with strong entry contribution rates, but lower point totals, could be had on the market for far less than a blue liner with gaudy point totals. Don’t pay a premium by chasing the headliner stats.
The immediate takeaway is that incorporating passing metrics into player evaluation continues to outperform existing public metrics that attempt to predict point production. This remains logical and a missing piece in player evaluation across the league. There is still significant value to be add in the trade and free agent market by exploiting uninformed teams that fail to properly evaluate their players. At the player level, passing needs to be accounted for.
PREDICTING GOALS
By accounting for shot quality via passing metrics we can more accurately predict a team and player’s on-ice goal-scoring rates.
The expected goals formula is derived from the expected primary points formula – a simple weighting of passing sequences based on the likelihood of a goal. Factors included were: Passes back to the point, royal road, behind the net, odd man, center lane, outer lanes, as well shot types like one-timers and rebounds.
An on-ice metric like this will capture much of a player’s off-the-puck movement to create lanes, find open ice, and generate offense through passing. These are all skills that separate great players from simply good ones. Also, it’s what separates great teams from good ones: how effective is the team’s system in manufacturing passing lanes and open ice?
Basically, identifying players that generate passing offense, especially Royal Road and Behind the Net shot assists, will improve a team’s offense. Simple as that.
It’s clear that by a simple weighting of a team’s or player’s on-ice passing attack will outperform current public shot and expected goal metrics. Not only are these more predictive, but some metrics are indicative of specific types of passes, highlighting certain playing styles for a team to analyze. Passing data allows us to take a closer look at the player environment and get a better sense of how they and their team are performing.
xG MODEL WITH PRE-SHOT MOVEMENT
The concept of an xG model is simple: look at the results of past shots to predict whether or not a particular shot will become a goal. Then credit the player who took the shot with that “expected” likelihood of scoring on that shot, regardless of whether or not it went in.
However, there remains additional room for improving these models. There are big gaps in information, and we know that filling them would make us better at predicting goals
Perhaps the biggest gap is pre-shot movement. We know that passes before a shot affect the quality of the scoring chance, but the pbp data does not include them. Thankfully, Corey Sznajder’s data does.
There is also past work that indicates that passing information will improve productivity. Ryan Stimson has done an enormous amount of work tracking and analyzing tactical play, with a particular focus on passes. He built a similar model with passing information and found that it was better at predicting a player’s future point production than past points and better at predicting a team or player’s on-ice point production than past point or shot rates.
Combined NHL pbp and Sznjader data:
· .For every shot, its type, location, distance from the net, and angle from the net
· For the event prior to every shot, the type of event and its location
· The time, distance, and angle change between the shot and its prior event
· Whether the shooter is on the home or away team and the score data
Next, I added info that Sznajder recorded about each shot and the passes that led up to it:
· How many passes (up to three) occurred before the shot, and how many of those passes were in the offensive zone
· Flags to indicate whether the shot was a scoring chance (i.e. from the “home plate” area), results from an oddman rush, and whether or not the goalie was screened
· For each preceding pass, its zone and lane, plus whether it was a royal road, behind the net, low-to-high, or stretch pass
We know that the number of passes before a shot influences the shooting percentage, as do the number of offensive zone passes. His project also defined a lot of the qualitative types of passes tracked and indicated that they would be valuable features.
Results:. The key takeaway is that models do a better job of predicting goals when they have access to pre-shot movement. Consequently, I believe that for games in which we have the data, the pre-shot model is more accurate than existing public models.
That said, you can see from the charts above that the difference is not enormous. The public xG models do a really great job with the data they have. So while we can expect tracking data to improve predictivity, it will likely be a marginal improvement over the xG models we have now, at the very least until an enormous amount of work has been done on incredibly detailed data on the movement of each player.
Variable Importance:. The single most important feature by a wide margin is how close the shot is to the net. Other pbp features like the time since the last event and the angle of the shot also remain very important.
Information about the specific preceding passes is also useful, though to a lesser degree. The first pass-specific variable is whether the pass immediately preceding the shot went across the “royal road” area. This makes sense and further validates that these types of shots are particularly dangerous scoring chances. The model also uses other variables around the location and type of the passes.
That said, this information is directional but should not be taken as gospel. For one thing, many of the variables are correlated with one another, so it is difficult to parse out exactly how important each one is compared to another. For example, the number of offensive passes is not a particularly useful variable, but the model is also looking at how many total passes there are. Much of the same information is included in both variables, so the former would be much more valuable on its own if we did not also have the latter. Keep this in mind as you look at results like the low importance for whether the preceding pass comes from behind the net. Shots with that type of pass are pretty likely to be close to the net and from a scoring chance area, so this correlation may reduce its importance while the pass type itself is still quite important.
we introduced a new expected goals (xG) model. It incorporates pre-shot movement, which made it more accurate than existing public xG models when predicting which shots would be goals.
However, we use xG models for far more than looking at individual shots. By aggregating expected goals at the player and team level, we can get a better sense of how each of them performs.
First, let’s look at each team. By adding up the 5v5 expected goals for and against each team, we can see each team’s total performance. This combines both how many shots the team gets and how dangerous those shots are.
In general, the pre-shot model performed well.
That said, we should not oversell the differences between the pre-shot model and its predecessors.
Some teams have meaningful differences between the models – noted by the distance between the two dots on its row. But these differences are much smaller than the variation between teams within the same model. That is, how a team performs is generally pretty similar between the two models, and that performance matters more than the tool we use to measure it.
If we split the expected goals into for and against and then look at the average per shot, we can get a better sense of each team’s shot quality.
Here, we are ignoring shot volume and just looking at the average quality of each shot for and against.
Success requires combining this with shot quantity as well as shooting and goaltender talent.
It’s interesting that the average xG for and against appear to be correlated. Intuitively, it makes some sense that teams that take dangerous shots may be playing a riskier, more aggressive style that leads them open to more dangerous against. However, this relationship is still much stronger than I would have expected, and it may be worth further investigation
Shooting from a scoring chance area, getting a screen, and shooting off an oddman rush are all important features that we intuitively expect would increase scoring. However, we don’t know the scale of the impact or how different types of oddman rushes compare. Let’s start by looking at the median xG for shots of each type in our dataset:
These results match our intuition: shots with a screen have double the shooting percentage as shots without, and scoring chances are more than triple as likely to go in. In addition, oddman rushes are extremely good scoring chances; the success of 3-on-2 rushes is surprising, but could be the result of a small sample; the rest of the values generally make sense.
Based on this, if I worked for an NHL team, I would have an analyst look much more deeply into understanding screens and oddman rush plays to see what they are doing and which players or teams are able to create them. The oddman rush variable also includes how many men are involved in the rush on each side, which may be more valuable than a binary “yes or no” flag that’s traditionally used for rush information.
We can see that the effect of a scoring chance or screen changes from shot to shot, but generally is a small increase. There’s plenty of variation: on some shots they increase the xG by 0.3, and in a few cases they even decrease xG. But typically there is a 1-2% increase in shooting percentage if there is a scoring chance and a 2-3% increase if there is a screen
Pass Type:
A pass is most dangerous if it comes from the center lane of the offensive zone. Other than that, the zone and lane do not matter very much. This isn’t surprising to anyone who has watched hockey: these passes require the defense to react and the center lane provides the most options. The more advanced methods back this up as well, so I won’t spend any more time on it.
The one that sticks out here is royal road passes: shots are much more likely to score if they come after a pass that moves across the ice in front of the goalie. Passes from behind the net do a bit better while stretch passes are no better or worse than other passes. Passes from low to high look bad here, but as we’ll see, there’s a bit more to it.
In general, almost all shots are improved if the royal road info is added, some dramatically so.
By design, low to high passes likely lead to shots from the point. These shots are far from the net and particularly unlikely to score, so they’ll generally have low xG values. However, that’s because of the location more than the presence of the pass. If we compare the same point shot with and without the pass, it is better off if the pass preceded it, presumably because it implies puck control and moving defenders around the zone. That said, the general low value of these plays implies that low to high passes generally don’t lead to success.
Comentarios