Player Development: Quantifying Skill Blending Patterns
What are the skills that can most easily be blended in women's hockey?
When reading Jack Han’s Hockey Tactics book for the first time in 2020, the following passage stood out to me:
“Barzal takes a collection of merely splendid skills and packages them into a transcendent whole. How? Darryl [Belfry] calls the concept “skill blending” – a player’s ability to multitask at game speed.
Many NHLers carve out successful careers by optimally blending an average set of skills. Others underachieve because they do some things extremely well but only in isolation. Barzal’s standout skills, stacked to perfection, make him nearly unstoppable.”
As we were building the SARAH models with Carleen Markey earlier this year, the importance of skill blending crossed my mind once again. In our analysis, we wanted to incorporate the effect that certain habits had on others, in order to simulate skill blending patterns.
In the initial iteration of our models, we only ended up looking at simple linear relationships and decided against incorporating them in our paper given the lack of correlations in our findings. However, simple linear models do not provide the best framework for our analysis.
For example, in a linear model with no interactions between variables, a player’s ability to get off the boards and her shouldering speed only have a correlation coefficient of 0.24 (which is low). But this analysis doesn’t incorporate a player’s ability to blend these skills in conjunction with other habits. In turn, this added component may result in a stronger blending potential between our 2 habits of interest.
For that reason, attempting to find non-linear relationships between habits would be better suited to perform this analysis.
Statistical Methodology
In order to find non-linear relationships between habits, we can use random forest classifiers, to stay consistent with our initial SARAH models. Random forest models are constituted of a (pre-determined) number of randomly built decision trees which are averaged to make predictions regarding our target variables.
Random forests are useful in this context as the added randomness will help test different variable combinations while integrating a proxy for variable interaction within our models.
Given that the SARAH models included 30 different habits, we will build 30 different random forests. Each model will try to predict the success probability of said habit based on the value of all the other habits below:
This way, we should be able to find non-linear relationships between different habits while adjusting for a player’s ability to blend these skills in conjunction with others.
However, the issue with random forests lies in the fact that they are often viewed as a black box. Random forests lack interpretability regarding how predictions are made: determining weights for each of the predictors is not as easy as in deterministic models such as linear regression.
To solve this issue, we have a few options. Random forests have a built-in “predictor importance” feature, which we used as part of our SARAH modelling process. However, in this case, we decided to explore another way of interpreting the results of our stochastic models: SHAP Values.
SHAP values, short for SHapley Additive exPlanations, allow for the interpretability of different types of statistical models, analyzing predictor impact on an observation-by-observation basis.
For instance, these are the explanatory values for 1 simulation of Maja Nylén Persson’s ability to leverage slip passes to add value to her team’s possessions:
We then average the SHAP values of all the observations to obtain the importance of different predictors in our models.
Skill Blending Patterns
When differentiating skills that can be highly blended, from those that cannot, there are 2 aspects to consider.
First, there are skills that drive other skills and facilitate the initiation of blending patterns. From a modelling perspective, such an impact can be quantified by analyzing the SHAP values of predictors (X variables) as the driver of other habits. For example, skill A (as a predictor) highly drives multiple other skills. Thus, skill A facilitates the initiation of blending patterns by driving multiple other skills.
Then, there are skills that facilitate the maintenance of blending patterns, by easily being driven by other skills. From a modelling perspective, this time, the impact can be quantified by summing the SHAP values pertaining to the predictors driving this given habit (y variable). For instance, skill B (as a predictor) does not drive other habits, but is highly driven by multiple other skills (as a target variable). Therefore, skill B facilitates the maintenance of blending patterns by being highly driven by multiple other skills.
In our analysis, we will consider both aspects of skill blending when differentiating skills that can be easily blended from those that are more difficult to blend.
Results within Skill Sets
When summing up the results of our analysis, while considering both aspects of skill blending, we get the following network graph for habits within their respective skill sets:
In the graph above, the size of points represents how easily a habit can be blended with others while the thickness of lines represents the strength of the relationship between two habits.
As such, in terms of skill sets, the shooting, passing and skating habits seem to be the ones that can most easily be blended together.
While the graph above solely depicts the relationships of habits within skill sets (to facilitate visual understanding), our analysis also includes relationships of habits within different skill sets.
Following the steps outlined above, the 5 habits that can most easily be blended with others are as follows:
(Shooting) Coordination
Deception with Puck
Slip Passes
Shouldering Speed
Getting off the boards
The network graphs below illustrate the strong habit-habit relationships (with overall SHAP values above 0.0325) of the top-5 mentioned above.
(Shooting) Coordination
In our SARAH project, we defined (shooting) coordination as a player’s ability to optimize feet placement and apply downward force on her shot to maximize accuracy & power.
Coordination of the shooting mechanics is a habit that is easily stacked with multiple others. Mainly, it can be well blended with other shooting habits: a good weight transfer combined with the application of downward force yields maximum power on a shot. Good feet positioning also allows for easier tips in front of the net.
In addition, dynamic puck catches – within weight shifts or crossovers – facilitate shooting coordination. Good feet placement on puck reception allows for a more dynamic first touch while initiating the optimization of feet placement on the shot.
Deception with Puck
Within the SARAH framework, deception with the puck is included in the stickhandling skill set. We defined this habit as a player’s ability to pull in players with the puck or give the illusion of making a specific play.
Deception with the puck can be blended with passing habits and can be stacked with elite playmaking abilities. By giving the illusion of preparing a specific play or pulling in players, time and space will likely open in good ice, which in turn, can be leveraged to create dangerous offensive chances.
Slip Passes
Then, we have slip passes, which are part of the elite passing skills that strong playmakers possess. In our project, we had defined a player’s slip passing competence as her ability to identify seams under or above the stick of opponents.
Slip passes are the third most easily blended skill thanks to their strong relationship with other elite playmaking and stickhandling habits (vision, deception w/ puck and handedness versatility). They are useful to add value to a team’s possession through lane changes on the transition or cross-ice movement in the OZ. But, in this case, it is more about the quality of skill blending patterns than the quantity of skills it can be blended with.
Shouldering Speed
Shouldering speed is an important component of a player’s overall fluidity on the ice. It encompasses movement patterns allowing smooth transition during changes of direction or to move from one play to the next.
Overall, shouldering speed is strongly related to different habits in a variety of different skill sets. But it is interesting to note that shouldering speed is a skill that facilitates the maintenance rather than the initiation of blending patterns.
Getting off the boards
Finally, a player’s ability to get off the boards completes the top-5 of most easily blended habits. For many skaters, a large share of first puck touches happen along the boards (in bad ice). Therefore, being able to retrieve the puck along the boards in a favourable posture is important to improve the condition of the puck.
Whether it is by blending this skill with the “feet in motion” habit following the catch (to create separation from the opponent) or by stacking it with passing skills, it all starts with a good posture to ensure control on the first touch along the boards.