Tactics: A hockey stat that we never talk about (Part 1)
An analysis of how ice time and rest time should be distributed in women's hockey
Back in December, at the Open Hockey Conference, Gilles Dignard presented the initial findings of his research on “game segments”. In his presentation, he defined a “game segment” as the time that elapses between puck drop and a whistle.
Gilles’ initial exploration of this metric got me thinking about another fundamental unit of time which we rarely talk about in hockey: “rest time”. Fast forward a few months and I finally found some time and the right data to analyze this metric.
We can define “rest time” as the number of seconds that passes between the end of one shift and the start of the next one. Basically, it represents the amount of time that a player spends on the bench between two shifts.
Understanding the importance of TOI
From my experience, time on ice and deployment patterns are among the stats that coaches care about the most. This is likely because coaches have direct control of both these aspects in a game that is as random as hockey. Ultimately, at any given time throughout a game, their goal is to ensure proper distribution of ice time, with the right combination of players on the ice.
However, no coach has ever asked to come up with a metric to evaluate whether players were resting enough between shifts to perform optimally.
Past research has shown that shorter shifts are desirable to maximize performance. In the conclusion of this article, Greg Revak puts it very simply: “tired players = bad players”. But, on a shift-by-shift basis, the tiredness of players may not only be dependent on shift length.
So, the question is the following. How does rest time prior to a shift influence on-ice performance?
Building the models
To answer this question, we can build a Ridge Regression model to make a prediction regarding Corsi For/60 (CF60), which will be our proxy for on-ice performance. Ridge Regression is useful to mitigate the multi-collinearity that exists between our predictors.
For this analysis, we will use InStat play-by-play data from the latest Women’s World Championship and Olympic Games. Our study will focus on defender data. The list of predictors in our model will include “rest time” as a variable, in addition to the following factors that may impact on-ice performance:
Player Overall Score (calculated as 67% of SARAH score & 33% of N-WHKYe)
Quality of Competition (opponents on the ice during the shift, average score calculated as above)
Quality of Teammates (teammates on the ice during the shift, average score calculated as above)
Group (A or B as categorical variables, excluding Group_B)
Score (-2, -1, 0, 1, 2 as categorical variables, excluding Score_0)
Zone Starts (DZ, NZ, OZ, On-Fly as categorical variables, excluding NZ)
Interactions between the “Group” variable and the “Overall Score”, “QoC” and “QoT” variables were added to our model to adjust for the difference in level of play between Groups A and B. Only even strength shifts were included in this study and the data was standardized prior to running the models.
Results and conclusions
Running separate models for the World Championship and the Olympics, and then combining the data together yields the following results:
We observe that the coefficients for the variable “rest_time” are positive in all 3 models. It is also interesting to note that the p-values in all 3 models for this variable were below the 0.05 threshold. All other things held equal, we can say with 95% confidence that the amount of rest prior to the start of a shift has a positive impact on CF60 (on-ice performance) during a shift. In other words, the more you rest before a shift, the more likely it is that your CF60 will be higher.
There could also be an intangible aspect to rest that isn’t currently captured in our models. That is the quality of “rest time”. The quality of rest would impact recovery, which in turn, would drive on-ice performance. I don’t currently have the data to quantify this aspect of rest time, but it is important for players to not only get the right amount of rest, but also quality rest between shifts.
However, even if we showed that rest time prior to a shift positively impacts performance, linear models may not be the best way to predict ice-time and rest time, on an individual player basis. This is due to the fact that deployment and rest time (for defenders, in this case) are constrained in a non-linear fashion by many circumstantial factors.
For example, a team with a lot of depth at the defense position will be able to use more of their defenders against a strong opponent. This should allow more rest time for every defender on that team.
On the contrary, a team which lacks depth might have to heavily rely on its top Ds to survive a game against that same strong opponent, especially in women’s hockey. This would give the top Ds less rest time but could also be the only way this team stays competitive in the game.
With this in mind, stay tuned for our next newsletter post in which we will build a set of non-linear models to simulate deployment patterns and distribute ice-time/rest time to defenders of 12 different women’s hockey teams, across 60 games.