Posted by Jaeson Rosenfeld
29. March 2011 16:23
One key metric in our goal scoring framework is shot quality. In this post we’re going to outline the key components of our measure of shot quality, look at the distribution of shot quality and see how much teams vary on shot quality.
Intuitively, we all know what makes a good shot. One modicum of proof is the rising voice of the announcer as a team works its way into a good scoring position. During working hours, we always have matches on television in the background at StatDNA, and I often find myself turning my head to watch just in time to see the last pass before a goal or a very close miss. Just by hearing the intonation of the announcer, you know something is about to happen. There’s an announcer in Brazil who sometimes uses the call “It’s going to be a goal! It’s going to be a goal!” which he calls before a shot is even taken, and a reasonably high percentage of the time, he actually gets it right.
So visually and from experience, we all know what a high probability shot is, and the shot quality metric is just a way to put our experience into numbers. To produce our shot quality metric, we fit a logistic regression to over 20 parameters that we thought had potential to significantly impact shot quality. The dependent variable was if a goal was scored in that possession (so rebounds get extra credit). The data set included over 4,000 foot shots, excluding penalties, from a sample of 160 Brazil Seria A 2010 games (we will discuss head shots in a separate post). There were 6 independent variables (and combinations of these) that from a shot creation/selection standpoint had a strong fit. The overall model fit is very strong with all coefficients being significant at the 5% or better level, and projected goals in the sample <1% difference to actual goals. The most important variables were (with more “+’s” meaning a stronger positive impact on shot quality and more "-'s" a stronger negative impact):
From our standpoint, most of these match what experience would tell us. We discussed a general metric of shot quality in a post a few weeks ago – shot distance – and we used a proxy of shots in the 18 yard box for shot quality. It turns out that shot distance is the single variable that influences shot quality most; however, the number of opposing players between the shooter and goal (including the goalie) and the level of defensive pressure on the shooter also end up being quite important. Goalie distance from optimal position is not important on a high % of shots, but there are a small percentage of the shots in the sample where the goalie is significantly off his line (sometimes due to a quick pass from a potential shooter) and this brings shot quality up significantly. Shooter body orientation also has a impact only on a small number of shots, by bringing the shot quality down when a player is not directly facing the goal – the extreme example of this is the bicycle kick where the player starts his shot execution with his back to the goal.
What’s quite interesting about the histogram of shots is that we have 68 shots in the sample that had a greater than 50% chance of being scored on. That means in almost half the games one of the two teams took a shot with a 50+% probability of going in! The extreme examples here are three shots with 90+% probabilities – in all three examples the player who took the shot was behind the goalie and less than 4 yards from the goal (and all three scored). Furthermore, 15% of non-penalty foot shots have a greater than 10% probability of being scored, with the average shot having a 7.4% chance of scoring.
Looking more granularly at the 85% of shots that have between a 0-10% chance of scoring, we see that the peak number of shots is at 2-3% goal scoring probability (about 20% of the total sample) and then the number of shots in each 1% range decreases steadily.
From the standpoint of performance measurement (and improvement), it’s interesting to look at how much teams vary on this metric. There are indeed very significant differences between teams in the sample:
The difference between the first and last place teams – Fluminense and Goias is extremely large. In fact a shot taken by Fluminense had an average foot shot quality/likelihood to score of 11% while Goias has a likelihood of 5%. Looking a bit deeper in the numbers, more than 10% of Fluminense’s foot shots came from positions behind every single defender and their shots were 20% closer to the goal than Goias. Fluminense and Goias took almost exactly the same number of foot shots per game (11.9 v 11.8) in this sample, yet Fluminense scored 1.4 goals per game vs. 0.5 for Goias. Shot quality is one key reason why, but as we see the advantage in shot quality (2.1x Goias) is actually less than the advantage in goals per game (2.8x Goias) and to understand that we need to look at two other metrics: finishing quality and goalie performance, which we will do in later posts. We also will examine shot quality at the player level to see which players are creating and getting open for high quality shots as a measure of player performance.