English Premier League Throw-in Data Analysis (as of 23/24,April)

-  This is an English translation of an article written in Japanese at DeepL.

In this article, I will present throw-in data from the English Premier League using data from the Wyscout API (by hudl). To be precise, the data is based on the Wyscout API source with logic adapted to my own analysis. Please refer to the following article for more information on the concept of this area.

The range of the data is from the 19/20 season onward to last weekend's games (with the exception of a few games from past seasons for which no data was available). 

The Value of Throw-in Attacks

Throw-ins themselves are a common occurrence in the Premier League, with about 20 per team per game. Throw-ins are actions that occur when a player breaks the touchline after an opponent touches the ball, and we believe that the higher the league level, the fewer the number of throw-ins. This is due to the number of actions on the side, the quality of the action itself (whether or not there is an error in passing or control of the side change), the number of clearances to the side, and other factors. Therefore, even between good teams, if there are many cases of safe clearing to the side, the number of such cases will increase, but nowadays, even players in back positions are required to have ball retention skills, so there are fewer such cases than in the past.

Ball possession often starts with a set play or a transition during in-play, and from there the ball is approached to the opponent's goal in different ways, The table below shows the number of set plays (excluding kickoffs) by start, the percentage of set plays that resulted in a shot within consecutive possession, and the expected goal value (xG) when a shot was taken. (Throw-ins and free kicks are divided into own team and opponent team because they cover a wide area.)

Number of shots by expected goal range. (The further to the right, the more shots are likely to be on goal.

The xG data is not explained here, as those who are interested in data to a certain extent will already have an idea of what it is about. The same is true of Wyscout's xG, so you can assume that the accuracy is somewhat reduced. (This is completely subjective, but I think the best xG output in the preliminary data is the CG of the Bundesliga local broadcast.) With this assumption in mind, the first table shows that the xG is low if you take the throw-in straight to the shot, and it is difficult to create a decisive point. I believe that this is due to the fact that the game starts with hands from the side and that it allows some time to re-set the defensive positioning. Therefore, it is probably not a priority in the strategy of the match. However, since it is a one-point game, it is not uncommon for the game to never end with such a throw-in tangle.

Data Analysis Design for Throw-ins

The data analysis design for throw-ins is shown in the figure below, taking into account the characteristics of Wyscout.

“Time to Throw-in” is the time from when the ball is out of play to when the throw-in is thrown. This is the time until the ball arrives and the time when communication with the umpires, such as substitutions and cards, is likely to occur, so there are cases in which this time cannot be controlled. In addition, when Wyscout checked the video, the timing of outplays tended to shift a little toward the late side, so the figures may be a little early. This makes the data unstable, but I want to cover this time because it is important in terms of whether or not it gives the defense time to set up again. 

The area and direction of the throw-in is obviously important because it affects the distance to the goal. If you are close to the goal, you can get closer to the goal with less action by launching long throws toward the penalty area. In this case, we defined a long throw-in as a throw-in that is sent to the central part of the penalty area (inside the half of the near zone).

It is also important to analyze whether or not a duel occurs after the throw-in. As discussed in a previous article, Wyscout has more duel inputs than any other data company, so it is easier to obtain a situation where players are in close proximity or not, even without player placement data. Dueling is divided into aerial, ground, and loose ball, but either way, the presence or absence of dueling is an important variable because ball possession can be unstable in either direction.

It is best to keep possession of the ball after a throw-in, but even if the duel is lost and the ball is taken away, it can be regained quickly if the players are in close proximity. In this case, we have chosen to use the recapture (recapture rate) when the player recaptures the ball within less than 5 seconds after the loss. If the player is good at recovering the throw-in, or if the player wants to attack counter-attacking rather than holding the throw-in, it can be said that the accuracy of the throw-in itself is not of particular concern. 

One of the KPIs is whether the player can continue to hold the throw-in for 5 seconds, so whether the player loses 5 seconds or continues to hold the throw-in. In this case, it is not included in the loss.

Overall Trend

Let's look at how the data changes with each trigger.

The data is clearly divided into two groups: the earlier the restart, the less likely a duel will occur and the more likely the team will be able to keep possession of the ball. The earlier the restart, the less likely a duel is to occur and the more likely a player is to keep possession of the ball. However, because the players are often not in a crowded situation, the percentage of players who regain possession when they lose the ball is lower.

Naturally, the throw-in will vary greatly depending on the area where the throw-in is made and towards which direction the throw is made. Long throws create a duel once every two times, but since the player can shoot closer to the goal immediately after the throw-in, the stats around the shot are more likely to increase. The closer the player is to the team's goal, the less likely he is to be shot at even if he loses, but if he is shot at, his xG will be higher. I think this is due to the fact that the shots are created by counterattacks. In other cases, the duel is more likely to occur when the shot is thrown forward, and almost never occurs when the shot is thrown backward. Therefore, if you want to be sure of retention, you should throw backward, but if you let your own team retreat and lose, the shooting rate will jump. It is rare, but sometimes seen, when a throw-in ball arrives with the backward receiver unprepared and the player panics.

Dueling immediately after a throw-in occurs at a rate of one in four; of the three types of dueling that Wyscout has, the ground game is the most common immediately after a throw-in. Although the throw-in is a ball thrown from above, the aerial duel is basically a head-to-head duel, so if it is a case of a chest-trough duel, it is a ground duel. Loose ball duels from throw-ins probably occur when there is a slight discrepancy between the distributor's distribution of the ball and the receiver's position. As shown in the table, the numerical trends seem to be case-by-case, each with its own advantages and disadvantages. 

These are the overall trends for the league after 19/20. Let us now look at the team data as of April ,15 for 23/24. The order is based on the highest throw-in game average.

Team Stats

The size of the circle is the rate of duel occurrence.

Not only around throw-ins, but Manchester City's (Manchester City) data is often outstanding. It seems that they (and their opponents) limit the occurrence of duels only to the point of victory.The restart time is the fastest, and from the throw-in, the priority is to keep the ball, and although the immediate recovery rate after a 5-second loss is low, they don't seem to be that conscious of the assumption that they will be robbed because they don't lose the ball in the first place.However, in the case of Man.City, the situation where the opponent forms a block at a low position and there is no choice but to move back seems to have been taken into account.

In the case of Man.City, the opponent forms a low block, so there is no choice but to bring the ball back to the back.West Ham and Bournemouth fall into this category.West Ham, in particular, also has a high shooting rate after losses.

Speaking of throw-ins, Arsenal's Tomiyasu's delay early in the season became a hot topic in Japan, and his median time to throw-in (14.4 seconds) tended to be the longest among the teams.This has affected the duel occurrence rate and the loss rate, but with the recovery rate of 76.9%, it has not become a major problem, at least in the league matches.Card trouble is a problem, though. 

Finally, here is some data on team exposure, i.e., against opponent throw-ins.

The teams that don't give their opponents throw-ins in the first place are lined up with big clubs and teams that are good at holding them. This is the opposite of the previous data, so the scatterplot tends to be better in the upper left and worse in the lower right.West Ham is also on the worse side, with a low takeaway rate and a tendency to lose possession quickly after taking it.As I mentioned at the beginning, throw-ins are not necessarily important, so even with this kind of data, West Ham can still be at the top of the list, but we may want to rethink the design around throw-ins a bit more.

Although their names are not shown due to Tableau's specifications, Chelsea is next to Arsenal in the upper left corner of the graph. These two teams can be said to have a tendency to have a high takeaway rate after an opponent's throw-in and not lose the ball immediately.Although the number itself is not large, Arsenal recorded an xG of more than 0.1 in all cases where they took a shot after taking a throw-in.Although the data presented here is from the Premier League, the 2-2 tie in the recent CL quarterfinal 1stLeg was also an attack after taking a throw-in from Bayern's own forward system.

I could go into more specifics, but it's a long story, so I'll stop here this time.