Get Ready to Throw It Back with the Ultimate Blast from the Past
Whether you have a strong affinity for long attacking throw-ins or a deep-seated dislike for them, it is impossible to deny that they have become both a crucial feature and a contentious issue in men's soccer over the past year. John Muller likely sparked a renewed interest in this tactic, as well as a potential Arsenal title, with his 2023 article for The Athletic, and Joe Lowery and I adopted his method for Backheeled when Minnesota United started employing long throw-ins in 2025. However, while each game typically features around 40 throw-ins on average, only approximately 10 of these throws occur close enough to the box to have a significant impact. Nevertheless, apart from jokes about consultant Thomas Grønnemark on Formerly Called Twitter, there has been a notable lack of commentary about the other types of throw-ins in popular media or public analytics circles. The only exceptions that I am aware of are Eliot McKinley's 2018 two-part article on this very website, and some recent academic research on the top 5 European leagues that, if you enjoy in-text citations and interpreting regressions, serves as an excellent preview for the remainder of this article.
Eliot conducted that research almost a decade ago, and I believed it was time to replicate and expand upon those findings with the impressive infrastructure that ASA has developed since the days of using CSV files on Dropbox. In addition to models that estimate throw completions and retained possession, I also analyzed the goals added for every possession following each throw to assess the value of throw choices. This allowed me to identify the MLS throw-in MVPs and provide an expanded set of general rules for approaching these often-overlooked moments of play. Before delving into the updated data, let's briefly review the first two seasons of Game of Throw-ins. In Part One, Eliot modified ASA's trusty xPass model for throws in two ways: adding a predictor for time since the previous event, known as xThrow, and using retained possession as another measure for a successful throw, referred to as xRetain. It is essential to recall some critical definitions: a successful throw, as measured by xThrow, is a throw that is touched by a teammate first, while a retained throw is a throw that results in the team maintaining possession afterwards. For instance, a player can throw the ball to a teammate's head, and even if the opponent deflects it, the team can still retain possession, or they can throw it off a defender's back and recover the ball themselves, which would be considered an unsuccessful throw but one that still results in retained possession.
After examining these models in detail, Eliot provided three key lessons for coaches to impart to their players. Firstly, it is crucial not to let the ball go out of bounds deep in one's own half if it can be prevented, as the likelihood of retaining possession after a throw-in in the defensive third is essentially a coin toss. Therefore, if there is no significant pressure from the opponent, it is a wise decision to make an effort to keep the ball in bounds. Secondly, when the opponent has a throw-in in their defensive third, it should be used as a pressing trigger, as the probability of a turnover following a throw-in is high, and the team should be prepared to capitalize on this opportunity. Thirdly, it is recommended to take throw-ins 5-10 seconds after the ball goes out of play, allowing the players to quickly get into position to receive the throw-in before the opponents can establish their defensive shape. I would also employ some gamesmanship, particularly in high-stakes matches, by having my players attempt to slow down the opponent's throw-in to give my defense time to organize.
One of the interesting developments when revisiting the original research is that the second point has indeed become a common tactic, with many teams now using throw-ins as a pressing trigger to start the game. In Part Two, Eliot grouped MLS teams based on their throw-in direction choices and evaluated teams and players on their retention success compared to model expectations. He discovered that while most teams lacked a defined style, some were clearly coached to throw the ball forward, while others were instructed to spread the ball around and often throw it backwards. Although the results were somewhat noisy, I noticed that two teams and their fullbacks had remarkable success retaining throws by spreading the ball around and throwing it backwards, and they were coached by Gregg Berhalter and Peter Vermes, who at the time favored possession-focused styles. More than anything, Part Two suggests that few coaches were paying particular attention to throws, and that throwing the ball backwards seems to be helpful for retaining possession, especially for teams that prioritize possession.
However, retaining possession is only one aspect of the game, and having the goals added for every possession following each throw allows us to push this analysis further and better understand how these decisions affect the immediate odds of scoring. Do throws down the line compensate for their low success odds with increased goal danger? Can teams reduce their risk of turnovers while also generating attacking value? Which are the best throw choices in different areas of the pitch? As I was working on the models, I realized that I was struggling to think through all the possible throws available, considering the various contexts in which they can be taken. For example, a throw that a player receives right on the half line could have been taken 30 yards behind them, 30 yards in front, or anywhere in between, and this context drastically changes the nature of the throw. Similarly, the distribution of throw targets in the attacking third is likely different from that in the defensive third. How can I visualize this complexity when there are so many overlapping patterns? The only solution is to make it interactive.
Before proceeding, take a moment to explore the interactive visualization, clicking around the touchlines, examining the different metrics, and noting the patterns of where throws go and how they tend to end up. Both sidelines exhibit roughly the same patterns, so we can focus on what happens when we move up and down the touchline. Taking a throw from one's own defensive corners is a rare but perilous situation, with no reliably good options, as retention rates and possession values are among the worst in the data. The next few origin zones start to reveal some patterns, with the most popular target zones being forward towards the goal, but these have low retention rates and possession values. Backwards passes begin to show favorable retention rates, and throws toward a team's own goal actually have the highest possession values by g+, which is an interesting finding that somewhat contradicts most traditional g+ evaluations, which often prioritize large field positions over everything.
As you approach midfield, these patterns evolve and solidify, with the overall spread of where throw-ins go growing, and it becoming increasingly clear that forward throws in the defensive half are not worth much, consistently having lower retention rates and possession values, while backwards throws yield better retention and g+ value. Forward throws just short of the midfield line are especially suspect, but possession values start to improve as these throws start to go past midfield. Zones with the highest g+ values are often the furthest, usually backwards but also in lateral and attacking directions, indicating that long throws, likely to teammates in space, are the highest upside choices. This makes sense, given that opponents are likely to condense the area as much as possible. The upside of trying to connect passes inside is a fascinating finding, as it mostly goes counter to many years of coaching orthodoxy, suggesting that, in one's own half, throwing the ball up the line, gaining yards, and then building from there can be a valuable strategy.
Once the ball thrower is in the attacking half, their options begin to open up even more, with putting the ball close to the box immediately becoming a high-value option, and new gaps near the goal line opening up as well. The retention rates of forward passes improve, and they actually surpass the completion rates when thrown into the box, but the risks also seem to increase, as this appears to be where most of the negative g+ average tiles appear, intermixed with highly valuable zones, likely driven by counter-attacks on unbalanced teams or juicy turnovers trying to play back to the middle of the field. Backwards passes lose some of their reliability, but there are pockets where they still generate dangerous possession when throws have reached the final third. Short throws also become viable, as while they are often pressing triggers for the opposing team when taken in the defensive half, the reduced distance to the goal may force defenders into withdrawn positions, creating more space. And, in an article about throw-ins, the line you all came for, long throws into the mixer remain very worthwhile.
In general, this tool demonstrates how the geography of the pitch is foundational for optimizing throw-ins, with team strategy influencing how many throws are taken and where, and players better understanding which areas to look and move to when putting the ball back into play. Unless your team is built to win 50-50 challenges and repress when you lose them, crowding everyone forward is a bad strategy, unless you can reach the box, in which case crowding everyone forward is a good idea. I follow Eliot's modeling framework but with a few tweaks of my own, using possessions as defined for goals added, which differs from Cheuk Hei Ho's possession definitions from his expected possession goals work. The results are consistent with both the old model results and the interactive visualization, which draws from 13 seasons of MLS throw-ins, showing that completion and retention rates are very low for throws down the line towards the goal, but increase as the angle opens and pans back towards a player's own goal.
These results are consistent with the idea that the distance thrown, as a continuous variable, seems to help model accuracy, as my current xThrow model tracks its observed values much closer than Eliot's did, and taking throws within 10 seconds of the previous on-ball action, before the defense can get in position, maximizes the chance of throw success. However, one of the seminal findings from the development of g+ here is that retention and completion rates are overrated, at least in the final third, as the outcome of an even improbable event, like a long throw into the box, can be so worthwhile that the completion rate is almost secondary. Consider distance thrown – xThrow says that the odds of success decline the longer you throw it, while xRetain says it doesn't have much effect on keeping possession, but long throws have more value in terms of g+, and it's basically a linear relationship. Incorporating goals added shows more context on throw angle – xThrow and xRetain are at their lowest with shallow, attacking angles, increase as the angle approaches a lateral throw, and plateau once throws pass 90 degrees and start going backwards. Projected g+ also starts at its minimum and increases rapidly, but once it hits a maximum value between 90 and 100 degrees, possession value starts to decline again as the angle continues to increase.
It's essential to acknowledge that these models are blurry pictures of complex dynamics, and event data alone can't capture the spatial context surrounding where a player chooses to send a throw-in or the resulting outcome, only tracing trendlines. Clubs could invest some effort peeling this onion a bit more by developing theories, testing them with film study and scouting, and operationalizing that to new players using tracking data. Finding the right footage to watch might be a challenge, but modeling possession g+ based on some reliable patterns of throw-ins lets us identify players that reliably generate more value than expected. To do this, I used 100 career throw-ins between 2013 and 2025 as a threshold to ensure that any players we find aren't low sample-size flukes.
First, I excluded strikers from these tables because only one of them has made more than 100 throws since the start of 2013, and perhaps I should have also excluded defensive and center midfielders. I chose to rank these tables by per-throw adjusted model overperformance to find players that can get the most value from each throw, and focus on what they do right. Djordje Mihailovic is the league leader in g+ per throw, while noted hurler Michael Boxall scores somewhat lower, and given the recency of Minnesota's aerial assault, perhaps if we segmented by more recent times, we'd see something different. However, this is only one half of the puzzle, as every successful throw-in needs a target, and coaching how teammates should run into space and create opportunities within a few seconds of the ball going out will be fundamental to maximizing team success with these restarts.
Thus, it makes sense to also identify which players maximize the value of throws they receive, and the receiver list is perhaps more satisfying, with many players getting their heads, shoulders, knees, and toes on long throws into the box, and Andrea Pirlo elegantly receiving throws inside and retaining possession to spray it about. So, let's wrap this up by returning to some data-driven strategic suggestions for throw-ins, despite finding these general patterns, it's essential to remember that, just like any shot can go in despite a low xG, or miss despite a high one, any throw-in could eventually result in scoring or conceding. These should be considered general tips for how to get more value from these moments in the long run, not a prescription for every throw-in situation. The first two rules are adapted from Eliot's original findings that I have replicated here, and they are: getting open as soon as possible is more important than getting forward, but if you can do both, that's great; avoiding letting the ball go out in your defensive third, and pressing when it goes out in your opponent's defensive third; prioritizing retaining possession over advancing field position on throws in your half; and looking for open teammates closer to the middle of the pitch, whether you can get it into the penalty area or not.
One final thought: I haven't done a team or coach-level analysis yet, and I'm sure that there are coaches, teams, and players that find success with a "chuck the ball forward" restart mentality, which might be teams that are confident in their ability to win 50-50s and exploit the chaos in dangerous areas that those plays can create. However, I couldn't help but notice that these findings seem to encourage a possession-style that uses space to progress rather than brute force, move the chains, field position ball, which is a fairly unique finding for g+, which absolutely loves field position. I have to assume that backwards passes have value exactly because teams can keep the ball and recirculate it to other parts of the pitch, and I'm frequently shocked when I watch matches at how compressed players get around the ball when it's a throw-in against a set defense. I haven't read ahead or anything, but I'd bet the key lessons to be gleaned from the scouting lists have to do with exploiting space with quick recognition and smart movement by both thrower and target, rather than raw length.
Eliot conducted that research almost a decade ago, and I believed it was time to replicate and expand upon those findings with the impressive infrastructure that ASA has developed since the days of using CSV files on Dropbox. In addition to models that estimate throw completions and retained possession, I also analyzed the goals added for every possession following each throw to assess the value of throw choices. This allowed me to identify the MLS throw-in MVPs and provide an expanded set of general rules for approaching these often-overlooked moments of play. Before delving into the updated data, let's briefly review the first two seasons of Game of Throw-ins. In Part One, Eliot modified ASA's trusty xPass model for throws in two ways: adding a predictor for time since the previous event, known as xThrow, and using retained possession as another measure for a successful throw, referred to as xRetain. It is essential to recall some critical definitions: a successful throw, as measured by xThrow, is a throw that is touched by a teammate first, while a retained throw is a throw that results in the team maintaining possession afterwards. For instance, a player can throw the ball to a teammate's head, and even if the opponent deflects it, the team can still retain possession, or they can throw it off a defender's back and recover the ball themselves, which would be considered an unsuccessful throw but one that still results in retained possession.
After examining these models in detail, Eliot provided three key lessons for coaches to impart to their players. Firstly, it is crucial not to let the ball go out of bounds deep in one's own half if it can be prevented, as the likelihood of retaining possession after a throw-in in the defensive third is essentially a coin toss. Therefore, if there is no significant pressure from the opponent, it is a wise decision to make an effort to keep the ball in bounds. Secondly, when the opponent has a throw-in in their defensive third, it should be used as a pressing trigger, as the probability of a turnover following a throw-in is high, and the team should be prepared to capitalize on this opportunity. Thirdly, it is recommended to take throw-ins 5-10 seconds after the ball goes out of play, allowing the players to quickly get into position to receive the throw-in before the opponents can establish their defensive shape. I would also employ some gamesmanship, particularly in high-stakes matches, by having my players attempt to slow down the opponent's throw-in to give my defense time to organize.
One of the interesting developments when revisiting the original research is that the second point has indeed become a common tactic, with many teams now using throw-ins as a pressing trigger to start the game. In Part Two, Eliot grouped MLS teams based on their throw-in direction choices and evaluated teams and players on their retention success compared to model expectations. He discovered that while most teams lacked a defined style, some were clearly coached to throw the ball forward, while others were instructed to spread the ball around and often throw it backwards. Although the results were somewhat noisy, I noticed that two teams and their fullbacks had remarkable success retaining throws by spreading the ball around and throwing it backwards, and they were coached by Gregg Berhalter and Peter Vermes, who at the time favored possession-focused styles. More than anything, Part Two suggests that few coaches were paying particular attention to throws, and that throwing the ball backwards seems to be helpful for retaining possession, especially for teams that prioritize possession.
However, retaining possession is only one aspect of the game, and having the goals added for every possession following each throw allows us to push this analysis further and better understand how these decisions affect the immediate odds of scoring. Do throws down the line compensate for their low success odds with increased goal danger? Can teams reduce their risk of turnovers while also generating attacking value? Which are the best throw choices in different areas of the pitch? As I was working on the models, I realized that I was struggling to think through all the possible throws available, considering the various contexts in which they can be taken. For example, a throw that a player receives right on the half line could have been taken 30 yards behind them, 30 yards in front, or anywhere in between, and this context drastically changes the nature of the throw. Similarly, the distribution of throw targets in the attacking third is likely different from that in the defensive third. How can I visualize this complexity when there are so many overlapping patterns? The only solution is to make it interactive.
Before proceeding, take a moment to explore the interactive visualization, clicking around the touchlines, examining the different metrics, and noting the patterns of where throws go and how they tend to end up. Both sidelines exhibit roughly the same patterns, so we can focus on what happens when we move up and down the touchline. Taking a throw from one's own defensive corners is a rare but perilous situation, with no reliably good options, as retention rates and possession values are among the worst in the data. The next few origin zones start to reveal some patterns, with the most popular target zones being forward towards the goal, but these have low retention rates and possession values. Backwards passes begin to show favorable retention rates, and throws toward a team's own goal actually have the highest possession values by g+, which is an interesting finding that somewhat contradicts most traditional g+ evaluations, which often prioritize large field positions over everything.
As you approach midfield, these patterns evolve and solidify, with the overall spread of where throw-ins go growing, and it becoming increasingly clear that forward throws in the defensive half are not worth much, consistently having lower retention rates and possession values, while backwards throws yield better retention and g+ value. Forward throws just short of the midfield line are especially suspect, but possession values start to improve as these throws start to go past midfield. Zones with the highest g+ values are often the furthest, usually backwards but also in lateral and attacking directions, indicating that long throws, likely to teammates in space, are the highest upside choices. This makes sense, given that opponents are likely to condense the area as much as possible. The upside of trying to connect passes inside is a fascinating finding, as it mostly goes counter to many years of coaching orthodoxy, suggesting that, in one's own half, throwing the ball up the line, gaining yards, and then building from there can be a valuable strategy.
Once the ball thrower is in the attacking half, their options begin to open up even more, with putting the ball close to the box immediately becoming a high-value option, and new gaps near the goal line opening up as well. The retention rates of forward passes improve, and they actually surpass the completion rates when thrown into the box, but the risks also seem to increase, as this appears to be where most of the negative g+ average tiles appear, intermixed with highly valuable zones, likely driven by counter-attacks on unbalanced teams or juicy turnovers trying to play back to the middle of the field. Backwards passes lose some of their reliability, but there are pockets where they still generate dangerous possession when throws have reached the final third. Short throws also become viable, as while they are often pressing triggers for the opposing team when taken in the defensive half, the reduced distance to the goal may force defenders into withdrawn positions, creating more space. And, in an article about throw-ins, the line you all came for, long throws into the mixer remain very worthwhile.
In general, this tool demonstrates how the geography of the pitch is foundational for optimizing throw-ins, with team strategy influencing how many throws are taken and where, and players better understanding which areas to look and move to when putting the ball back into play. Unless your team is built to win 50-50 challenges and repress when you lose them, crowding everyone forward is a bad strategy, unless you can reach the box, in which case crowding everyone forward is a good idea. I follow Eliot's modeling framework but with a few tweaks of my own, using possessions as defined for goals added, which differs from Cheuk Hei Ho's possession definitions from his expected possession goals work. The results are consistent with both the old model results and the interactive visualization, which draws from 13 seasons of MLS throw-ins, showing that completion and retention rates are very low for throws down the line towards the goal, but increase as the angle opens and pans back towards a player's own goal.
These results are consistent with the idea that the distance thrown, as a continuous variable, seems to help model accuracy, as my current xThrow model tracks its observed values much closer than Eliot's did, and taking throws within 10 seconds of the previous on-ball action, before the defense can get in position, maximizes the chance of throw success. However, one of the seminal findings from the development of g+ here is that retention and completion rates are overrated, at least in the final third, as the outcome of an even improbable event, like a long throw into the box, can be so worthwhile that the completion rate is almost secondary. Consider distance thrown – xThrow says that the odds of success decline the longer you throw it, while xRetain says it doesn't have much effect on keeping possession, but long throws have more value in terms of g+, and it's basically a linear relationship. Incorporating goals added shows more context on throw angle – xThrow and xRetain are at their lowest with shallow, attacking angles, increase as the angle approaches a lateral throw, and plateau once throws pass 90 degrees and start going backwards. Projected g+ also starts at its minimum and increases rapidly, but once it hits a maximum value between 90 and 100 degrees, possession value starts to decline again as the angle continues to increase.
It's essential to acknowledge that these models are blurry pictures of complex dynamics, and event data alone can't capture the spatial context surrounding where a player chooses to send a throw-in or the resulting outcome, only tracing trendlines. Clubs could invest some effort peeling this onion a bit more by developing theories, testing them with film study and scouting, and operationalizing that to new players using tracking data. Finding the right footage to watch might be a challenge, but modeling possession g+ based on some reliable patterns of throw-ins lets us identify players that reliably generate more value than expected. To do this, I used 100 career throw-ins between 2013 and 2025 as a threshold to ensure that any players we find aren't low sample-size flukes.
First, I excluded strikers from these tables because only one of them has made more than 100 throws since the start of 2013, and perhaps I should have also excluded defensive and center midfielders. I chose to rank these tables by per-throw adjusted model overperformance to find players that can get the most value from each throw, and focus on what they do right. Djordje Mihailovic is the league leader in g+ per throw, while noted hurler Michael Boxall scores somewhat lower, and given the recency of Minnesota's aerial assault, perhaps if we segmented by more recent times, we'd see something different. However, this is only one half of the puzzle, as every successful throw-in needs a target, and coaching how teammates should run into space and create opportunities within a few seconds of the ball going out will be fundamental to maximizing team success with these restarts.
Thus, it makes sense to also identify which players maximize the value of throws they receive, and the receiver list is perhaps more satisfying, with many players getting their heads, shoulders, knees, and toes on long throws into the box, and Andrea Pirlo elegantly receiving throws inside and retaining possession to spray it about. So, let's wrap this up by returning to some data-driven strategic suggestions for throw-ins, despite finding these general patterns, it's essential to remember that, just like any shot can go in despite a low xG, or miss despite a high one, any throw-in could eventually result in scoring or conceding. These should be considered general tips for how to get more value from these moments in the long run, not a prescription for every throw-in situation. The first two rules are adapted from Eliot's original findings that I have replicated here, and they are: getting open as soon as possible is more important than getting forward, but if you can do both, that's great; avoiding letting the ball go out in your defensive third, and pressing when it goes out in your opponent's defensive third; prioritizing retaining possession over advancing field position on throws in your half; and looking for open teammates closer to the middle of the pitch, whether you can get it into the penalty area or not.
One final thought: I haven't done a team or coach-level analysis yet, and I'm sure that there are coaches, teams, and players that find success with a "chuck the ball forward" restart mentality, which might be teams that are confident in their ability to win 50-50s and exploit the chaos in dangerous areas that those plays can create. However, I couldn't help but notice that these findings seem to encourage a possession-style that uses space to progress rather than brute force, move the chains, field position ball, which is a fairly unique finding for g+, which absolutely loves field position. I have to assume that backwards passes have value exactly because teams can keep the ball and recirculate it to other parts of the pitch, and I'm frequently shocked when I watch matches at how compressed players get around the ball when it's a throw-in against a set defense. I haven't read ahead or anything, but I'd bet the key lessons to be gleaned from the scouting lists have to do with exploiting space with quick recognition and smart movement by both thrower and target, rather than raw length.
#news, #football