Goals Added and The Great Possession Shift
/This is part three of our series on Goals Added (g+). Here is part one, where John first introduced it and here is part two, where Matthias broke our brains.
Since data analytics tools first began to be applied to soccer, the field has moved in waves. In the beginning, there was a focus on goals. Then attention shifted to look at other things we can count, like shots, passes, tackles, and all the other events we know and love. Eventually, there was a coalescing around possession percentages, pass completion, and other second-order statistics. From there, everyone went all-in on shots. Shots give us goals, and goals are so rare that it’s difficult enough to learn much from them alone, that looking at shots makes sense.
Adapting metrics created for hockey, soccer analysts first introduced total shot ratio and PDO. Eventually, more probabilistic things like shot-based expected goals and expected assists were introduced. This was a significant step, signaling a shift from valuing not only where the shot came from, but also the context around buildup and where the ball and sequence ended up.
More recently, people started asking about things that aren’t shots. Aren’t passes important too? What about dribbles? We started to see ball progression metrics, final third entries, penalty box entries, and eventually non-shot expected goals. We saw things that measure getting into dangerous areas and why some passes are more dangerous than others.
While it’s insightful to look at how many and how valuable the shots a team generates and concedes are, that only encapsulates two-thirds of the pitch. Although ball progression and non-shot xG give you insight into dangerous ball movement, sometimes the goal of a possession isn’t to score, it’s to not concede. This is where today’s larger “value-type” models come into play. How can we evaluate actions that indirectly lead to (or prevent) shots and goals? Sarah Rudd may have been the first person looking at this, almost a decade ago with some Markov Chain based work. There have been many more analyses looking at this recently, including but absolutely not limited to KU Leuven’s Valuing Actions by Estimated Probabilities (VAEP), Karun Singh’s Expected Threat (xT), Opta’s Possession Value (PV+), our own Dave Laidig’s Points Above Replacement (PAR) and Cheuk Hei Ho’s Expected Possession Goals (xPG), as well as Luke Bornn and Javier Fernandez’s Expected Possession Value (EPV). Each of these attempted to take everything we know about each event and model it into something that tells us how valuable it was towards scoring (or not conceding). With our new goals added (g+) model, we attempt to address a lot of the same problems, and it’s important to contextualize g+with the existing work, and what distinguishes it from the rest.
What Is Goals Added (g+)?
If you want the methodology, read Matthias’ post from earlier this week. If you don’t, here’s resident funny man and renowned Ronny Deila enthusiast, John Muller, to explain:
The “…model that values any touch of the ball anywhere on the field in terms of goals. Not just how likely a pass is to lead to a goal, although it does do that. Not just how likely a tackle is to prevent a goal. For every single on-ball action—a header won, a dribble lost, an awful corner kick that sails over everyone’s head, you name it—the model digs through a bunch of contextual data and calculates how much the play improved that possession’s probability of ending in a goal and, just as importantly, how much it reduced the other team’s chance of scoring on the next possession. Which means we can now compare all kinds of plays with the same unit of account: their likely effect on the scoreline.
Why Do We Think Our Model Is Good, and How Does it Compare With Others?
It’s more than just shots
Right off the bat, one of the nice advantages of Goals Added is that it doesn’t look just at shots, it looks at everything that happens on the ball. Passes, crosses, dribbles, shots, carries, you name it. Everything is evaluated in terms of how it alters the probability of scoring during this possession before the action (the pre-value), and after the action (the goals added).
Take this pass from Marky Delgado to Jacob Shaffelburg. Delgado receives the ball in a position with a prior possession context, meaning the Toronto possession from this point has a 1.7% chance of scoring, or 0.4% chance of them conceding. He then successfully plays a clever through pass for Shaffelburg into a situation with 13.9% chance of scoring. +0.122 “goals” have been added to the likelihood of scoring in this attack, so the pass itself scores +0.122 g+. We’ll touch upon how that +0.122 is divided up between Delgado and Shaffelburg a bit later, but immediately, we can see the advantages compared to shot-based xG, which would have only assigned value to the final action in the sequence - the shot. While we eventually get to the same place, a blocked pass across goal which would result in a high xG shot, how that portion is divided up is a lot different than xG to Altidore on a would-be tap in and xA to Shaffelburg on a would be square.
It treats different kinds of touches differently
While non-shot xG offers similar insight on the value of moving the ball into different zones, it can be agnostic to how it moves. There are many things we can infer from how a ball is moved: a carry that moves five yards upfield is often unopposed, while a dribble or take on in the same location removes a defender from the play. A pass may or may not split a defender along the way. Each different touch has implications on how the possession evolves, even if the ball starts and ends in the same two locations. Similarly, the velocity of the possession (both vertically towards goal and horizontally across the pitch) says a lot about how dangerous it actually is. Lastly, what happened previously? A pass into the striker after a long spell of possession may be less valuable against a set defense than the exact same pass in a counter attacking opportunity.
This contextual information is reflected in the model, which examines the prior actions in the possession and adjusts the goal probability of the situation and the g+ of the action accordingly. While in the short term these differences may appear small, extrapolating this out over a whole dataset allows us to find obvious value differences between g+ and something like non-shot xG. Similarly, we don’t throw away the shot value, it’s still apportioned, just differently.
It looks at both sides of the ball
Models like xT, created by Karun Singh, give us a really nice insight into how we proportion a lot of the credit from that movement, even if a given action doesn’t directly lead to a shot. While xT, breaks the pitch down into a Markov model with movement between each zone, by looking at all successful moves (that end in a shot) we are able to see how valuable actions that move from one zone to another are, outside of the lens of just xG. But xT only considers the attacking side of the ball. While some say that the ideal purpose of every possession is to score, sometimes soccer is messy and that’s just not true. Just ask any supporter who has cursed at the television while their team clings to tenuous lead late in stoppage time and the striker takes a shot instead of dribbling to the corner to waste time. Other times the most beneficial play for a team is to smash it out of the penalty area. While not retaining possession is bad for their chances of scoring a goal, it is really good for reducing their chances of conceding.
This is one of the great benefits of our model, namely that every action is evaluated based on the change in probability of a team’s possession resulting in a goal, but also the change in probability of their opponent's next possession resulting in a goal. We saw similar ideas in our own xPG, but that only considers the chance of the turnover immediately becoming a shot, rather than the entire possession chain. We also see this in Opta’s PV+, VAEP, and EPV, and it allows us to actually evaluate the cost and benefit of defensive actions beyond just counting them. As we know, bad teams usually rack up clearances and tackles and blocks purely from being bad and having to do lots of defending. With our model, we can still see the value gain from those defensive actions, but it’s very much tempered by the fact that they are also huge negatives involved in not retaining possession. We can also properly evaluate defensive midfielders who do lots of dirty work and ball progression but are traditionally undervalued in shot-based xG models.
Holistic Models
So that leaves ASA’s g+, Bornn et al.’s EPV, Jan van Haaren’s VAEP, and Opta’s PV+ as models that examine the entirety of event data, consider the context of each event, and consider both sides of the ball. While all attempt to do similar things, there are crucial differences to note before we continue. EPV is based on a combination of tracking data and event data, meaning not only can you evaluate the possession value added by each action, but you can evaluate the decision making of each player as you know the true options available. The great tragedy of EPV is that most of the inputs involved are proprietary (i.e. they’re not available to plebs like us), and may not even exist for many leagues or teams. While we all hope to have the same resources as Barcelona and Liverpool one day, we’ll keep it to the world of public-ish analytics for now.
So how is g+ different from VAEP or PV+? It’s a good question, because they all talk about value and possessions and probabilities and actions and honestly all sound quite similar.
First off, a small disclaimer: (1) while we compare to PV+ and VAEP, kudos to those guys for putting it out (mostly) publicly and driving discussion forward in this space, (2) these are comparisons to the specific iterations of their models in the text released. Much like g+, these things are in constant development, growing and working to describe the game as accurately as possible.
Possession Context
When attempting to value actions, the context in which it takes place is extremely important. We already see that the information gain we see from these possession models in comparison to context-less models is huge. PV+, for example, takes some context indicators from the last 5 actions in the possession chain, while VAEP takes the last 3, and g+ takes into account the specific last action, with some modifiers for features of the whole possession. Similarly, the actual contextual clues the model considers influences how action values are determined. It is unclear what context flags PV+ uses, while VAEP uses the following complex features and game context features: (1) distance and angle to goal from start and end location of the action, (2) the distance covered in the x and y direction during the action, (3) the number of goals scored by the possessing team after the action, (4) the number of goals scored by the defending team after the action, (5) the goal difference after the action. Goals added considers similar contextual clues (you can find a comprehensive list here,) with both models using the distance covered during the action and prior actions to give some idea about the “velocity” of a possession both laterally across the pitch and vertically towards goal. Goals Added (g+) only considers the action immediately prior to the action currently being valued, but with some greater context about the possession as a whole. The model includes flags for the total number of actions in the possession thus far, with indicators for the number of actions since the last dead ball, and what said dead ball was (goal kick, throw-in, corner etc.).
Receptions
One of the key differences between g+ and other similar frameworks is the use of receptions. When a player makes a pass from A to B, how do we proportion the credit to the passer and the receiver? If a center back passes to his partner, did he do anything of value to receive the pass? Probably not, so assuming the receiver gets 0 pass value might feel fair. But what if your attacking midfielder plays a slick through ball for a striker, making a perfectly timed run off the back shoulder of the defender? It becomes a lot harder to convince yourself the passer deserves all the credit in that situation. Under VAEP and PV+, the passer gets all the credit for that pass, even though the clever run of the attacker is arguably as important as the pass itself. Opta has looked at the PV+ of passes received as a measure of strikers, given they don’t often participate in earlier phases of the possession as much, but they aren’t explicitly dividing the credit up between passer and receiver.
Editors Note: Shortly prior to publication of the Goals Added methodology there was an adjustment to SPADL and VAEP to include receiving value with Atomic-SPADL.
In our g+ model, pass value-added is split between the passer and receiver, according to our xPass% model. xPass% allows us to estimate the probability of a pass being successful based on some contextual factors and the start and end locations, allowing us to evaluate passes. If a pass has an xPass% of 20%, the passer gets 20% of the credit, the receiver gets 80% of the credit. Let’s go back to the Delgado-Shaffelburg example from before. xPass% tells us the average passer completes that pass approximately 72% of the time and g+ tells us it added 12.2% goal probability to the possession (and is worth +0.122), so Delgado gets +0.088 (0.72*0.122) worth of pass value and Shaffelburg gets +0.034 (0.28*0.122) worth of receiving value.
On very high completion rate passes the receiver gets very little credit because they didn’t really do anything. The passer gets all the (minimal) credit because they didn’t mess it up. But players who are very good progressive pass receivers, receive many long passes from the goalkeeper, or receive dangerous passes in behind the defense get a big chunk of credit for getting open enough to receive and maintain possession. On the face of it, one potential limitation is backwards passes; while some may have intrinsic positive values as they open the pitch up more, they are often negatively valued. This way a player can play backwards and lose some of the value of the possession but not all, and players receiving backwards passes aren’t being dinged as they generally have a quite high completion rate.
Turnovers
Just as g+ credits offensive players for completing difficult passes, it also factors in the cost of those same passes getting intercepted. When we look at a turnover in PV+ and VAEP, we see that a player loses the value of their possession, and loses the value of an opposition possession beginning from that context. To use the example Opta used in their roll-out announcement, a possession begins inside the possessing team’s penalty area and has a 1% chance of resulting in a goal for the possessing team. The defender attempts to pass out and it is intercepted, realizing a 14% chance of the opposition scoring. This would be a -0.15 PV+ (-0.01+(-0.14)) turnover. However, Opta caps these turnover values at -2.5% (the average value of a possession) because fully attributing turnovers to the passer really penalizes high risk/high reward creative attackers. As they bear a lot of the creative load, they rack up negative value turnovers very quickly, and any model that tells you your best attackers are actually the worst is probably missing something.
We combat this in two ways: (1) scaling passing by xPass% allows us to also scale turnovers by xPass%, meaning difficult passes are penalized less for not coming off. By scaling by xPass% we can see an incomplete pass from a 5% area, generating a 3% opposition attack, with a 25% xPass%, as only a -0.02 g+ hit instead of a -0.08 g+ hit.
This really helps out attackers who try to unlock defenses, and it doesn’t lead to g+ inflation because they see the same mediation on successful passes. But if a player fails a very easy (high xPass%) pass that gives a dangerous counter, it’s a big hit.
(2) Turnovers look across three possessions. Sometimes a failed pass is attempted with the thought that even if it doesn’t come off, the team has good structure to take advantage of the opposition's next possession, whether that is through a counter-press or allowing them to set up appropriately or just gaining field position. We’ve even seen this with teams choosing to play kickoffs out of play near the corner flag, allowing them to set up a press effectively or by playing balls into the channels to chase and move up the field. By looking one possession further ahead, the turnover calculus is as follows: an incomplete pass from a 5% goal probability attack, generating a 3% opposition attack, with a 25% xPass%, will on average have a next possession value of 2%. The pass value is now (-0.05-0.03+0.02)*0.25 or -0.015 g+ instead of the -0.02 g+ it was before. Again, it’s a small tweak, but it makes a lot of difference in evaluating those high value attackers.
It is important to recognize that the initial possession value is the value of the whole possession pre-action. More simply, the possessing team’s probability of scoring minus the defending team’s probability of scoring. The Intercepting team’s possession value is the realized possession value after the turnover. The third term is the value of the next possession after this. Turnovers, as such, look over three possessions, while all other actions look over two.
Taking both of these to the end of the same Delgado-Shaffelburg clip above, Shaffelburg fails on the square to Altidore for a simple finish in a very high value scenario (12.6%). However because the xPass% probability is so low, the opposition possession value is low, and your next possession will likely start quite high up the pitch, Shaffelburg only sees a -0.013 g+ on the pass.
Another important area to think about is how these models evaluate shots. While they do look at so much more, it’s important to get the value of shots right. PV+ awards the xG value of the shot, while VAEP awards the difference between the xG value of the shot and the value of the goal (1-xG) as a look at finishing skill.
g+ approaches the value of shots a little bit differently. While getting in good positions to take shots is inherently valuable, a lot of that value comes with receiving the ball (and g+) in good areas. We assess the shooting value added as the difference between the xG of the shot, and the possession value from that location. This lets us do a few things: first, is shooting actually the best decision? If you are in a great assist zone (with high possession value but low xG) and choose to shoot, that’s a stupid thing to do! You lose g+. Second, what work did the player do between receiving and shooting? You get that value too, but as dribbling. This diminishes the actual value of shooting in comparison to being able to generate the shot, but there are some shooting tweaks, too. Instead of using the standard shot-based xG for the shooting g+ calculation, we use a modified xG which includes the likelihood to generate rebounds based on pre and post shot information.
Goals Added and Player Comparisons
The final really interesting thing to think about is how values work in relation to an average or replacement level player. If we think about WAR in baseball or basketball, it gives us an idea of A) where the very tip top of players are and B) how good someone is compared to a player fresh out of college. In MLS, this is a nice comparison given we have the SuperDraft, and well, lots of kids fresh out of college and newly professional. While VAEP and PV+ give you broad general ideas, Goals Added Above Average (g+avg) and Goals Added Above Replacement (g+r) allow us to make tight comparisons to a real roster filling method for MLS clubs. Traditionally, you would use some percentile guideline (say 5th percentile) to define replacement level, and take the 5th percentile of goals added. However, 5th percentile goals added players really don’t play very much because soccer has so few substitutions (probably less than 1% of total minutes). As such, if we define the replacement level by the 5th percentile of the total minutes played, we can see what a replacement level player (who actually plays) looks like. This ends up giving a replacement level (with some variance position to position) around the 20th percentile of goals added.
By doing this for each positional group we can eliminate any artifacts of comparing between positions (like holding center backs to league wide passing average/replacement levels, instead of just other center backs) and see how players actually stack up against their league-wide peers. Similarly, players have their average and replacement levels determined by where they play their minutes. An attacker who spends some time as a winger and some time as a striker will see their replacement and average level as the minutes weighted average of both positional baselines league wide. In theory, this could be broadly applied to multiple leagues to give you a more accurate baseline of what is replacement level. For example, we could see how a replacement level MLS player compares to a replacement level player in the Bundesliga II, though the answer would be a sad one for Terrence Boyd.
What’s the Gist of All of This?
I think it’s important to look at how this fits into the timeline of soccer analytics. We went from counting goals and passes to more holistic and probabilistic non-shot and possession value models. But even the best existing models are just missing some important features that mesh with how we think about and watch the game, and what we know about what makes the best players the best, and why.
Goals Added (g+) allows us to examine the value added by players in all phases and events of the game in the context of not only scoring on their possession, but also preventing the opposition from scoring (and in the case of turnovers, getting the ball back) through the most important measurement unit in soccer, goals. While there have been many attempts to model the entirety of event data, g+ allows us a unique insight into what valuable things players do along the way during possessions, both for and against. By comparing it to a real, identifiable benchmark through goals added above average (g+avg) and goals added above replacement (g+r) it gives us an opportunity to examine a whole cadre of interesting things in the MLS sphere (and you’ll be seeing some of them in the coming days on ASA). It’s still a work in progress, but it’s a really exciting start.