Goals Aren’t Created Equal. Just Ask These Two.
Before we really get into this piece, have a go at answering this question: Which was Manchester United’s most memorable goal of the 2020/21 Premier League season?
Maybe it was one of those Pogba stunners. Maybe it was one of those Cavani tap-ins that came from his oh-so-satisfying off-ball runs (or his chip at Old Trafford). Maybe it one of Maguire’s two headed goals, or maybe, just maybe — for the rival fan tears — it was one of Bruno Fernandes’ penalties.
Whatever you went with, it’s highly unlikely you picked Mason Greenwood’s second goal against Burnley from April this year:
So why have I focused on this goal? United aren’t playing a big rival, the goal isn’t particularly great (there’s a fairly large chunk of luck involved, with the deflection) and there isn’t even a crowd reaction or anything out of the ordinary to make this stick in the mind.
Simply put, this goal was the most valuable one of Manchester United’s league season. Why and how? Read on.
Let’s look at this intuitively first. Can we think of examples where a certain goal would be considered more important than another one? The simple answer is yes. Obviously, if you score when it’s 0–0 and put your team in the lead, thereby tilting the scales in your favour, it’s a lot more useful than scoring when you’re already 7–0 down, and you just really want to go home by that point.
This isn’t anything that anyone who follows football won’t be able to tell you. But let’s ramp it up. Let’s say your team is 2–0 down, and you score the goal that makes it 2–1. Later on a teammate of yours makes it 2–2, and completes the comeback. Which goal is more valuable? Maybe you think about this for a few seconds, but you end up landing, pretty firmly, on the side of the second goal. After all, it’s the one that secured the result for your team, right?
Let’s take it up one last notch. Is it more beneficial to score if your team has been dominating the shot count throughout, or if you’ve been under the pump and you’ve managed to nick one against the run of play? Bit tougher to respond intuitively, isn’t it?
We’ll go through the process together, and hopefully, by the end of this piece, you get a good idea of how we can answer these questions.
Problem Number One: How do we measure the value of a goal? What’s our currency?
So, this little issue is solved with the help of our good friend xG. Using expected goal values for individual shots, we can simulate entire matches and look at the result probabilities, instead of just going on the result itself. For example, maybe a team won 1–0, but the xG was 0.45 to 2.21 in favour of the other team (looking at you, Brighton). In such a case, on simulating the game virtually, you might get a result that says the actual winning team would only win that game 4.3% of the time (let’s say), essentially making what happened a freak event.
For more on how the process actually happens, visit this page , run by FC_rstats.
Now, from result probabilities, we can extract a statistic called expected points (xPts), by weighting the probabilities by the number of points obtained for each result. Here’s an example:
Earlier this season, when Manchester United visited the Etihad Stadium to take on Man City, the shot count was 23–8 in City’s favour and the xG (according to Understat) was 1.28 for City and 1.35 (plus one penalty, making it 2.11) to United. [Highlights here]
Simulating this match 10,000 times using the xG data, the probabilities of the following results were generated as:
Man Utd win: 59.64 %
Draw: 22.27 %
Man City win: 18.09 %
To generate expected points for each side, all we have to do is multiply their respective win percentage by three, add it to the draw percentage multiplied by one and the loss percentage multiplied by zero (which, obviously, will always be zero).
So, for United, that comes out to: (3 * 59.64%)+(1 * 22.27%) = 2.01 xPts
And for City, (3 * 18.09%)+(1 * 22.27%) = 0.76 xPts
Essentially, if this event (the match) had taken place enough times for it to strictly follow the distribution of the xG, United would have averaged 2.01 points per game across all the games, and City 0.76. This particular instance saw United get 3 points, and City none, as they won 2–0, with goals from Bruno Fernandes and Luke Shaw.
[If you have studied statistics at any level, this will be familiar to you as the calculation of the expectation (or mean) of an outcome.]
Problem Number Two: Okay, now we know the currency. How do we assign value to a goal?
This process is extremely simple. We basically split the match into two distinct parts; pre- and post-the goal in question. We can obtain expected points values for the teams involved, using the shots and goals that have already happened. We assign goals an xG value of 1 in this process, because in essence, we want the simulation to know that every time it iterates, that shot has a 100% chance of hitting the net.
Then, we run the simulation again, adding the goal in to our list of values, (again, assigning it an xG value of 1, as we want the simulation to know what we know; that that shot has already been scored).
Again, we look at the expected points and note the change in the values. This is how we assign value to a goal. Allow me to make this more clear with an example. Let’s take Greenwood’s goal, which I referenced earlier.
The goal was scored in the 83rd minute of the game. What had happened until that point? Well, Man United had taken 12 shots, scoring with one (Greenwood scored the opener). At the same point in time, Burnley had taken 10 shots, also scoring with one, as James Tarkowski headed home from a corner to bring them level. [Highlights here]
Now freeze.
Looks like a pretty even game, right? It is, according to the 10,000 simulations run. Let’s look at the result probabilities at that point:
Man Utd win: 28.08%
Draw: 45.15%
Burnley win: 26.77%
xPts for Manchester United: 1.29
Basically, if the game had stopped before the goal in question here, the most likely outcome would have been a draw, which was also what was happening then, with the score at 1–1.
But with the next shot in the game, Greenwood gets a lucky deflection and the ball goes in. Goal, 2–1 to United. Let’s add this goal into the simulation and look at our new result probabilities.
Man Utd win: 72.91%
Draw: 22.85%
Burnley win: 4.24%
xPts for Manchester United: 2.42
Do you see the difference that goal made? A game which was pretty close until that point, suddenly became loaded in United’s favour with that goal. So, for this goal, we would assign it a value of (2.42–1.29)=1.13 xPts.
Problem Number Three: Aren’t you missing something, Rahul?
Why, yes. Yes I am.
Let’s think about the previous example for a moment. Greenwood scored that goal in the 83rd minute, to put United 2–1 up. Now, would that goal be as important if it had happened in the 60th minute? Or, would it not have been more important if he’d scored it in the 91st minute?
The way we’re measuring goal value at the moment, there’s no room to allow for timing. If Greenwood scores that goal 23 minutes earlier, that’s 23 minutes more that Burnley have available to score an equaliser. If he scores 8 minutes later, that’s 8 minutes less that Burnley have to get an equaliser. The two situations aren’t equal, and here comes the last part of our puzzle.
We apply a simple adjustment factor for time, by multiplying the xPts difference by the minute in which the goal was scored (limiting the value to 90), divided by 90.
So, the actual value we would assign to Greenwood’s goal would be:
1.13 * (83/90) = 1.04 xPts
If he’d scored in the 60th minute:
1.13 * (60/90) = 0.75 xPts
And if he’d scored in the 90th minute+:
1.13 * (90/90) = 1.13 xPts
Conveniently, this adjustment helps us work around another potential issue. When a team scores with the first shot of a particular game, what is the goal value, in terms of expected points?
Before the shot was taken, there had been zero shots, which means the only possible result was a draw (xPts = 1). After the shot, the only possible result is a win for the scoring team (xPts = 3), because effectively, there’s been only one shot in the whole game.
If this situation plays out in the first minute of the game, the value, without time-adjustment, would come out to 2.00 xPts, the maximum possible value for any goal. But that doesn’t make sense, does it? Surely, just scoring with the first shot of the game in the first minute can’t be the most valuable goal possible?
Applying the time adjustment in such a scenario gives us a final goal value of 0.022, which is far more in line with what common sense tells us.
Putting the Concept into Practice:
So, now that we’ve got all the boring stuff out of the way, let’s get back to our favourite use of statistics: r̶u̶n̶n̶i̶n̶g̶ ̶p̶r̶o̶p̶a̶g̶a̶n̶d̶a̶ collecting valuable insights.
Here is a look at the Manchester United squad’s goal tallies and the value of the goals they scored, both as a total and as an average per goal. Before anything else, have a look at the second column.
Only one player added more value with their goals than Mason Greenwood did; Bruno Fernandes. The two are astronomically apart in terms of the number of goals they scored (7 vs 18), but over the course of the season, Fernandes’ goals were barely 50% more valuable than Greenwood’s.
Here’s a list of some of Bruno’s goals:
Making it 7–0 vs Southampton (minute 86)
Making it 3–0 vs Leeds United (minute 19)
Making it 6–1 vs Leeds United (minute 69)
Making it 1–0 vs Tottenham (minute 1)
Making it 1–0 vs Man City (minute 1)
These 5 goals were all worth less than 0.1 xPts, when factoring in game state and timing, our two parameters for valuing goals.
Here’s a list of all 7 of Greenwood’s goals:
Making it 2–1 vs Burnley (minute 83)
Making it 2–1 vs Brighton (minute 82)
Making it 2–1 vs Aston Villa (minute 55)
Making it 1–0 vs Burnley (minute 47)
Making it 2–1 vs West Ham (minute 67)
Making it 3–1 vs Tottenham (minute 95)
Making it 1–1 vs Leicester City (minute 14)
Every single one of Greenwood’s goals is worth at least 0.15 xPts, with 5 being game-leading goals and one equaliser. That, it must be said, is pretty impressive stuff, particularly for a 19 year-old.
Now, the purpose of this comparison is not to say that Greenwood is better than Fernandes, or not even to say that he is more important to the team. Indeed, if we just consider Fernandes’ 7 most valuable goals, he ends up coming out slightly ahead of Greenwood.
Fernandes plays more, and scores more, and consequently has more ‘junk’ goals, which affect his overall tally. Plus, a player who gives a team 18 league goals in a season cannot be considered anything but valuable.
Rather, the question here is the why, at least with respect to Greenwood. Part of the reason for him almost exclusively scoring high-value goals is that he often comes on as a substitute, so he is relatively fresh. Also, as we’ve seen, late goals are intrinsically more valuable than early ones.
Even if we try to cut out some of the noise, and just compare Greenwood to players who have scored the same number of goals as him, he still comes out head and shoulders above the rest. Admittedly, the sample size is still small enough that this could simply be a coincidence, but could there be more to it?
In this list of Manchester United’s top 10 most valuable goals this season, Greenwood appears 3 times, including twice in the top 5. No one else is on it more than once. (Data isn’t available for his time, but something tells me Ole Gunnar Solskjaer’s goal value numbers would’ve been right up there with the best).
Is there something mental, or biological, that causes certain players to experience an uptick in performance when the ‘time is right’? Is this a reliable measure of ‘clutchness’? Can opposition teams work with this information, and concentrate their defensive attention on players who consistently score high by this metric?
Is it possible, that, like finishing, and movement, value addition is also a skill? And if so, does it hold the key to solving the dilemma of the late substitution? This is a topic that has long been considered to be more or less esoteric in nature; managers having a ‘sixth sense’, ‘smelling the right time’, and so on and so forth. But can we crack it using data? That, perhaps, is the question.
Of course, the applications here are not limited to just examining goalscorers. Creators, high-value passers and potentially even goalkeepers could be measured by this yardstick in one way or another.
The Drawbacks
Now, no model is perfect, and this one is no different.
The primary issue we have here is in the evaluation of game state, using the xG. Since xG is a shot-based metric, we can only evaluate game state on the basis of that singular event, while in reality, the balance/equilibrium of a game depends on a heck of a lot more, simply because there’s so much other stuff going on.
The incorporation of non-shot and post-shot xG would also improve this, as well as the potential use of territory mapping, with the aid of tracking data, which has yet to really enter the mainstream. These would help us take more accurate ‘screenshots’, if you will, of the game states during our evaluation of goal value.
It must also be kept in mind that this method of goal valuation considers the goal only in the context of the game in which it takes place. This is not something that can tell us if a player is consistently performing in finals, derbies and other important matches. That would require the addition of match context in the scheme of a season.
All in all though, this is hopefully a step forward into a relatively unexplored area of data analysis in football, and the landscape looks promising.
