This week we are taking a closer look at some excellent work being done on possession value, and the recent development of a model that has the potential to be the “next xG”, Goals Added (G+).
When they first published G+ earlier this year, American Soccer Analysis (ASA) also shared a wealth of information to help understand the model, how it compares to its competition, and its application to football in the real world. I will give a brief overview, but if you’d like more information, I’d recommend checking out everything that ASA have written about the model, which has been linked throughout this article.
Introducing Goals Added
Goals Added is a model developed by American Soccer Analysis that attempts to quantify a player’s contributions to their team’s performance. G+ uses on-ball actions, both offensively and defensively, and measures the impact each touch has on the chance of their team scoring or conceding, over the following two possessions.
Goals Added uses data about the ball location and movement, and preceding actions, and computes how the probability of scoring or conceding changes as result of a player’s action, using historical data as a comparison.
The Methodology Behind G+
For a more detailed discussion of the methodology behind G+, I’d recommend taking a look at ASA’s Methodological Deep Dive. However, the following is their summary of the method:
1. Create distinct “plays” referred to as possessions or chains.
2. Determine the expected goals earned over the rest of the possession at any stage of that possession.
3. Use statistical modeling techniques to estimate that rest-of-possession value in #2 above, based on details of the “game situation.” A game situation is defined based on where the ball is, which team has it, and how it got there. We use that model to estimate the value of each possession, as of just before every action.
4. Take the xG earned for the rest of this possession and subtract the xG earned for the opposition on the subsequent possession. Thus our possession value is actually a net two-possession value. More detail on this later.
5. Use the difference in expected possession values (from #4 above) before and after an action occurs (e.g. a pass) to derive the value of that action.
6. Assign those values to players involved in the action. We believe that in this step we’ve derived some value allocation methods to push these Goals Added methodologies forward, toward more accurately valuing all players at all positions.
The method they use for calculating the xG value of possessions is a machine learning algorithm called XGBoost (nothing to do with Expected Goals, despite the name). I’ll avoid going too deep into the weeds here, but XGBoost is a form of gradient boosting, which makes predictions about a variable of interest by generating multiple model predictions, sequentially, and using the results to update the algorithm’s final prediction.
ASA made several interesting conceptual decisions when developing G+. First, it measures possessions as the sequence of individual player actions in a possession chain. Goals Added groups individual player actions into pairs of sequences, meaning they take a team’s possession chain (every action in that possession), and the subsequent possession chain for the other team.
Treating possessions as pairs means that both sides of the ball can be measured, and each action can be measured in terms of both the contribution it makes to the probability of scoring a goal, and the probability of conceding. This not only rewards players for their contribution to their team’s offensive threat, but also holds them accountable for helping the opponent.
Another element which I really appreciate is the way they credit players for successful passes. The passer and the receiver share the credit, and the amount of credit that each receives depends on the way that ASA’s Expected Pass% (xPass%) model evaluates the difficulty of the pass itself. They also treat turnovers in a similar fashion, giving more or less credit based on the difficulty of the turnover and the threat of the opposition play.
How G+ Compares with Other Possession Value Models
Though it has received quite a lot of attention, Goals Added is not the first of its kind. The starting point for a lot of this work is Sarah Rudd’s work producing a framework for analyzing offensive production, using Markov chains, and in recent years there have been various attempts to measure the value of possession and player actions. Some of the standout examples include:
- KU Leuven’s Valuing Actions by Estimating Probabilities (VAEP) framework
- Stats Perform’s (formerly Opta) Possession Value Framework
- Karun Singh’s Expected Threat (xT)
- Luke Bornn and Javier Fernandez’s Expected Possession Value (EPV)
There are a number of differences between G+ and these other approaches to measuring possession value, which have been discussed at great length in ASA’s roundtable discussion of the model, and their article looking at the alternative approaches. Further, KU Lueven have compared Goals Added with their VAEP framework, and it is a great look at the differences, and is valuable if you’re seeking to better understand G+ and its competitors.
There are some conceptual and methodological differences between all of these models. However, the biggest differences are the number of actions included in the measurement of a possession’s value, the way credit is distributed for successful passes and turnovers (as previously discussed), and the credit shooters are given beyond their finishing.
With regards to the type of actions that are included in the valuation of possession, each model varies slightly. The VAEP framework, for example, uses a snapshot of play, training a model to predict the probability of a goal in the 10 actions following the action being measured, while xT is primarily interested in ball progression, only attributing value to actions that advance the ball, such as passes and dribbles. Goals Added, on the other hand, includes the entire possession chain, and the opposition possession chain that follows.
The way that the different models treat shots varies as well. The Goals Added approach gives a player credit for getting in a good shooting position (because of the way pass value is shared), and then the added value from the shot itself is computed by taking the difference between the xG of the shot, and the possession value from that location. They argue that this makes it possible to assess whether shooting was the right choice in that situation, and it helps assess what the player did between receiving the ball and taking the shot. Finally, they also modify xG to calculate the probability that the shot produces rebounds, so as to attribute credit to the player for creating valuable opportunities from the shot being saved.
Applying G+ to Football in the Real World
American Soccer Analysis have done a great job of explaining how their model works, and how it compares with the competition, but they’ve also spent quite a bit of time looking at the application of their G+ model in the real world, analyzing the MLS using the model’s results.
This not only shows how G+ can be applied in the real world, and how it can contribute to the advancement of our knowledge of the sport, it also serves as a validity test. If G+’s results were wildly different from the eye test, or from other metrics for measuring player performance, this would require some investigation. But on the whole, they find that it is pretty consistent.
For example, they demonstrate that G+ is more predictive of future points per game than xG, shots, goals, and past points per game, as demonstrated in the following figure.
They also compared Goals Added with the trained eyes of analysts and clubs. First, ASA recruited a group of analysts from a number of outlets, and asked them to watch a Seattle Sounders game and rate the 10 best and 10 worst plays by Nicolas Lodeiro.
While they found some differences between the way analysts assessed plays and the way the G+ model assessed Lodeiro’s actions, there was broad agreement, which is promising.
They also compared MLS player wages with goals added above replacement. This essentially tests whether clubs are paying more for players with higher total goals added values.
Sure enough, they find that MLS clubs are paying more for players that G+ values highly. Assuming that, on average, MLS clubs are paying good players more, this is a positive for the Goals Added model.
Measuring possession value and the value of on-ball actions is a huge step in football analytics, and Goals Added is probably the best effort that anyone has managed so far. It’s unclear whether G+ will ultimately become a dominant metric in the sport in the future, but it’s certainly a really valuable contribution to our knowledge of the game.
The hope is that the work being done by ASA and its competitors will eventually lead to publicly available metrics for measuring player actions, which will further our knowledge of the game significantly.