I disagree. It is pretty decent overall considering how many things they cant actually measure yet. For example. For Euro 2024 I tried an experiment for two games. Trying to rate every player on the field for a single team as objectively as possible. My intention was to test out how feasible it is for journalists to watch a match and give consistent ratings for all players on the pitch the next day in newspapers. First I picked out the match Belgium vs Romania and I was rating Belgium players. All of them at the same time, as well as I could by noticing every bit of performance, off and on the ball. I was having a phone with me, taking notes and adjusting ratings in rral time while mentally tracking performance of each player throught a match. These are final results: Casteels - 6.5 / 8.0 Theate - 7 / 7.3 (subbed off 77') Vertonghen - 7.5 / 7.3 Faes - 7.5 / 7.3 Castagne - 7.5 / 7.8 Onana - 7.5 / 6.9 Tielemans - 8 / 7.6 (subbed off 72') Doku - 8.5 / 7.6 (subbed off 72') De Bruyne - 8.5 / 8.6 Lukebakio - 6.5 / 6.9 (subbed off 56') Lukaku - 7.5 / 7.1 --- Trossard - 6.5 / 6.8 Mangala - 6.5 / 6.9 Carrasco - 7 / 6.6 Debast - 6.5 / 7.0 --- Tedesco - 7 The first rating is my rating with 0,5 precision and the second number is sofascore rating (after the match, at the time, it might have adjusted later on). The 1-10 rating scale I use is this one: 10 - brilliant 9 - excellent 8 - strong 7 - positive 6 - okay 5 - mediocre 4 - bad 3 - disappointing 2 - terrible 1 - abysmal I wasnt checking sofascore detail during the match at all. I couldnt even if i wanted to because this was extremely challenging to do right. I had to be 100% on and still felt i could keep up with dynamics of the match. Either way, i concluded that newspaper ratings are unreliable because they are extremely difficult to be made consistent and faithfull to everything that happens on the field. Plus, i was surprised that somewhat my ratings are similar to sofascores. I did one more game for germany like that against switzerland: Neuer - 7.5 / 6.6 Mittelstadt - 7 / 7.2 (off) Tah - 5.5 / 6.5 (off) Rudiger - 7 / 6.9 Kimmich - 6 / 6.8 Kroos - 7.5 / 8.2 Andrich - 6.5 / 7.1 (off) Wirtz - 7.5 / 7.4 (off) Gundogan - 6.5 / 6.6 Musiala - 8 / 7.2 (off) Havertz - 6.5 / 6.8 --- Schloterback - 6 / 6.6 Raum - 7.5 / 7.8 Beier - 5.5 / 6.6 Sane - 6.5 / 6.9 Fullkrug - 7.5 / 7.3
Using basic statistics (goals, assists, pre-assists), the closest thing to something like this that I could come up with is to take the team’s goal difference and subtract the player’s G/A. Obviously this is very imprecise, but at least it gives an interesting perspective 1. GF/GA You only take the matches in which the player actually played. You calculate: GF/GA = (team goals scored) / (team goals conceded) This measures how dominant the team was with the player on the pitch. 2. Minus Player (GF/GA without the player) From the team goals scored (GF), you subtract the player’s goals/assists. Goals conceded (GA) remain the same. Then you recalculate: (GF − player G/A) / GA 3. Diff This is simply: Diff = GF/GA − (GF/GA without the player)
I dont understand GF/GA formula. How do you get 73.6% for example? If team scored 100 goals and conceded 40? What would that be? 60%
Sir Alex Ferguson: "I don't mean to criticize any of the players who played for me at United, but there were only four who were world-class: Cantona, Giggs, Ronaldo and Scholes. And of the four, Cristiano was like the ornament on top of a Christmas tree. A genius." Let's note Ferguson is considered by many as the greatest manager of all time so his gradations are significant - he's coached Rooney, etc. so he clearly elevates those four into an elite level and says Ronaldo is above even that level. He's seen these players every day for literally decades so with his experience he can create these gradations in abstract levels in his head easily and Ronaldo is at the top and nobody else is there with him. A good summary from Gemini AI on what it means and the significance of the words: When Sir Alex Ferguson called Cristiano Ronaldo the "ornament on the top of a Christmas tree," he meant Ronaldo was the unique, exceptional, and most gifted player among a group of world-class talents he managed, highlighting his genius and beauty as the final, shining touch that completed the whole picture of greatness, a step above even other legendary players like Cantona, Giggs, and Scholes. It wasn't a criticism but the ultimate compliment, signifying Ronaldo's extraordinary, almost divine, talent that elevated the entire team. What the Metaphor Means The Star of the Show: Just as the star on a Christmas tree is the most prominent and beautiful decoration, Ronaldo was the standout, most brilliant element in Ferguson's collection of great players. Completeness & Perfection: The ornament crowns the tree, making it complete and perfect; Ronaldo's genius was the pinnacle that made Manchester United's team truly exceptional. Genius: Ferguson explicitly called him "a genius," emphasizing that Ronaldo possessed a special, almost magical quality beyond even other top players. This is I mean by him being an efficient cause - he transforms and completes the actuality of his team's potential. Even Ferguson knows it.
Yeah, it’s actually (GF) / (GF + GA) * 100 In this case, I’m counting Messi’s club games in europe. He has 853 games, the team scored 2109 goals and conceded 755 So it’s (2109) / (2109 + 755) = 73.6%
I disagree totally from you I realized while watching all of Ruud Gullit's entire matches in the 1988 Euro Cup Final Phase with 100% attention and super focused SofaScore Ratings, in some metrics they invent the statistics and put in totally fake numbers like in Defensive Headers Won % or Offensive Headers Won % like in Interceptions and tackles won % as well and winning the second ball for the Team or defensive covers volume % or dribbles completed % or Crossing ... or Long Range Balls .... or Killer Balls .. they watch pretty well Clearances Slide blocks Defensive blocks because they observe the Key-Passes pretty well, passing completion rate % and Shots Attempts
Interesting mate. Are the goals/assists figures the OPTA ones (with penalties scored, and without pre-assists and non-OPTA assists etc)? More Alex Ferguson quotes on 'best ever' players in general: Alex Ferguson Quote: “Critics have always questioned whether players like Pele from the 50s could play today. Lionel Messi could play in the...” “Critics have always questioned whether players like Pele from the 50s could play today. Lionel Messi could play in the 1950s and the present day, as could Di Stefano, Pele, Maradona, Cruyff because they are all great players. Lionel Messi without question fits into that category.” Sir Alex Ferguson pays tribute to Alfredo Di Stefano | Football News | Sky Sports “There is a phalanx of great ones – Cruyff, Maradona, Pele, Puskas and Di Stefano. Di Stefano was one of the greatest in my mind.” Sir Alex Ferguson Named His Greatest Footballer of All Time Sir Alex On The GOAT Debate..”I was a fan as a Kid..” #football #goat #story #pele #maradona
It is an interesting spin on gc% to be honest, but doesnt quite capture expected performance in a way it is unambigous. It seems to be gc% with inclusion of relative number of goals conceded as a proxy for a level of opposition. Yeah, it is very difficult to measure expected performance, especially with stats only while player is on the pitch, because his presence changes the whole dynamic from inside out. However, it does measure how much of a difference player is in absolute terms.
Applying the delta logic, the ideal formula would be soemthing like: GC% - xGC% = delta performance Where GC% is calculated as it is (non-opta assists, pre-assists seem to be good inclusion here), and xGC% is one of those difficult to measure metrics that would have to capture most important variables of player involvement like role, volume of attacks (team attacking strength), quality of chances, etc. It depends how far we go. xGC% can be (manually) empirically calculated but only after enough players have GC% data, similar to making an xG model for shots out of actual goal data. Machine learning brings it to another level.
You just preceded me by a few minutes. I was going to write that you were both trying to reinvent the wheel in my opinion (absit iniura verbis of coarse). Expected goals is the metric professional scouts consider, and there are analysts paid to calculate xg numbers and make them available. The ability of a goalscorer is probed by ratio of xgs and actual goals. Team goal chances creation prowess can be quantified by average xgs. That is the metric that is believed to be the most reliable by professionals.
The reason xG has became so popular is exactly because it makes sense in principle. It is a proper conceptualization of performance. Expected goal is a metric that says what is expected probability for a player to score a goal given a quality of a shoting opportunity. It answers, what is an expected probability that would an average player (we can define what average player means - it is a reference point) score from this specific circumstances on the pitch. Then there is xGoT, based on a kind of shot player generated, what is expected probability that it will be scored. It uses historical data to assess these probabilities. How it comes about is that people freeze 100 000+ shots in a moment shot is taken, then they label each shot based on various observables (this is very xG models differ and what determines quality of xG model) like movement of ball (in air, on the ground, fast, slow..), proximity of defenders, number of defenders, percantage of goal covered by those defenders, proximity of a shot to goal, gk positioning and percantage of goal covered by gk, etc. So they take each shot, label it however they seem to matter and then they assign outcome to it (goal or no goal). They do that for super large number of shots and then they train computer to spot patterns. Computer runs the data (trains on it) then creater an internal model of which variables lead to goals at which probability rate. Then you give to the trained model a new shot that is not in a database, label it as well based on categories you choose, and model tells you: "based on what I saw in my training, how frequently such goals are being scored, this shot has X probability of being scored." The same, independently is done for xGoT, and any other expected metrics. You train a computer based on real data, it creates an internal model then uses that model to predict likelihood of any shot that is given to it. The problems are related to training: 1. Model depends on database you use. If you use too small data set, it will not be able to calibrate itself towards real probabilities and pick up the noise. Furthermore, expected probability relates to level of competition it is chosen. Computer can be trained on data at mateur level then it produces expected probability that amateur player scores this type of shot. Choice of database changes what "average player" means. This is something we can't know regarding xG models. We don't know what database they used. Presumely, they used top level football database, but this is one of choices you have to decide beforehand, before you start training. What you want expected probability to represent. And depending on which database you use, expected probability will reflect average of that database. 2. Categorization and labeling of each shot is arbitrary. This is where quality of xG models vary different from a provider to provider. Because you have to tell computer what parameters of shots are important to consider when creating a model based on these parameters. This is a true bottleneck in machine learning that determines quality and realism of a model. If you start labeling shots based on irrelevant parameters, like the air temperature in the moment of shot being taken, computer will start looking for patterns on how much temperature affects probability of shot being scored. Since database is not infinite, so that computer has a chance to "conclude" that temperature has 0% influence on outcome of a shot, it will probably spot some non-zero pattern between temperature and shot being scored. This will be noise, not reality. So you can't simply overwhelm training with nonsense categories for each shot. You have to cleverly pick parameters that are deemed important (like distance of a shot, etc.) so that it can model based on these parameters. Everything depends on how computer is trained. Poorly trained xG model, will be nonsensical. 3. The crux of the issue How do you label for 100 000+ shots relevant parameters like the spin of the ball, level of bounce, easiness of shot being taken? These are real bottlenecks. If through ball is a bit overhit for 1v1 with on-rushing GK and player is struggling to catch up to it, but still manages to generate a shot from it out of his own effort, this will be very difficult shot to convert. We all understand low expectation of that shot being converted because it was very difficult to reach the ball. But how do you capture this in the process of labeling when training a model? Early xG models where agnostic to such things. Now, they are becoming better at it, but it is still a training issue that would require extremely clever labeling process. What can you do about spin for example? You can label each shot based on low, medium, high spin? But some spins can be high, but still an easy shot to take. Some low spins are very difficult to adjust to in a moment. These are real challenges with expectation models across the board, but in principle, expectation models are the future of football, which is evident by prominent use of them professionally and commercially. Here I would like to make a distinction between commercial products like xG models, Sofascore ratings, etc., and actual models that reflect reality. These companies don't have incentives to capture nuance of performance or be accurate, they are making commercial products that people will use and buy. So I wouldn't expect for such companies to solve these issues in detail, as their goal is not football accuracy, but commercial use. Also xG models are only related to goals. This is not quite true, as player can generate higher xG opportunity by dribbling, set upts, etc. So it is not purely reflective of team's ability to create chances for a player. It could be true that is the best way atm, but I am thinking where it will lead. Expected performance is a correct way to go and it will continue to go that way. Natural extension and progression of it is, just like xG models are produced, doing it for every type of action player takes. This applies to off the ball movement as well. It is possible in principle to create a model that in real-time produced expected probabilities that player increases chances of scoring based on circumstances (on the pitch configuration) he is currently occupying, and then judging his actual action against this expected probability. Obviously, this is a bit distant future. It requires significant advancements in tracking players actions throughout the whole match, plus clever classification of actions. But frankly, we already have technologies to do all of that like GPS tracking (think of offsides, how they create models of players), computers, AI, etc. It is all possible tbh, it just takes resource, expertise, time, incentives. XG models are good analogy of what future evaluation of performances will look like. Then just like we measure 30 goals - 23,3 xG to determine how much player overpeforms expectations of shots, this translates to every action at arbitrary level of granularity. This is why I suggest that a more accurate way of conceptualizing player's performance is in the -/0/+, delta model, where: 0 = player performed exactly based on expectations given his circumstances on the pitch +1 = his performances has an expected probability of increasing goal difference of his team by 1 -1 = the opposite This can be done directly by a model who observes everything that is actually going on on the pitch (like xG model) and it doesn't require middlemen like proxy for how good their team is, opposition, etc., or outcomes of the match. XG model doesn't need outcome of a shot to calculate a probability of shot being scored. It comes down to completely cutting off proxies and indirect evaluations, to direct evaluations of what is going on on the pitch. So the 0 delta rating, wouldn't require any additional consideration of how good his team is or opposition or what teammates do, etc. It is already embedded in the model. This would normalize anything and make performances directly comparable without considering wider context and proxies. Purely based on football mechanics and dynamics.
There is another similar method on the web MessivsRonaldo called “points won via G+A”. https://www.messivsronaldo.app/detailed-stats/points-contribution/ It basically takes how many points the team won in the match (3, 1, or 0) and asks “how many points would the team have won without the player’s goals and assists?” For example, if the team wins 3–0 and the player scores 1 goal, that is 0 points won by the player, because without his goal the result would still be 2–0. Another case: if the team wins 3–2 and the player scores 1 goal, that is 2 points won by the player, since without his goal the result would be a draw. What is interesting about this method is that it can, in a way, benefit players who face stronger opponents, because if your team wins by a large goal margin (and assuming that in those cases the opponent is weaker), the player does not gain many points. On the other hand, in tighter matches the player can gain more points. Here I did the calculation myself, using non-penalty goals, Opta assists, non-Opta assists, and pre-assists.
Yet another flawed method that leaves aside the true goal of a football team (and consequently of any serious football player), that is winning matches and overvalues players that monopolize teamplay at the expenses of team result.
Coincidentally, the order of the list is consistent with what one would expect to find in a list of the best players, the strange thing would be to see Zico ranked above Messi or Pelé, and that doesn’t happen.
Something interesting I counted some time ago is the player’s involvement in the team’s shots, using the same criteria as with GC% Goals: (Shots) Assists Opta: (Key Pass) Assists Non-Opta: (Indirect Key Pass) Pre-Assists Opta: (Pre Key Pass) Pre-Assists Non-Opta: (Indirect Pre Key Pass) The sample is limited, but I don’t think it would change much with more matches. Pelé was involved in 54.66% of his team’s shots, and you know what? In those matches, he was involved in 68.75% of the goals (which is a GC% consistent to his career total). That means the shots Pelé was involved in had a higher probability of ending in a goal. Cristiano Ronaldo, in those matches, was involved in 50.2% of his team’s shots (if we remove those early ones from 2006, his percentage rises to 56%, which is a range similar to Pelé’s). In those matches, Cristiano was involved in 41% of the goals. Wasn’t the supposed reason for Cristiano having a low GC% that he didn’t participate much in the play or some nonsense like that? So why is his involvement in his team’s shots similar to Pelé’s?”
is Interesting yes this one , on my view Pelé was much more complete & versatile than Cristiano Ronaldo as simple as that . ! Pelé, also a veteran, between 1968 and 1973 seasons became an excellent defender as well. with a very high defensive volume for his position.
Ac Milan 5 .......vs ............. 0 Real Madrid Home-Game European Champions Cup 1988-1989 season ! Ac Milan 16 Shots Attempts 8 Shots on Target ratio % 0, 50 % accuracy % Real Madrid 10 Shots Attempts 1 Shots on Target ratio % 0, 10 % accuracy %
Are shots equally correlated to a high GC% figure, compared to a key pass? I would assume a key pass to assist ratio would be better than shots taken to goal ratio for most. That would be the first obvious answer that comes to mind, when discussing the weird question over similar shot involvements, despite the GC% difference. For most cases, I would imagine any player with a high number of key passes and playmaking attempts, would do well for GC%, over a player with similar levels of shot-involvements, mostly via shots-attempted on goal. Wasn't this built in to the statistic in the first place? Why are you asking us? In the end, these metrics are for me, biased for high usage rate, high proportion of energy spent on game-changing plays (not just shots, but playmaking attempts also) with the ball at the feet players, who are also fully backed-up by the system (with the added caveat that such privileges aren't universally applicable for all situations). I wouldn't be that surprised if Cristiano Ronaldo ranks lower than senior players such as Luis Figo in terms of usage rate during the 2006 World Cup run. You can chop up the numbers any way you wish, but this sort line of questioning makes me feel you have zero ideas about the limitations and biases of your own proposed statistical models, or you don't care at all, as long as it aligns with your previously conceived notions of player effectiveness.
To remove any confusion, I am going by what Trachta10 details out himself, not what I think makes sense. It ends up being the case that Cristiano Ronaldo himself should try to playmake more, because his GCA to key pass ratio, is better than his goals to shots-attempted ratio, appease Trachta10 and his particular methodology. As in, all the inaccurate shot attempts are tallied, yet all inaccurate passes that were initially attempted as assists, pre-assists are not. If you only count the ones that do connect, the passing related goal-creating actions (assists, deflected assists, and pre-assist) will usually have a better ratio when compared to key pass tallies. Once we tally creative passes attempted tallies, it may become a different matter, but since Trachta10 purposefully left them out, and asks the question, I am forced to deal with his methodology.
I'll just use Robert Lewandowski as a proxy, and use his GCA figures without the rebound shot-related GCA, to make a rough approximation of the kind of GC% raising creative actions that includes pre-assists. The numbers will not be a precise like-for-like, but we will get the ridiculous nature of how it is tallied, and why using it to judge players is foolish, especially from a poster who is known prioritize his beliefs over statistical integrity. Robert Lewandowski (domestic league statistiscs 17/18 ~ 25/26): 97 goal-creating actions (assists, indirect assists, penalties won, pre-assists) that are not shot-related in the creative process (shots that lead to a rebound). 261 key passes -> 37% 239 goals 1044 shots attempted -> 23% It turns out Robert Lewandowski should try more playmaking actions, instead of wasting his time trying to maximize his goal-count, on the off-chance that Trachta10 decides to tally the creative process in the most blatantly biased manner, and uses it to judge players holistically.
Cause he takes a lot of difficult shots (35+ yards, headers on crowded corners, difficult angles, with his weak foot, etc...) that makes up for the difference in key passes.