This season's A-League shows that the table does lie

Nice work, but you forget to mention Perth Glory in your contenders. Surely way above Adelaide over the past 2 seasons.

Great post

It would certainly be interesting to observe how some extra after-hours one-on-one coaching targeting his finishing might effect Hoffman's chance conversion rate. Would depend on Hoffman's willingness and the quality of the coaching though, the quality of the feedback and how much active learning time there is in the session affects skill acquisition. Because arguably the tougher and harder to teach component of being a forward is actually being able to make the right runs and be in the right spot to even create those chances. Shame his age profile isn't a bit younger if he could improve the chance conversion, Newcastle would be set for the future.

I can't get the line breaks to work for me even on my unedited comments at the moment. Tried it across two different devices too, so no idea if I'm just double cursed or what. :laughing:

As annoying as it might be to admit, achieving buy in from people typically requires a bit more than just some evidence. You need to be able to sell people on it with a good pitch, which requires being a good marketer/salesperson too. In sport nothing causes widespread interest, acceptance and adoption quite like quantifiable results (e.g. table position). So the challenge is to convincingly sell the link between the method and helping to achieve positive results; such as incorporating resistance training into a marathon runner's training program improving running economy (and therefore their race results).

Yes - I read all of that stuff long ago. I have no problems with using stats to work out what usually works. You would be insane not to. But, the stats are not perfect and depend on lots of subjective judgements. My particular issue with this article is the assumption that a historical view of what worked in the past can reveal a lie about current table positions. Surely the way the stats are compiled has to be continually revised to reflect changes in tactics. If not, the tactical innovations and resultant successes of people like Cruyff, Bielsa, Guardiola and Klopp would never be reflected by the stats. Mourinho's successes too are highly relevant, because his approach was so different to those four. The xG stats for 17/18 had Mourinho's Manu team 5 positions and 20 points short of their actual results. And who knows how the stats would have reflected Ferguson's successes, which owed more to his people management skills than any sophisticated tactical ideas.

Shabab I would have thought a good use for these kinds of stats is to help coaches work out what successful teams do that makes them successful. The problem is 'success' is defined by where teams finished on the table in past seasons. So the stats, at best, represent what worked in the past, not what's working right now. When the table positions according to the stats don't align with the actual table positions then it's the stats that need to be adjusted, not the table position. The table position is the objective fact the stats aspire to predict. The EPL this season is a great example of how the stats are deficient. Liverpool is 25 points clear of Man City in the real world, but 4 points behind them in xG /xGA world. With such a disparity in outcomes, the stats obviously need adjustment because Klopp is clearly doing something with Liverpool that the stats are missing.

I don't think so Shabab - I've found no amount of refreshing gets the line breaks back for edits of an already posted comment.

I see what you mean Matthew, but whatever we say it's hard to argue with the stats sceptics who say "the table doesn't lie".

It's not your fault Matthew - the edit tool messes up the formatting when you edit a comment. A really basic fail for an edit function.

Oof, sorry for any lack of linebreaks in there. They show up when I edit the comment but aren't displaying for me in the published post.

To be fair, at least in the case of the big European leagues (those being the data points that I have examined so far) players that someone might consider a lethal, clinical goalscorer regularly: 1) Typically only outperform their xG tally by no more than one to three goals across a season, and 2) Struggle to consistently outperform their xG tally over multiple seasons, often reverting to the mean (their xG) the season after if they did outperform it the season before. So for all the weaknesses in methodology that one could debate with some of the current predictive modelling being used, the xG model not taking into account a player’s finishing ability is something I am less concerned about currently. A bigger issue for example would be the definition of what equals a ‘Clear Cut Chance’, since you can end up with three different figures for the same EPL game depending on which provider you use due to differences in their methodology for defining a ‘Clear Cut Chance’. Along that same note, those differences in methodology mean that you can end up with variances in assigned xG totals for the same EPL game across the different providers (Opta, Wyscout, etc), so Inter-rater Reliability can still have some issues. Which at the end of the day means that the magnitude of Syndey’s overperformance against the predictive modelling for the season so far would be open to some variance if we had more than one data provider for the A-League currently (Alas Opta doesn’t do A-League currently). Having not looked extensively into how players rated as poor finishers perform against their xG tally for one, or multiple seasons, I’ll pass up offering an informed opinion on that subject for now. Based on the methodology though the hypothesis would be that a player like Hoffman shouldn’t be massively under-performing in actual goals scored versus their xG tally season upon season, but he certainly could have a season where he achieves such a feat. That is one weakness with the predictive (expected) modelling used at the moment though; an individual or team that is underperforming/outperforming the predicted modelling typically will revert to the mean eventually but giving an actual accurate timeline for when it will happen is a much harder feat. Which can throw things out a bit when a Goalkeeper outperforms their expected ‘Goals Saved’ statistic for the season and helps earns their team a bunch of extra points (David de Gea went on one such run for Manchester United during a previous EPL season).

Yes Mark, totally agree with your comment. What if's count for nothing at the end of the day.

Based on this article, it's obvious xG doesn't account properly for the quality of a team's goal-scorers and goalkeepers. That is, if you have Hoffman on the end of what would normally be an xG then you should multiply the value of that xG by .02 (or thereabouts). If it was Fornaroli you would multiply by 1.5. If these stats accurately reflected the real world then over time the stats would tend to coincide with end-of-season table positions. If you were training an AI system to predict the winner of a comp then the only real measure of the system's quality would be how closely it matches the end-of-season table positions. If you compare this to how you train an AI system to recognise a disease then the end-of-season table is like the diagnosed disease. The known diagnosis is an objective fact and the system is trained to zero in on those things that point to instances of the disease prior to actual diagnosis. Concordance with the actual cases is how you measure the system's quality. By analogy, the objective measure of a football AI system's quality would be concordance with the end-of-season table. The problem with football though is every competition has a range of different factors affecting the eventual outcome. Omitting any of those factors might affect the reliability of any prediction tool when applied to a particular competition. That is perhaps why overseas coaches have such a chequered history in the A-League. Maybe their predictive stats haven't been calibrated for A-League conditions. There is a tactical element that I don't think the stats capture either. For example, as a regular Sydney FC watcher I could also argue that the raw stats don't account for Sydney's ability to shift gears during a game. Because they can do that Sydney's stats are lower than they might be under real pressure and they regularly beat teams without necessarily dominating them for a whole game. You would have to incorporate something like the heart rate readings of each player to capture something like that in predictive stats.

You’re stretching your credibility wafer thin with analysis like this.

Sigh. It was pretty predictable that this article was going to go into a discussion of expected goals. Expected goals are interesting to a point, but they are like economic models and other forecasts. They can tell you what probably should have happened, but looking back they are no substitute for what actually happened. Saying the table lies because it doesn’t match up with what expected goals predicts should have happened is nonsense. One serious question I have which you might be able to answer, how does expected goals differentiate between Cristiano Ronaldo having the exact same chance at goal as, say, Elvis Kamsoba?