iezzex said: Do club rep affect result? Expand On 'no detail' processing, possibly.
On 'full detail', can't remember now, but I don't think so. I know for certain that player rep doesn't. I don't think club attributes affect the results either, but I suppose it does come into consideration when AI managers are deciding what team to play against you (you know, how after mid-season, suddenly games get a lot harder once they adjust to your actual team quality).
Yarema said: Way to misrepresent the data. Why don't you add the steps from 10-15 and 15-20 as well? Oh right, because they are quite similar as 1-6 Expand If you ignore 6 > 10 for dribbling, then dribbling appears just about linear as well:
You choose to see what you want to see. I choose to see what is actually there, and what I see are two different attributes, showing the same pattern of non-linearity at the 6-10 level, which reduces quite significantly the chance that it is a statistical error.
Can't be bothered digging it up, but there's more deductions you can make like this for other attributes by comparing his 6 > 18 data to his 1 > 20, 10 > 20 data and whatnot
Yarema said: Sorry what? Other than work rate, the other attributes are pretty much linear.
Obviously we aren't looking for the 6th decimal, there is margin of error. And they might even behave slightly nonlinearly but a linear function can be a very good approximation to reduce complexity. Expand Acceleration 1 > 6 = +42.7 (+8.54/extra point) Acceleration 6 > 10 = +58.6 (+14.65/extra point)
The non-linearity is not as profound as for work rate, but it is nonetheless significantly non-linear.
Additionally I would posit that the league quality itself introduces a further level of non-linearity (or perhaps the non-linearity is a reflection of it, and would be more pronounced in a realistic league). HarvestGreen used a league that from memory had players with 10 in everything. In the Premier League, you need at least ~13 acceleration just to survive, so I would expect to see a large disparity between say acceleration 6 > 10 and acceleration 10 > 14 in the Premier League.
keithb said: So again as always you have made a false claim. Nowhere has harvest green said attributes are non linear. Expand
harvestgreen22 said: Non linear Its influence is relatively significant from 1 to 8, but it becomes much smaller from 14 to 20
If he has, for instance, 16 ，18 , or even 20, it might waste a considerable amount of ability, but the effect actually only increases very little Expand
As for Mbappe, it's true I underestimated him a little as ST. I had him as 12th; in the latest best fit to the objective results I'm working on, he's currently 6th. The difference between us is not that I don't make mistakes and you do, but that I correct my mistakes, while you continue to insist that Mbappe is equal to Haaland even when reality shows otherwise time and time again.
And no, I'm not going to remove the results of your file from my list, so stop DM'ing me about it.
keithb said: So that was it. Where does he say speed, anticipation, dribbling and jumping reach for example aren't linear. Again we knew a lot of attributes don't matter at all, some don't matter after they reach a certain value. But the high scoring ones do.
I can't work out if you're stupid or if you think we're stupid? Maybe both?
As for being a liar when you released premier league 1 you didn't set any hidden attributes. Therefore genie scout wasn't producing skewed numbers. Yet you repeatedly lied that genie scout was at fault for giving Mbappe such a low score. Again this was a lie it was your appalling ratings. Expand What actually first came to my mind was HarvestGreen's excel tables, which show the non-linear effect of multiple attributes very clearly and in detail. Look at the right side:
Here we see that speed, work rate, among others, have non-linear effects.
Dribbling follows more of an inverted U-shaped curve, so in practice is mostly linear, but not always. I've always said that myself:
GeorgeFloydOverdosed said: HarvestGreen22, using a very artificial setup method (all attributes 10 bar the one being tested) and assessing without regard for position, claims dribbling is 4th most impactful attribute and follows a largely but not entirely linear benefit through 1-20. Expand
You've got the GS narrative backwards. I created the Premier League 1.0 weightings (without releasing them), then was informed by someone else that there was a problem with GS, and deduced it was the hidden attributes causing the majority of the issues because GS doesn't calculate things correctly with them. The weights work as intended in other programs (i.e. FM26 scoring system), including the hiddens (i.e. FM SuperScout). When I released Premier League 1.0 for GS, I removed the hiddens so it could work as best it could, since we can't really do away with GS entirely yet.
Objective testing has shown now that Mbappe is worse than 34-year-old Lewandowski and 36-year-old Messi gives him a run for his money. Mbappe is no where near Haaland on any measure - team position, goals, rating. And your own rating file turned out be worse than FM Genie Scout's default on key measures. Your file shows that apparently Dybala is as good as Lewandowski or Kane 🤡
So as I posted in another thread, I've ran some controlled tests of specific players and found they finish in this ranking order (ranked by team position):
Erling Haaland = 1.666 Robert Lewandowski = 2.666 Kylian Mbappe = 3 Daizen Maeda = 3 Harry Kane = 3.5 Dusan Vlahovic = 3.571 Victor Osimhen = 3.833 Lionel Messi = 4 Donyell Malen = 4.666 Robert Glatzel = 6.5 Paulo Dybala = 6.833 Adam Le Fondre = 7.5
6 full season samples each, all relevant competitions on full detail, and this isn't your run-of-the-mill plop Haaland at Arsenal and see how he goes - the rest of the team are all identical non-exceptional players in order to minimize variation while remaining realistic and tested to finish ~5 on average by themselves (we replace the STs with 4 duplicates of the player we want to test)
If you use Falbraav's file which puts the weightings into Genie Scout, it doesn't align with the real results at all:
If we remove the hiddens, which are known to be problematic in GS, not much changes. Lewandowski remains behind Mbappe & Osimhen; Maeda remains near the bottom.
The rankings remain incorrect if we plug Falbraav's GS figures into FM26 scoring system (to rule out GS being the problem):
While I was doing this and looking at OP's table, I realized what could be a big part of the problem. He's using points to measure success rather than team position. Points are very unreliable, because they can be 57 for one season sample and 83 the next - what doesn't change much is the team position, so that team will likely finish 2nd both times even though the point tally is all over the place. You would still get some valid correlations obviously, i.e. no pace/acc = no points at all.
I suggest to the OP if he intends to have another crack at it, to assess by the team position and not the team points.
And another problem with points vs team position could be that it neglects defensive advantages. It could explain why positioning does so terribly, for perhaps better defense favors draws as opposed to victories and losses - but a draw is only worth 1 point; a win 3. We want victories anyway, so this doesn't matter all too much either way.
He said it here, about Work Rate. Not sure about the other attributes, but I remember reading that and I've had a filter set at 6+ Work Rate ever since. Lowering the Work rate 1 , it have very serious consequences (goal difference -110),
and it should be ensured that a player has at least 6 and preferably 10 Work rate. A player with a Work rate is not desirable. But continuing to increase Work rate (from 10->20) is low effect. The difference between Work rate 10 and Work rate 20 is relative small.
The 6-point attribute seems to be the threshold for some attributes.
Core attributes are these: Special : Work rate (need to reach 10, higher is useless) Expand
keithb said: Well if thats all he said then he didn't say it did he. But no surprise if Bozo is quoting that. He's a liar. Expand Amazing.
Yarema said: If an attribute point has a value of 10 from 4 to 5 and 1 from 14 to 15, how exactly do you set a fixed value for scoring in GS? That is the fundamental issue. You need linear scaling at least in the vast majority of values. You can't just say oh it's an average of lets say 3 because it'll be right only in a narrow bracket or maybe even never.
What you can do for attributes that have minimal or no value above a certain point is to set it's value to 0 (or whatever minimal value the calculations show) with a caveat that it only applies above X and that you need to filter for it. I am not disputing that some attributes might act that way, but that assigning them some value (by feel?) is misleading. And again FM-Arena tests showed that going from 18 to 19 to 20 pace is more or less just as valuable, it keeps it's scaling. Can you win with 18? Sure, but 20 is still better. Expand With respect, my post was my answer to that very question.
You could argue the method is inadequate, but you can't say it achieves nothing at all, and I'd challenge you with the following question: If not this method, then what? Assume attributes are linear and get assessments that are even more wrong?
It's true that filters can be part of the solution, and I've suggested to people before to use filters in combination with GS weightings if they can, but obviously this isn't a very good solution. Filtering is too blunt a tool for one thing; it's better to use weighted weightings instead usually.
I don't believe in FM Arena testing much, because they obviously contradict the findings of HarvestGreen, Orion, and myself. Even aside from the results, I didn't like some of the methodology that I could see, in the same way that I don't like how HarvestGreen uses an artificial league where everybody has the same attributes, instead of using a real league like the Premier League.
We see something similar reveal itself in this guy's testing. I would characterize it as a kind of blase 'let's apply a python script to this' mentality. Sometimes, the kind of info that gets can be very useful, but I think more typically it leads to skipping over the necessary nuances that only someone who has waded deep into it with hours of trial and error and reflection and critical thought can discern.
For example, he says his results show that attributes are linear. Ok, so why hasn't he reflected and deduced why this contradicts HarvestGreen's findings? The most likely answer is that he simply hasn't put in the time in doing that. By his own account the data was from an '8 hour overnight session', and I don't know how long he spent on it overall, but my impression is that this is probably something he's done over a week or two.
I'm not trying to throw shade, and it's no crime to not spend your life on analyzing FM, I'm just saying it's an approach that has certain downsides. My downsides are that I'm not a programmer, so I can't use such a methodical and fast approach, and because I'm always coming up with different ideas, I don't say 'this is the result I got, that's it from me'.
Now you say that assigning values by feel is misleading. Well let's look at a concrete example:
I brute force test players until I know who ranks where.
Starting with informed guesses (i.e. start pace at '100' not '0', and passing at '0' not '100' ), I adjust the weightings in Genie Scout until the player rankings all correspond exactly to their actual performance ranking.
Because some or many attributes are not linear, there may be a 1% error here, a 3% error there, and even a handful of ranks out of the thousands out of alignment, but for the most part, it does its job.
I could continue the example further, but I think the point is already clear - where exactly is the error or impossibility in it?
OpticFawn said: can you give us the pics for each position on genie scout with the data for the new files for best fit gs file Expand I'll be posting it soon, but it's still a work in progress right now.
Translating the real results to Genie Scout has been proving to be really tricky. As ChatGPT has confirmed to me, there's no straightforward way of deriving the specific combination of attribute weightings from the results alone. For example say I find drib 30 + ant 30 + pace 70 is what works.. how do I know it's not just ant 30 + pace 70, and dribbling has 0 effect?
Thankfully I can rule out a lot of attributes from the get-go from previous testing, but I still have to just manually fiddle around with all sorts of different combinations to find out what works and what doesn't until I have something that at least aligns with the real rankings as closely as possible.
I started with the following process:
1. Start with Premier League 1.0 'weighted weights' version 2. Compare players who have completely different sets of attributes but had similar results 3. Adjust weighting of 1-3 attributes that I know matter (i.e. pace, anticipation, work rate, etc.) starting with the most likely attribute (i.e. pace) until all players are roughly where they should be ranked. 4. If one player's rank is fixed, but then another falls out, return to step 3. 5. Find a player with high rep or high CA who has unusually low % rating with the changed weights. Test performance in controlled test first, then do step 3 with them if performance doesn't match rank. 6. Keep repeating step 5.
Right now I've got a functional set of weights, but everything has to be changed around again the next time a sample finds a player changes their ranking. I think 6 samples is giving us a fairly accurate picture now to work with, but for true accuracy I think it's going to require at least 10.
Then there is the matter of position variation. But my first result in another position so far seems to suggest the weights could be the same for every outfield position. Once I've done all positions, with perhaps 10 samples for STs for calibration, then I'll post it.
I mean it's kind of the basis of all these weighting systems, because if a point of pace was vastly different from 9 to 10 as opposed to 15 to 16 there would be no point in having a fixed value for an attribute weight.
So you either accept they are pretty much linear (with a few exceptions) or throw GS ratings out the window. Expand It's not the basis of my weightings, as non-linear values do not make useful weightings in Genie Scout impossible.
Here's how I handle the problem of non-linearity:
Finishing - I know from testing it has no advantage above ~7, and is only of moderate value even below that all the way down to one. Therefore I give it a very low weight, or perhaps even no weight at all.
Work Rate - It's very clear that work rate is crucial going from 1 to 6 but has greatly diminishing returns beyond that, and pretty much tops out at around ~11-13. If you don't believe the testing done on this, just look at player like Haaland, Mbappe and Messi - their work rate is 13, 12, and 9 respectively. The compensation in GS for this is fairly straightforward, instead of weighting Work Rate equal to pace or acc if we were just looking at the 1-to-6 effect, we give it around 20-40 instead of 70-100.
Jumping Reach - This has one of the most difficult curves of all to accommodate. It is very important going from 1 to ~7, then moderately important ~8-13, has a massive jump in performance once going above the league threshold (~14-17 for Premier League), but is then near useless going from ~18 to 20 since probably no other player will challenge above 17 anyway. Here we target either ~8-13 (moderate rating), or the ~14-17 threshold (high rating, perhaps the highest even above pace/acc).
Of course you run into situations where GS considers 1 pace 20 acc equal to 10 pace 10 acc, but those 1 pace 20 acc players are the exceptions not the rule, so it doesn't render the whole enterprize futile.
keithb said: haahah. You lie and misrepresent what I say often. You're a clown.
No Mbappe and Haaland are not rated the same. And when you increase TS to 107, like I said in the post, the gap becomes even bigger. Its 2% and then becomes 4%. They are not rated the same. Is there actually something wrong with you?
AGAIN how can you create a file and use and not see this?? There's a worrying disconnect. Expand To be honest I didn't even check if you had a TS rating, because I took one look at the FS rating and knew it was rubbish.
The screenshot I provided is, you'll agree, the FS ratings you've given.
So do you want me to use the FS or TS rating?
Looking at your TS ratings, you have Osimhem above Lewandowski even though both having 15 jumping reach, so more of the same 💩 either way.
keithb said: Ps I dont have Mbappe and Haaland rated the same. You dont even know how to use and read GS. How can you make a file for it? I love that you lie all the time, it just makes it funnier. Expand I wouldn't want to misrepresent you. This is what your GS file says does it not?
Default database, start of game (England - 3/7/2023)
There are now 6 samples per player. That's 2,736 Premier League matches simulated on full detail + all other relevant first team competitions (i.e. Community Shield, Champions Cup, Club World Cup, etc.) simulated on full detail as well.
Edit: I've added a color-coded column to compare 4 different GS rating files. 1st is new Best fit file, 2nd is Premier League 1.0, 3rd is keithb, 4th is Falbravv's META FMA 2.0 file based on skawkclsrn's recent data. Green = 1% or less difference. Yellow = 2-4% difference. Red = 5%+ difference. Ignore Harry Kane, he needs a slight readjustment right now due to the latest samples.
There are a lot of new deductions that can be made.
One is that now we have a rough idea what GS % corresponds to in terms of concrete performance numbers, and another is that we can tell at what point random variation makes distinguishing between % numbers pointless.
We can always do with more samples, but if we just take the difference between whoever is highest and lowest between Lewandowski, Mbappe, Maeda, Kane, Vlahovic, Osimhen, and Messi (total +/- 1 position difference maximum) we get:
((4.08/80.70)x100) = 5.06% - My best fit of the data so far in GS (not to minimize %, but to correct player rank) ((14.30/91.99)x100) = 15.54% - Premier League 1.0 'weighted weights' version ((14.94/88.11)x100) = 16.96% - Premier League 1.0 ((14.14/92.25)x100) = 17.14% - 'Pure Performance' FM24 ((13.45/77.59)x100) = 17.33% - My first/original FM24 ratings file ((15.89/84.94)x100) = 18.71% - 'Blended' FM24 ((18.34/83.29)x100) = 22.02% - ykykyk05251's weights file ((22.88/88.01)x100) = 26.00% - Orion's coefficients file ((29.27/91.47)x100) = 32.00% - Genie Scout Default
That is where we are accounting for extreme cases. In other words, a 4-5% difference in GS is almost certainly an objectively better player.
But what if we make it a bit leniant, and tolerate ~10-20% being exceptions we miss. In other words, this is the more typical uncertainty range. To discover this, I exclude the bottom % player from the list:
((4.03/80.70)x100) = 4.99% - My best fit of the data so far in GS (not to minimize %, but to correct player rank) ((6.21/88.11)x100) = 7.05% - Premier League 1.0 ((6.00/83.29)x100) = 7.20% - ykykyk05251's weights file ((7.34/75.95)x100) = 9.66% - Falbravv's 'META FMA 2.0' file based on skawkclsrn's data this week ((10.39/89.51)x100) = 11.60% - Falbravv's 'META FMA' file based I think FM Arena's testing(?) ((11.67/91.47)x100) = 12.76% - Genie Scout Default ((13.45/77.59)x100) = 13.20% - My first/original FM24 ratings file ((12.61/95.25)x100) = 13.24% - 'Pure Performance' FM24 ((12.42/91.99)x100) = 13.50% - Premier League 1.0 'weighted weights' version ((13.21/94.59)x100) = 13.97% - keithb's weights 💩 ((13.64/88.01)x100) = 15.50% - Orion's coefficients file ((14.14/84.94)x100) = 16.65% - 'Blended' FM24
So we see that whatever the case, ~4-5% is going to be the range within which the player who is better becomes uncertain, but also the difference is going to be pretty meaningless objectively.
Wigo said: Guys i read all threads about atributes testing ant meta ones... this topic is the most intresting to me and i think i have a question that nobody answers. We found that speed in general is the king, but everytime i start a new save i see that (fm26) first season Harry Kane with speed of turtle is scoring unbelievable amount of goals and is gets player of the year award. So technically is speed is everything how come this happening??? so still plenty theories in my head that maybe for some positions speed is not so overpowered as we all think? Expand I can answer this for you because I'm currently doing isolated testing and analyzing the attribute differences closely for a number of players, and one of them is Harry Kane.
So far my testing finds that Kane is in fact objectively one of top STs in FM24.
I've already come up with a draft set of weights that allows Kane's low acc/pace and other quirks to align smoothly with all other players tested so far, but I'll just tell you a more abstract version of what I'm seeing.
This is my impression so far (for ST at least specifically):
1. Pace/acc, and stamina to keep it going through a game, are still the top factors. 2. Pace/acc is not *strictly* necessary to do well. A player with 9 acceleration can function decently as ST in the Premier League (I've tested this). If you have a team of them, you'll be relegated, but as individual exceptions to the rule they can do the job. Although I haven't tested it, I would expect similar with a team of Harry Kanes. You need at least ~13-14 average for the team as a whole in Premier League just to survive. So pace/acc remains very important. 3. Only about one third of attributes matter. Anticipation and dribbling are in this third for ST, but all of these are at least a bit less crucial than pace/acc. For instance, Messi has 20 dribbling, and while he is not no.1, it does appear to be a key factor in him being one of the top STs according to testing even at age 36 in FM24. None of these attributes are strictly necessary either except perhaps 1 or 2, see: Daizen Maeda 4. This is still dubious, but through analyzing examples such as Kane vs. Lewandowski specifically, it would appear that some attributes such as passing may actually have negative performance impact.
Here is something I already wrote that I was thinking of posting earlier:
There is no reason why Lewandowski should be better than Harry Kane, and yet he is. In fact, Lewandowski is slightly inferior on a number of things.
Lewandowski has advantages in the following areas:
+3 aggression +3 first touch +3 balance +2 anticipation +2 concentration +2 natural fitness +1 bravery +1 off the ball +1 acceleration +1 agility +1 strength +1 important matches
None of those are adequate to explain the difference, given Kane has the following advantages (I'll list just a few):
5 years younger +7 crossing +5 passing +5 vision +4 teamwork +3 left foot +2 consistency +1 determination +1 composure +1 work rate
I suspect that high passing and possibly crossing reduces team performance. Perhaps what is happening in the engine is that Kane is passing to a team-mate do to the job, who is always an inferior player in this case.
---
Note: A possible conclusion of course is that +1 acceleration or +2 anticipation and so forth is just very important, but what you'd be missing is that I'm not just comparing these two players together, I'm comparing them to a dozen others, who each all have to align with the rest of each other, in line with the objective performance results. So if I weight anticipation more say, that would throw 3 others out of whack with each other. There is no simply no way I can see that Lewandowski is better than Kane, and yet after 6 seasons of testing for each, Lewandowski in fact is. The only thing that seems possible is that passing and perhaps other attributes Kane is better in, are having a negative effect on his performance.
Then again, it could be that we just need a few more samples to set things straight. And here we are talking the minor differences of who is say 7th and who is 6th.
Fraudiola01 said: Guys I am a returning player, I already have FM24, but I can get FM26 for cheap. How good is the ME compared to FM24? I know the UI sucks but I’ve heard the Tactical Creator gives you more options due to the OOP/IP formations and that you can play more varied styles of play. Is that true? Expand If you mean the ME under the hood rather than the visuals, from what I've seen and read, it's about the same, the new OOP/IP is almost placebo, and it gets dominated by strikerless tactics.
Yarema said: I think FM-Arena attribute tests for the past few editions show linear effect for pretty much all of the attributes. Expand HarvestGreen, Orion, and my own findings don't though.
And the effects can be pretty massive not subtle, for example Orion finds pressure 1 to 10 is +47.33% but pressure 10 to 20 is +10.18%, and HarvestGreen confirmed this with pressure 1 to 6 18.2% and pressure 6 to 18 8.9%. I've found the same.
I think the indisputable contribution of this data is that it tells us what matters in FM26, and it seems like there's not much difference to FM24.
I don't know why you're finding every attribute effect to be linear, I think something going wrong somewhere in regards to that. I suspect it could be that you're randomizing attributes of every player at every club in the league, so that there's no longer set levels to compete against (i.e. in a real league, most players will have ~14 pace for example, or most GKs will have ~14 aerial reach, producing thresholds - it's not necessarily innate to the attribute itself, but a product of the relative competition. So 13 > 14 pace isn't a threshold in Vanarama South, but it is in Premier League). It's not entirely clear though if you're randomizing every player at every club, or just one club.
The part that intrigues me the most are the position weights. Have you simply divided the general results by the CA costs for each position, or have you done comparative testing for each position here?
Although I don't agree with the reasoning given as critique of my weightings, I do believe the underlying idea that the precise rankings may be somewhat impaired.
I attempted to address this with weighting the weights, but it produced worse results, or at least no better.
I've tried comparative deduction by comparing several players match ratings & goal tallies with their attributes. Some key problems with this approach:
- Match ratings have been demonstrated to favor presence of technical/mental ability over actual wins - Goals scored can simply reflect finishing, and poor finishing has minimal effect on the team win rate (it simply means someone else with higher finishing in the team scores the goals) - I actually managed to get the players to 'line up' with the ratings/goals a lot better, but it took a very strange set of GS weightings to get there. Either I'm on the wrong track, or even if it's right, I would be building a house without being able to know the foundations.
I concluded that what's actually needed is more result-based data, but on individual players. It's more tedious, and will require a fair amount of samples for good accuracy, but I think it's the best approach.
Method
I use my meta team at Man City, but replace a position with duplicates of a particular player. If I'm testing ST, I remove proficiency in other positions so as not to mess up the results, as what I'm most interested in is deducing what constitutes the precise ranking of STs. Full detail is used for all first team competitions.
Results
Here's some data I've collected so far (ranked by team performance with player as ST):
The reason I picked Le Fondre to include is because I saw him compared in a Youtube video, and even though he's a 36-year-old with 9 pace/acc, he managed the results you see here. A takeaway from this for me is that pace/acc isn't strictly necessary. Using the comparative deduction I mentioned before, I found that anticipation and composure are more critical than pace/acc.. as I said, these are strange conclusions, so rather than make the facts fit the theory, I'm going to be getting the facts straight first and then seeing what theory fits the data.
I'll probably do about 5-6 samples each of ~10 players for ST and then see if I can deduce consistent patterns to create accurate weightings.
iezzex said: So heres my question If i still win, but i wanna find a keeper that actually saves, what i shall look for Rn Chevalier looks good for me, best keeper for me yet, but he's old in my save, so i wanna find or make a good regen, waht to look for? Expand I may have a look into it later, but I suspect that it's fundamentally your tactic and overall level of quality of your players, not the attributes of your GK, that determine the number of goals he concedes.
We can influence who scores goals in a team, but that's because at the end of the day a goal is going to be scored. But you can't reduce the number of goals your GK concedes, because he's the only one who can concede them.
iezzex said: Is it all about team overall performance or not? Cuz what i've seen on my long real save, as example, Courtouis were bad (with my high def line at least) Yeah i won alot, and gd was insane, but he's avg rating was like 6.8 or something like that haha Expand To clarify, my template, and therefore the weightings, are based on team position at end of season, not match ratings.
On 'no detail' processing, possibly.
On 'full detail', can't remember now, but I don't think so. I know for certain that player rep doesn't. I don't think club attributes affect the results either, but I suppose it does come into consideration when AI managers are deciding what team to play against you (you know, how after mid-season, suddenly games get a lot harder once they adjust to your actual team quality).
If you ignore 6 > 10 for dribbling, then dribbling appears just about linear as well:
Dribbling
1 > 6 = +15.1 (+3.02/extra point)
6 > 10 = +2.7 (+0.675/extra point)
10 > 15 = +19.5 (+3.9/extra point)
15 > 20 = +14.0 (+2.8/extra point)
You choose to see what you want to see. I choose to see what is actually there, and what I see are two different attributes, showing the same pattern of non-linearity at the 6-10 level, which reduces quite significantly the chance that it is a statistical error.
Can't be bothered digging it up, but there's more deductions you can make like this for other attributes by comparing his 6 > 18 data to his 1 > 20, 10 > 20 data and whatnot
Obviously we aren't looking for the 6th decimal, there is margin of error. And they might even behave slightly nonlinearly but a linear function can be a very good approximation to reduce complexity.
Acceleration 1 > 6 = +42.7 (+8.54/extra point)
Acceleration 6 > 10 = +58.6 (+14.65/extra point)
The non-linearity is not as profound as for work rate, but it is nonetheless significantly non-linear.
Additionally I would posit that the league quality itself introduces a further level of non-linearity (or perhaps the non-linearity is a reflection of it, and would be more pronounced in a realistic league). HarvestGreen used a league that from memory had players with 10 in everything. In the Premier League, you need at least ~13 acceleration just to survive, so I would expect to see a large disparity between say acceleration 6 > 10 and acceleration 10 > 14 in the Premier League.
harvestgreen22 said: Non linear
Its influence is relatively significant from 1 to 8, but it becomes much smaller from 14 to 20
If he has, for instance, 16 ，18 , or even 20, it might waste a considerable amount of ability, but the effect actually only increases very little
As for Mbappe, it's true I underestimated him a little as ST. I had him as 12th; in the latest best fit to the objective results I'm working on, he's currently 6th. The difference between us is not that I don't make mistakes and you do, but that I correct my mistakes, while you continue to insist that Mbappe is equal to Haaland even when reality shows otherwise time and time again.
And no, I'm not going to remove the results of your file from my list, so stop DM'ing me about it.
I can't work out if you're stupid or if you think we're stupid? Maybe both?
As for being a liar when you released premier league 1 you didn't set any hidden attributes. Therefore genie scout wasn't producing skewed numbers. Yet you repeatedly lied that genie scout was at fault for giving Mbappe such a low score. Again this was a lie it was your appalling ratings.
What actually first came to my mind was HarvestGreen's excel tables, which show the non-linear effect of multiple attributes very clearly and in detail. Look at the right side:
Here we see that speed, work rate, among others, have non-linear effects.
Dribbling follows more of an inverted U-shaped curve, so in practice is mostly linear, but not always. I've always said that myself:
GeorgeFloydOverdosed said: HarvestGreen22, using a very artificial setup method (all attributes 10 bar the one being tested) and assessing without regard for position, claims dribbling is 4th most impactful attribute and follows a largely but not entirely linear benefit through 1-20.
You've got the GS narrative backwards. I created the Premier League 1.0 weightings (without releasing them), then was informed by someone else that there was a problem with GS, and deduced it was the hidden attributes causing the majority of the issues because GS doesn't calculate things correctly with them. The weights work as intended in other programs (i.e. FM26 scoring system), including the hiddens (i.e. FM SuperScout). When I released Premier League 1.0 for GS, I removed the hiddens so it could work as best it could, since we can't really do away with GS entirely yet.
Objective testing has shown now that Mbappe is worse than 34-year-old Lewandowski and 36-year-old Messi gives him a run for his money. Mbappe is no where near Haaland on any measure - team position, goals, rating. And your own rating file turned out be worse than FM Genie Scout's default on key measures. Your file shows that apparently Dybala is as good as Lewandowski or Kane 🤡
So as I posted in another thread, I've ran some controlled tests of specific players and found they finish in this ranking order (ranked by team position):
Erling Haaland = 1.666
Robert Lewandowski = 2.666
Kylian Mbappe = 3
Daizen Maeda = 3
Harry Kane = 3.5
Dusan Vlahovic = 3.571
Victor Osimhen = 3.833
Lionel Messi = 4
Donyell Malen = 4.666
Robert Glatzel = 6.5
Paulo Dybala = 6.833
Adam Le Fondre = 7.5
6 full season samples each, all relevant competitions on full detail, and this isn't your run-of-the-mill plop Haaland at Arsenal and see how he goes - the rest of the team are all identical non-exceptional players in order to minimize variation while remaining realistic and tested to finish ~5 on average by themselves (we replace the STs with 4 duplicates of the player we want to test)
If you use Falbraav's file which puts the weightings into Genie Scout, it doesn't align with the real results at all:
If we remove the hiddens, which are known to be problematic in GS, not much changes. Lewandowski remains behind Mbappe & Osimhen; Maeda remains near the bottom.
The rankings remain incorrect if we plug Falbraav's GS figures into FM26 scoring system (to rule out GS being the problem):
Haaland = 16.8
Mbappe = 16.1
Osimhen = 15.6
Lewandowski = 14.5
Maeda = 14.3
Messi = 14.1
While I was doing this and looking at OP's table, I realized what could be a big part of the problem. He's using points to measure success rather than team position. Points are very unreliable, because they can be 57 for one season sample and 83 the next - what doesn't change much is the team position, so that team will likely finish 2nd both times even though the point tally is all over the place. You would still get some valid correlations obviously, i.e. no pace/acc = no points at all.
I suggest to the OP if he intends to have another crack at it, to assess by the team position and not the team points.
And another problem with points vs team position could be that it neglects defensive advantages. It could explain why positioning does so terribly, for perhaps better defense favors draws as opposed to victories and losses - but a draw is only worth 1 point; a win 3. We want victories anyway, so this doesn't matter all too much either way.
He said it here, about Work Rate. Not sure about the other attributes, but I remember reading that and I've had a filter set at 6+ Work Rate ever since.
Lowering the Work rate 1 ,
it have very serious consequences (goal difference -110),
and it should be ensured that a player has at least 6 and preferably 10 Work rate.
A player with a Work rate is not desirable.
But continuing to increase Work rate (from 10->20) is low effect. The difference between Work rate 10 and Work rate 20 is relative small.
The 6-point attribute seems to be the threshold for some attributes.
Core attributes are these:
Special : Work rate (need to reach 10, higher is useless)
keithb said: Well if thats all he said then he didn't say it did he. But no surprise if Bozo is quoting that. He's a liar.
Amazing.
What you can do for attributes that have minimal or no value above a certain point is to set it's value to 0 (or whatever minimal value the calculations show) with a caveat that it only applies above X and that you need to filter for it. I am not disputing that some attributes might act that way, but that assigning them some value (by feel?) is misleading. And again FM-Arena tests showed that going from 18 to 19 to 20 pace is more or less just as valuable, it keeps it's scaling. Can you win with 18? Sure, but 20 is still better.
With respect, my post was my answer to that very question.
You could argue the method is inadequate, but you can't say it achieves nothing at all, and I'd challenge you with the following question: If not this method, then what? Assume attributes are linear and get assessments that are even more wrong?
It's true that filters can be part of the solution, and I've suggested to people before to use filters in combination with GS weightings if they can, but obviously this isn't a very good solution. Filtering is too blunt a tool for one thing; it's better to use weighted weightings instead usually.
I don't believe in FM Arena testing much, because they obviously contradict the findings of HarvestGreen, Orion, and myself. Even aside from the results, I didn't like some of the methodology that I could see, in the same way that I don't like how HarvestGreen uses an artificial league where everybody has the same attributes, instead of using a real league like the Premier League.
We see something similar reveal itself in this guy's testing. I would characterize it as a kind of blase 'let's apply a python script to this' mentality. Sometimes, the kind of info that gets can be very useful, but I think more typically it leads to skipping over the necessary nuances that only someone who has waded deep into it with hours of trial and error and reflection and critical thought can discern.
For example, he says his results show that attributes are linear. Ok, so why hasn't he reflected and deduced why this contradicts HarvestGreen's findings? The most likely answer is that he simply hasn't put in the time in doing that. By his own account the data was from an '8 hour overnight session', and I don't know how long he spent on it overall, but my impression is that this is probably something he's done over a week or two.
I'm not trying to throw shade, and it's no crime to not spend your life on analyzing FM, I'm just saying it's an approach that has certain downsides. My downsides are that I'm not a programmer, so I can't use such a methodical and fast approach, and because I'm always coming up with different ideas, I don't say 'this is the result I got, that's it from me'.
Now you say that assigning values by feel is misleading. Well let's look at a concrete example:
I brute force test players until I know who ranks where.
Starting with informed guesses (i.e. start pace at '100' not '0', and passing at '0' not '100' ), I adjust the weightings in Genie Scout until the player rankings all correspond exactly to their actual performance ranking.
Because some or many attributes are not linear, there may be a 1% error here, a 3% error there, and even a handful of ranks out of the thousands out of alignment, but for the most part, it does its job.
I could continue the example further, but I think the point is already clear - where exactly is the error or impossibility in it?
I'll be posting it soon, but it's still a work in progress right now.
Translating the real results to Genie Scout has been proving to be really tricky. As ChatGPT has confirmed to me, there's no straightforward way of deriving the specific combination of attribute weightings from the results alone. For example say I find drib 30 + ant 30 + pace 70 is what works.. how do I know it's not just ant 30 + pace 70, and dribbling has 0 effect?
Thankfully I can rule out a lot of attributes from the get-go from previous testing, but I still have to just manually fiddle around with all sorts of different combinations to find out what works and what doesn't until I have something that at least aligns with the real rankings as closely as possible.
I started with the following process:
1. Start with Premier League 1.0 'weighted weights' version
2. Compare players who have completely different sets of attributes but had similar results
3. Adjust weighting of 1-3 attributes that I know matter (i.e. pace, anticipation, work rate, etc.) starting with the most likely attribute (i.e. pace) until all players are roughly where they should be ranked.
4. If one player's rank is fixed, but then another falls out, return to step 3.
5. Find a player with high rep or high CA who has unusually low % rating with the changed weights. Test performance in controlled test first, then do step 3 with them if performance doesn't match rank.
6. Keep repeating step 5.
Right now I've got a functional set of weights, but everything has to be changed around again the next time a sample finds a player changes their ranking. I think 6 samples is giving us a fairly accurate picture now to work with, but for true accuracy I think it's going to require at least 10.
Then there is the matter of position variation. But my first result in another position so far seems to suggest the weights could be the same for every outfield position. Once I've done all positions, with perhaps 10 samples for STs for calibration, then I'll post it.
I mean it's kind of the basis of all these weighting systems, because if a point of pace was vastly different from 9 to 10 as opposed to 15 to 16 there would be no point in having a fixed value for an attribute weight.
So you either accept they are pretty much linear (with a few exceptions) or throw GS ratings out the window.
It's not the basis of my weightings, as non-linear values do not make useful weightings in Genie Scout impossible.
Here's how I handle the problem of non-linearity:
Finishing - I know from testing it has no advantage above ~7, and is only of moderate value even below that all the way down to one. Therefore I give it a very low weight, or perhaps even no weight at all.
Work Rate - It's very clear that work rate is crucial going from 1 to 6 but has greatly diminishing returns beyond that, and pretty much tops out at around ~11-13. If you don't believe the testing done on this, just look at player like Haaland, Mbappe and Messi - their work rate is 13, 12, and 9 respectively. The compensation in GS for this is fairly straightforward, instead of weighting Work Rate equal to pace or acc if we were just looking at the 1-to-6 effect, we give it around 20-40 instead of 70-100.
Jumping Reach - This has one of the most difficult curves of all to accommodate. It is very important going from 1 to ~7, then moderately important ~8-13, has a massive jump in performance once going above the league threshold (~14-17 for Premier League), but is then near useless going from ~18 to 20 since probably no other player will challenge above 17 anyway. Here we target either ~8-13 (moderate rating), or the ~14-17 threshold (high rating, perhaps the highest even above pace/acc).
Of course you run into situations where GS considers 1 pace 20 acc equal to 10 pace 10 acc, but those 1 pace 20 acc players are the exceptions not the rule, so it doesn't render the whole enterprize futile.
No Mbappe and Haaland are not rated the same. And when you increase TS to 107, like I said in the post, the gap becomes even bigger. Its 2% and then becomes 4%. They are not rated the same. Is there actually something wrong with you?
AGAIN how can you create a file and use and not see this?? There's a worrying disconnect.
To be honest I didn't even check if you had a TS rating, because I took one look at the FS rating and knew it was rubbish.
The screenshot I provided is, you'll agree, the FS ratings you've given.
So do you want me to use the FS or TS rating?
Looking at your TS ratings, you have Osimhem above Lewandowski even though both having 15 jumping reach, so more of the same 💩 either way.
I wouldn't want to misrepresent you. This is what your GS file says does it not?
Default database, start of game (England - 3/7/2023)
90% | 90% | 90% | 90% | Erling Haaland - 1st, 1st, 4th, 1st, 2nd, 1st = 1.666 position, 1.119 goals/match, 7.80 rating
84% | 84% | 80% | 79% | Robert Lewandowski - 2nd, 2nd, 2nd, 3rd, 5th, 2nd = 2.666 position, 0.721 goals/match, 7.33 rating
84% | 81% | 90% | 87% | Kylian Mbappe - 2nd, 3rd, 3rd, 3rd, 5th, 2nd = 3 position, 0.701 goals/match, 7.48 rating
80% | 69% | 76% | 72% | Daizen Maeda - 2nd, 3rd, 4th, 4th, 2nd, 3rd = 3 position, 0.586 goals/match, 7.23 rating
83% | 84% | 80% | 78% | Harry Kane - 5th, 2nd, 3rd, 3rd, 1st, 7th = 3.5 position, 0.563 goals/match, 7.29 rating
80% | 81% | 77% | 78% | Dusan Vlahovic - 1st, 4th, 4th, 5th, 1st, 6th, 4th = 3.571 position, 0.634 goals/match, 7.23 rating
80% | 81% | 84% | 85% | Victor Osimhen - 3rd, 3rd, 7th, 3rd, 3rd, 4th = 3.833 position, 0.714 goals/match, 7.29 rating
80% | 78% | 85% | 80% | Lionel Messi - 4th, 2nd, 1st, 5th, 9th, 3rd = 4 position, 0.588 goals/match, 7.16 rating
75% | 72% | 77% | 74% | Donyell Malen - 2nd, 2nd, 7th, 5th, 5th, 7th = 4.666 position, 0.514 goals/match, 7.22 rating
75% | 74% | 69% | 72% | Robert Glatzel - 5th, 7th, 6th, 3rd, 12th, 6th = 6.5 position, 0.634 goals/match, 7.16 rating
72% | 72% | 79% | 71% | Paulo Dybala - 5th, 5th, 6th, 10th, 3rd, 12th = 6.833 position, 0.492 goals/match, 7.07 rating
61% | 63% | 62% | 61% | Adam Le Fondre - 8th, 11th (sacked), 5th, 7th (sacked), 6th (sacked), 8th (sacked) = 7.5 position
Edit: I've added a color-coded column to compare 4 different GS rating files. 1st is new Best fit file, 2nd is Premier League 1.0, 3rd is keithb, 4th is Falbravv's META FMA 2.0 file based on skawkclsrn's recent data. Green = 1% or less difference. Yellow = 2-4% difference. Red = 5%+ difference. Ignore Harry Kane, he needs a slight readjustment right now due to the latest samples.
There are a lot of new deductions that can be made.
One is that now we have a rough idea what GS % corresponds to in terms of concrete performance numbers, and another is that we can tell at what point random variation makes distinguishing between % numbers pointless.
We can always do with more samples, but if we just take the difference between whoever is highest and lowest between Lewandowski, Mbappe, Maeda, Kane, Vlahovic, Osimhen, and Messi (total +/- 1 position difference maximum) we get:
((4.08/80.70)x100) = 5.06% - My best fit of the data so far in GS (not to minimize %, but to correct player rank)
((14.30/91.99)x100) = 15.54% - Premier League 1.0 'weighted weights' version
((14.94/88.11)x100) = 16.96% - Premier League 1.0
((14.14/92.25)x100) = 17.14% - 'Pure Performance' FM24
((13.45/77.59)x100) = 17.33% - My first/original FM24 ratings file
((15.89/84.94)x100) = 18.71% - 'Blended' FM24
((18.34/83.29)x100) = 22.02% - ykykyk05251's weights file
((22.88/88.01)x100) = 26.00% - Orion's coefficients file
((29.27/91.47)x100) = 32.00% - Genie Scout Default
That is where we are accounting for extreme cases. In other words, a 4-5% difference in GS is almost certainly an objectively better player.
But what if we make it a bit leniant, and tolerate ~10-20% being exceptions we miss. In other words, this is the more typical uncertainty range. To discover this, I exclude the bottom % player from the list:
((4.03/80.70)x100) = 4.99% - My best fit of the data so far in GS (not to minimize %, but to correct player rank)
((6.21/88.11)x100) = 7.05% - Premier League 1.0
((6.00/83.29)x100) = 7.20% - ykykyk05251's weights file
((7.34/75.95)x100) = 9.66% - Falbravv's 'META FMA 2.0' file based on skawkclsrn's data this week
((10.39/89.51)x100) = 11.60% - Falbravv's 'META FMA' file based I think FM Arena's testing(?)
((11.67/91.47)x100) = 12.76% - Genie Scout Default
((13.45/77.59)x100) = 13.20% - My first/original FM24 ratings file
((12.61/95.25)x100) = 13.24% - 'Pure Performance' FM24
((12.42/91.99)x100) = 13.50% - Premier League 1.0 'weighted weights' version
((13.21/94.59)x100) = 13.97% - keithb's weights 💩
((13.64/88.01)x100) = 15.50% - Orion's coefficients file
((14.14/84.94)x100) = 16.65% - 'Blended' FM24
So we see that whatever the case, ~4-5% is going to be the range within which the player who is better becomes uncertain, but also the difference is going to be pretty meaningless objectively.
I can answer this for you because I'm currently doing isolated testing and analyzing the attribute differences closely for a number of players, and one of them is Harry Kane.
So far my testing finds that Kane is in fact objectively one of top STs in FM24.
I've already come up with a draft set of weights that allows Kane's low acc/pace and other quirks to align smoothly with all other players tested so far, but I'll just tell you a more abstract version of what I'm seeing.
This is my impression so far (for ST at least specifically):
1. Pace/acc, and stamina to keep it going through a game, are still the top factors.
2. Pace/acc is not *strictly* necessary to do well. A player with 9 acceleration can function decently as ST in the Premier League (I've tested this). If you have a team of them, you'll be relegated, but as individual exceptions to the rule they can do the job. Although I haven't tested it, I would expect similar with a team of Harry Kanes. You need at least ~13-14 average for the team as a whole in Premier League just to survive. So pace/acc remains very important.
3. Only about one third of attributes matter. Anticipation and dribbling are in this third for ST, but all of these are at least a bit less crucial than pace/acc. For instance, Messi has 20 dribbling, and while he is not no.1, it does appear to be a key factor in him being one of the top STs according to testing even at age 36 in FM24. None of these attributes are strictly necessary either except perhaps 1 or 2, see: Daizen Maeda
4. This is still dubious, but through analyzing examples such as Kane vs. Lewandowski specifically, it would appear that some attributes such as passing may actually have negative performance impact.
Here is something I already wrote that I was thinking of posting earlier:
There is no reason why Lewandowski should be better than Harry Kane, and yet he is. In fact, Lewandowski is slightly inferior on a number of things.
Lewandowski has advantages in the following areas:
+3 aggression
+3 first touch
+3 balance
+2 anticipation
+2 concentration
+2 natural fitness
+1 bravery
+1 off the ball
+1 acceleration
+1 agility
+1 strength
+1 important matches
None of those are adequate to explain the difference, given Kane has the following advantages (I'll list just a few):
5 years younger
+7 crossing
+5 passing
+5 vision
+4 teamwork
+3 left foot
+2 consistency
+1 determination
+1 composure
+1 work rate
I suspect that high passing and possibly crossing reduces team performance. Perhaps what is happening in the engine is that Kane is passing to a team-mate do to the job, who is always an inferior player in this case.
---
Note: A possible conclusion of course is that +1 acceleration or +2 anticipation and so forth is just very important, but what you'd be missing is that I'm not just comparing these two players together, I'm comparing them to a dozen others, who each all have to align with the rest of each other, in line with the objective performance results. So if I weight anticipation more say, that would throw 3 others out of whack with each other. There is no simply no way I can see that Lewandowski is better than Kane, and yet after 6 seasons of testing for each, Lewandowski in fact is. The only thing that seems possible is that passing and perhaps other attributes Kane is better in, are having a negative effect on his performance.
Then again, it could be that we just need a few more samples to set things straight. And here we are talking the minor differences of who is say 7th and who is 6th.
If you mean the ME under the hood rather than the visuals, from what I've seen and read, it's about the same, the new OOP/IP is almost placebo, and it gets dominated by strikerless tactics.
HarvestGreen, Orion, and my own findings don't though.
And the effects can be pretty massive not subtle, for example Orion finds pressure 1 to 10 is +47.33% but pressure 10 to 20 is +10.18%, and HarvestGreen confirmed this with pressure 1 to 6 18.2% and pressure 6 to 18 8.9%. I've found the same.
I don't know why you're finding every attribute effect to be linear, I think something going wrong somewhere in regards to that. I suspect it could be that you're randomizing attributes of every player at every club in the league, so that there's no longer set levels to compete against (i.e. in a real league, most players will have ~14 pace for example, or most GKs will have ~14 aerial reach, producing thresholds - it's not necessarily innate to the attribute itself, but a product of the relative competition. So 13 > 14 pace isn't a threshold in Vanarama South, but it is in Premier League). It's not entirely clear though if you're randomizing every player at every club, or just one club.
The part that intrigues me the most are the position weights. Have you simply divided the general results by the CA costs for each position, or have you done comparative testing for each position here?
I attempted to address this with weighting the weights, but it produced worse results, or at least no better.
I've tried comparative deduction by comparing several players match ratings & goal tallies with their attributes. Some key problems with this approach:
- Match ratings have been demonstrated to favor presence of technical/mental ability over actual wins
- Goals scored can simply reflect finishing, and poor finishing has minimal effect on the team win rate (it simply means someone else with higher finishing in the team scores the goals)
- I actually managed to get the players to 'line up' with the ratings/goals a lot better, but it took a very strange set of GS weightings to get there. Either I'm on the wrong track, or even if it's right, I would be building a house without being able to know the foundations.
I concluded that what's actually needed is more result-based data, but on individual players. It's more tedious, and will require a fair amount of samples for good accuracy, but I think it's the best approach.
Method
I use my meta team at Man City, but replace a position with duplicates of a particular player. If I'm testing ST, I remove proficiency in other positions so as not to mess up the results, as what I'm most interested in is deducing what constitutes the precise ranking of STs. Full detail is used for all first team competitions.
Results
Here's some data I've collected so far (ranked by team performance with player as ST):
Erling Haaland 👍 - 1st, 1st, 4th = 2nd position, 1.263 goals/match, 7.93 rating
Kylian Mbappe 💩 - 5th, 1st, 3rd = 3rd position, 0.708 goals/match, 7.54 rating
Daizen Maeda - 2nd, 3rd, 4th = 3rd position, 0.533 goals/match, 7.28 rating
Harry Kane - 5th, 2nd = 3.5th position, 0.429 goals/match, 7.25 rating
Victor Osimhen - 3rd, 3rd, 7th = 4.333th position, 0.699 goals/match, 7.27 rating
Adam Le Fondre - 8th, 11th (sacked), 5th = 8th position, 0.416 goals/match, 6.82 rating
The reason I picked Le Fondre to include is because I saw him compared in a Youtube video, and even though he's a 36-year-old with 9 pace/acc, he managed the results you see here. A takeaway from this for me is that pace/acc isn't strictly necessary. Using the comparative deduction I mentioned before, I found that anticipation and composure are more critical than pace/acc.. as I said, these are strange conclusions, so rather than make the facts fit the theory, I'm going to be getting the facts straight first and then seeing what theory fits the data.
I'll probably do about 5-6 samples each of ~10 players for ST and then see if I can deduce consistent patterns to create accurate weightings.
If i still win, but i wanna find a keeper that actually saves, what i shall look for
Rn Chevalier looks good for me, best keeper for me yet, but he's old in my save, so i wanna find or make a good regen, waht to look for?
I may have a look into it later, but I suspect that it's fundamentally your tactic and overall level of quality of your players, not the attributes of your GK, that determine the number of goals he concedes.
We can influence who scores goals in a team, but that's because at the end of the day a goal is going to be scored. But you can't reduce the number of goals your GK concedes, because he's the only one who can concede them.
Cuz what i've seen on my long real save, as example, Courtouis were bad (with my high def line at least)
Yeah i won alot, and gd was insane, but he's avg rating was like 6.8 or something like that haha
To clarify, my template, and therefore the weightings, are based on team position at end of season, not match ratings.