Orion said: The second one - as said in disclaimer - uses Jumping Reach to power of two. So just take Jumping Reach x Jumping Reach x the last coefficient. Polynomial regression can show features 'synergie' that is represented as two features multiplication - or one feature multiplied by itself. Expand
Amazing, thanks so much for taking the time there to keep me right on this
Middleweight165 said: Can you say the same thing again but pretend you are explaining it to a 5 year old? Expand
The training data was transformed to show player attributes not as 1-20 but I calculated average value for each attribute for every league. Then for every player I changed his attribute value as (actual_attribute_value - average_attribute_value_in_player_league). This way attributes can also be negative - if player has lower attribute value than average for his league. And that was the data that I used to training the model. So model says 'hey, take this 'new attribute value' for the player and multiply it by his coefficient'. Because of this in theory we should check in what league you are playing, check average attributes values for the players in that league, calculated this new 'difference' attribute and then use coefficient to calculate player 'rating'. To make it easier, and because we compare players that will play in the same league, we will just use original attribute values of players. This may introduce some additional error in the rating calculation, but since this makes calculation so much easier we accept this flaw of the method.
CBP87 said: @Orion Are you saying if a FB has 15 pace, then the way to work out the rating is 15 x 15 x 0.018991??? Expand
To get player rating/score you use all the coefficients so you make a sum Pac x 0,018991 + Jum x 0,014582 + Acc x 0,012670 + Ant x 0,012269 + Cnt x 0,009911 + Dri x 0,008246 + Cmp x 0,007720 + Cro x 0,006812. And then you can compare this 'sum' between the players. In theory the higher the sum the better the player should be. Additionally when you take into consideration this test you can include in your filter out players with 6+ or 10+ Work Rate since very low Work Rate results in extreme decrease in players performance.
Orion said: The training data was transformed to show player attributes not as 1-20 but I calculated average value for each attribute for every league. Then for every player I changed his attribute value as (actual_attribute_value - average_attribute_value_in_player_league). This way attributes can also be negative - if player has lower attribute value than average for his league. And that was the data that I used to training the model. So model says 'hey, take this 'new attribute value' for the player and multiply it by his coefficient'. Because of this in theory we should check in what league you are playing, check average attributes values for the players in that league, calculated this new 'difference' attribute and then use coefficient to calculate player 'rating'. To make it easier, and because we compare players that will play in the same league, we will just use original attribute values of players. This may introduce some additional error in the rating calculation, but since this makes calculation so much easier we accept this flaw of the method. Expand
Btw how does this work with double attributes? Like Jum^. If the model uses adjusted values then squaring the full number seems odd and potentially overshadowing everything else.
Yarema said: Btw how does this work with double attributes? Like Jum^. If the model uses adjusted values then squaring the full number seems odd and potentially overshadowing everything else. Expand
In linear regression you basically check every single feature and select the one that fits to the data the most. Then you exclude that one and check the model with this first feature + the second one. In polynomial regression you allow model to look for 'single' feature but also combination like feature1 x feature2 or feature1 x feature1. If it fits the data better than 'single' feature it will pick that feature. If 'single' feature fits the data better than 'combined features' it will keep the single one. In this model I allowed searching for quadratic equations. And for some cases it turned out that at some point, like for the Forward, Jum x Jum fits the data better than any single feature that was left in the pool. But for example Jum x Jum gives worse fit that just Jum.
They used different methodology. Keep in mind that for a striker coefficient for Jumping Reach is ~0,024 while for Pace it's ~0,021 so high pace is 'almost' the same 'good' as high Jumping Reach but having both high is even better. They use very different approach to the testing. If they proved that Jumping Reach above 17 doesn't give significant increase in the outcome so it be. If you want to be precise in model results interpretation it basically says that every single point of Jumping Reach higher than the average for a certain league gives a striker +0,024768 for his average rating. So this +3 point difference (20 compared to 17) gives around +0,074 to average rating. If in reality it doesn't it's not much off. Especially if we consider that it represents features mostly in linear way. I was really surprised that basically a linear regression model was able to be that accurate in this particular application.
Orion said: In linear regression you basically check every single feature and select the one that fits to the data the most. Then you exclude that one and check the model with this first feature + the second one. In polynomial regression you allow model to look for 'single' feature but also combination like feature1 x feature2 or feature1 x feature1. If it fits the data better than 'single' feature it will pick that feature. If 'single' feature fits the data better than 'combined features' it will keep the single one. In this model I allowed searching for quadratic equations. And for some cases it turned out that at some point, like for the Forward, Jum x Jum fits the data better than any single feature that was left in the pool. But for example Jum x Jum gives worse fit that just Jum. Expand Yes but the model was working with adjusted numbers. Let me give an example. Average JR in the league is 12, a striker has 17. The way you explained stuff in the original post and later on basically says you take 17-12=5 so you do the calculations with 5. And isn't it possible that quadratic fits well for adjusted number (5) but not the actual number 17? Squaring numbers like 5 is a lot different that 17. Above 8 it basically dominates every other attribute basically double dipping (or triple since it's also the best attribute for a striker anyway).
Edit: Or to put it another way. Comparing JR contribution in above scenario if striker has 17 or 18 JR. Taking numbers at face up value you'd get roughly 1,27 vs 1,40. So a difference of 0,13 which would roughly be the effect of +6 pace. Using adjusted numbers of 5 and 6 makes the scores 0,49 and 0,55 respectively, a difference of 0,06 so half of the previous difference. Still equals +3 pace which feels a lot.
What made you think that the thread indicates Jumping Reach is capped in the sense that it’s not effective beyond 17? The explanation actually says they deliberately limited it to 17 in the test to avoid an unrealistic scenario, not that its effectiveness is capped in gameplay.
CBP87 said: @Bafici can you share how you got those GS ratings please? It might just be me but I can't make sense of the Coefficient scoring on the opening thread Expand
I match the highest number to 100 and scale the rest according to that. Like this for example.
GK Agi (0,014640/0,014640)*100 = 100 Ref (0,012837/0,014640)*100 = 88 Aer (0,011812/0,014640)*100 = 81 Thr (0,007465/0,014640)*100 = 51 Com (0,007436/0,014640)*100 = 51 Han (0,006255/0,014640)*100 = 43 Dec (0,005089/0,014640)*100 = 35
Yarema said: Yes but the model was working with adjusted numbers. Let me give an example. Average JR in the league is 12, a striker has 17. The way you explained stuff in the original post and later on basically says you take 17-12=5 so you do the calculations with 5. And isn't it possible that quadratic fits well for adjusted number (5) but not the actual number 17? Squaring numbers like 5 is a lot different that 17. Above 8 it basically dominates every other attribute basically double dipping (or triple since it's also the best attribute for a striker anyway).
Edit: Or to put it another way. Comparing JR contribution in above scenario if striker has 17 or 18 JR. Taking numbers at face up value you'd get roughly 1,27 vs 1,40. So a difference of 0,13 which would roughly be the effect of +6 pace. Using adjusted numbers of 5 and 6 makes the scores 0,49 and 0,55 respectively, a difference of 0,06 so half of the previous difference. Still equals +3 pace which feels a lot. Expand
Well, I guess this is derived from the idea of this experiment which is determining which attributes have the biggest impact on high match score. Hence, winning headers for ST has the biggest impact probably because striker has many oppurtinities to increase "key headers" stat during the match. Blind guess though.
kvasir said: What made you think that the thread indicates Jumping Reach is capped in the sense that it’s not effective beyond 17? The explanation actually says they deliberately limited it to 17 in the test to avoid an unrealistic scenario, not that its effectiveness is capped in gameplay.
Glennson said: So what do you guys think is the best genie ratings file to follow for the meta: this one from these findings or the findings from Zippo? Expand
Would like to know this aswell guys, great work btw!
Yarema said: Yes but the model was working with adjusted numbers. Let me give an example. Average JR in the league is 12, a striker has 17. The way you explained stuff in the original post and later on basically says you take 17-12=5 so you do the calculations with 5. And isn't it possible that quadratic fits well for adjusted number (5) but not the actual number 17? Squaring numbers like 5 is a lot different that 17. Above 8 it basically dominates every other attribute basically double dipping (or triple since it's also the best attribute for a striker anyway).
Edit: Or to put it another way. Comparing JR contribution in above scenario if striker has 17 or 18 JR. Taking numbers at face up value you'd get roughly 1,27 vs 1,40. So a difference of 0,13 which would roughly be the effect of +6 pace. Using adjusted numbers of 5 and 6 makes the scores 0,49 and 0,55 respectively, a difference of 0,06 so half of the previous difference. Still equals +3 pace which feels a lot. Expand
Ok. I get your point. You are right. In such case you can exclude feautres that are quadratic. As said in the topic the results are shown related to the average attributes in a league and in perfect scenario they should be applie that way. My proposal was to simplyfy players comparison. I'll try to figure something out for this scenario. I might even just run the model for those 2 positions again and allow only 'single' feautres since those are majority of the results anyway.
So I've been testing this on my journeyman save, previously I was a spreadsheet with calculations for attacking and defending based on attributes weighted by the FM Arena new attribute testing.
That got me three promotions and then three years title wins on the top league (Japan)
I'm currently in Korea and with the weighting based on Orions approach, I'm still winning at the same rate but the football seems better to watch and average player ratings are higher. I'm claiming more player & goals of the month.
In general, my players are slower than than what I would sign on my previous set up but using harvestgreen22 training, I could build up pace and acceleration over time if needed.
One thing that does feel off is the jumping reach ratings. I didn't understand the calculations at first so I'm obviously not an expert on any of this but my suspicions are you need one or two players with high ratings there to score from set pieces and it's those goals that boost the average rating rating - potentially any position could have a high attribute.
It might be along the same line of high free kicks/corners/penalty taking having an impact, but that's purely guesswork on my part
This was in my head for several weeks before this post reveal the answer. I sold my Youssef Chermiti who has 140CA with striker rating 79 and upgrading my Striker to Karim Konate 160CA with 84 striker rating. However, Konate never perform better than chermiti and i keep asking... how come? with new rating in this post, its clear that chermiti rating is 77%TS comparing to konate 73%TS.
RFC said: So I've been testing this on my journeyman save, previously I was a spreadsheet with calculations for attacking and defending based on attributes weighted by the FM Arena new attribute testing.
That got me three promotions and then three years title wins on the top league (Japan)
I'm currently in Korea and with the weighting based on Orions approach, I'm still winning at the same rate but the football seems better to watch and average player ratings are higher. I'm claiming more player & goals of the month.
In general, my players are slower than than what I would sign on my previous set up but using harvestgreen22 training, I could build up pace and acceleration over time if needed.
One thing that does feel off is the jumping reach ratings. I didn't understand the calculations at first so I'm obviously not an expert on any of this but my suspicions are you need one or two players with high ratings there to score from set pieces and it's those goals that boost the average rating rating - potentially any position could have a high attribute.
It might be along the same line of high free kicks/corners/penalty taking having an impact, but that's purely guesswork on my part Expand
Good to hear that it works for you. I'm really looking forward to feedback since as said, the model is one thing, we know how it works and in theory I can calculate the average error for the model but knowing that it actually works in 'real life' is something different.
In terms of rating - FM Arena test is based on team Points. harvestgreen22 test is using goal difference. And my model is using players rating as a target variable. Hence my model will look for attributes that increase rating, but not team performance per se. We know that in general average rating will be somehow connected with team results. And I know about a flaw of this model that since it looks for players for high rating it will usually prioritise offensive player but that's due to FM rating system.
pek80 said: This was in my head for several weeks before this post reveal the answer. I sold my Youssef Chermiti who has 140CA with striker rating 79 and upgrading my Striker to Karim Konate 160PA with 84 striker rating. Konate never perform better than chermiti and i keep asking... how come? with new rating in this post, its clear that chermiti rating is 77%TS comparing to konate 73%TS. Expand
Just to clarify. So at first you used GenieScout default 'rating' and then one from my model?
Orion said: Just to clarify. So at first you used GenieScout default 'rating' and then one from my model? Expand
Yes bro, the one in your model is the only "rating model" that rate Chermiti higher than Konate since all other model value pace acceleration, and finishing but not weight much in jumping reach!
pek80 said: This was in my head for several weeks before this post reveal the answer. I sold my Youssef Chermiti who has 140CA with striker rating 79 and upgrading my Striker to Karim Konate 160CA with 84 striker rating. However, Konate never perform better than chermiti and i keep asking... how come? with new rating in this post, its clear that chermiti rating is 77%TS comparing to konate 73%TS. Expand
Could Chermiti’s higher goal tally be because he was your team’s main aerial threat, with extra goals coming from set pieces?
These ratings have worked well for me too, though I find the high Jumping Reach requirement for full-backs a bit off-putting. Just finished my first season with Schalke and comfortably won the 2. Bundesliga—curious to see if I can win the top division as well. In a previous save using FM-Arena ratings, I nearly won La Liga with Alavés in the first season, finishing just one point behind a dominant Real Madrid, who only lost once.
pek80 said: Yes bro, the one in your model is the only "rating model" that rate Chermiti higher than Konate since all other model value pace acceleration, and finishing but not weight much in jumping reach! Expand
Just out of curiosity can you post the screenshot of attributes for both player?
kvasir said: Could Chermiti’s higher goal tally be because he was your team’s main aerial threat, with extra goals coming from set pieces?
These ratings have worked well for me too, though I find the high Jumping Reach requirement for full-backs a bit off-putting. Just finished my first season with Schalke and comfortably won the 2. Bundesliga—curious to see if I can win the top division as well. In a previous save using FM-Arena ratings, I nearly won La Liga with Alavés in the first season, finishing just one point behind a dominant Real Madrid, who only lost once. Expand
One of my wild guesses behind Jumping Reach being high rated attributed for other positions, beside obviously scoring a lot from set pieces, was that it has high correlation with Strength. So player with high jumping reach usually have also high height, high weight and high Strength. So maybe strength also plays part in it because player with high strength will be able to just push opposition player and get the ball back.
Thanks. I really looks on the first glance that Konate should be far superior to the Chermiti. Like really the only thing that Chermiti has better than Konate is solely jumping reach and nothing else.
Yarema said: Yes but the model was working with adjusted numbers. Let me give an example. Average JR in the league is 12, a striker has 17. The way you explained stuff in the original post and later on basically says you take 17-12=5 so you do the calculations with 5. And isn't it possible that quadratic fits well for adjusted number (5) but not the actual number 17? Squaring numbers like 5 is a lot different that 17. Above 8 it basically dominates every other attribute basically double dipping (or triple since it's also the best attribute for a striker anyway).
Edit: Or to put it another way. Comparing JR contribution in above scenario if striker has 17 or 18 JR. Taking numbers at face up value you'd get roughly 1,27 vs 1,40. So a difference of 0,13 which would roughly be the effect of +6 pace. Using adjusted numbers of 5 and 6 makes the scores 0,49 and 0,55 respectively, a difference of 0,06 so half of the previous difference. Still equals +3 pace which feels a lot. Expand
For features using multiple attributes my suggestion is to multiply them and then take a square before multiplying by coefficient. So when you have for MLR Acc, Agi and coeff 0,003329 do: SQR(Acc * Agi) * Coeff This way you can balance the impact of this double attribute multiplication and you keep the 'partial score' properly scaled so for example when I tried said MLR using the same attribute value for all features, the 'partial score' for the double attribute feature stayed proportional to it's coefficient compared to other features.
FREVKY said: That's what I thought, so I created pretty simple excel spreadsheet that make mass player comparison possible and easy.
Here's how it works: First, you need to import specified views. I created two sets of these: one for your team (for squad view) and one for scouting tab - both with CA & PA hidden - and another set with CA & PA visible for those who like to spoil the fun a little bit. Download and paste them into "views" folder in your Documents (C:\Users\your_name\Documents\Sports Interactive\Football Manager 2024\views is the default path).
When you load the view, you need to select every player, so click on one player and than ctrl+a to select everyone in the team or in the scouting range. Just bear in mind the more players you select, the more time it takes, so if you're about to select over 1000 players, give it a few seconds to work. Then, press cltr+p to "print" the selection into HTML file. Save it wherever you want, name it whatever you want. Then, you need the spreadsheet (MS Excel file). Open it and then in the Excel go to to File->Open and select the html file with your set of players. Copy it's whole contents (ctrl+a, then ctrl+c) and paste them into my spreadsheet in the blank sheet called "IMPORT" then switch the sheet to the one called "CALCULATION" and it should automatically calculate values for every player for each position using coefficients from this thread. Additionalli I added sections with CA, PA and difference between them (it will work only when you used views with PA and CA obviously).
Of course you can use whatever filters you want on the scouting section to narrow down the amount of players to whatever you really need.
At first glance it may sound complicated a bit but it's pretty easy to use. If you find any trouble using it, I'll try to help.
The spreadsheet is editable so do whatever you want with it, if you find any room for improvements, go for it. Expand
Bafici said: I made a genie scout file. Biggest difference from fmarena ratings is this is bit more realistic. Great players in real life takes higher ratings.
Just to keep you guys up to date I've decided to make a small adjustment for the features that use two attributes - so the last one for MLR and the last one for Striker. I've decided to take square of a product from multiplication those features (SQR(Attribute1 * Attribute2) * Coefficient). This way you can keep it proportional to the coefficient. I don't know how it works with your files. I assume that for spreadsheet it's not an issue. What about GenieScout? Does it support other mathematical operations? If not I'd suggest just to get rid of those 2 features (since they are last important for given positions) and adjust others to sum up to 100% just like you did for the other positions.
Just want to say how much I appreciate the work you put into testing and letting us see behind the curtain. I like the influx of new research regarding attributes and training and whatnot.
I just find it hard to wrap my mind around the fact that this is currently the best Left back I can have (playing as Malmö 3rd season)
Orion said: The second one - as said in disclaimer - uses Jumping Reach to power of two. So just take Jumping Reach x Jumping Reach x the last coefficient.
Polynomial regression can show features 'synergie' that is represented as two features multiplication - or one feature multiplied by itself.
Amazing, thanks so much for taking the time there to keep me right on this
Middleweight165 said: Can you say the same thing again but pretend you are explaining it to a 5 year old?
The training data was transformed to show player attributes not as 1-20 but I calculated average value for each attribute for every league. Then for every player I changed his attribute value as (actual_attribute_value - average_attribute_value_in_player_league). This way attributes can also be negative - if player has lower attribute value than average for his league. And that was the data that I used to training the model. So model says 'hey, take this 'new attribute value' for the player and multiply it by his coefficient'.
Because of this in theory we should check in what league you are playing, check average attributes values for the players in that league, calculated this new 'difference' attribute and then use coefficient to calculate player 'rating'.
To make it easier, and because we compare players that will play in the same league, we will just use original attribute values of players. This may introduce some additional error in the rating calculation, but since this makes calculation so much easier we accept this flaw of the method.
@Bafici can you share how you got those GS ratings please? It might just be me but I can't make sense of the Coefficient scoring on the opening thread
@Orion Are you saying if a FB has 15 pace, then the way to work out the rating is 15 x 15 x 0.018991???
CBP87 said: @Orion Are you saying if a FB has 15 pace, then the way to work out the rating is 15 x 15 x 0.018991???
To get player rating/score you use all the coefficients so you make a sum
Pac x 0,018991 + Jum x 0,014582 + Acc x 0,012670 + Ant x 0,012269 + Cnt x 0,009911 + Dri x 0,008246 + Cmp x 0,007720 + Cro x 0,006812. And then you can compare this 'sum' between the players. In theory the higher the sum the better the player should be.
Additionally when you take into consideration this test you can include in your filter out players with 6+ or 10+ Work Rate since very low Work Rate results in extreme decrease in players performance.
Orion said: The training data was transformed to show player attributes not as 1-20 but I calculated average value for each attribute for every league. Then for every player I changed his attribute value as (actual_attribute_value - average_attribute_value_in_player_league). This way attributes can also be negative - if player has lower attribute value than average for his league. And that was the data that I used to training the model. So model says 'hey, take this 'new attribute value' for the player and multiply it by his coefficient'.
Because of this in theory we should check in what league you are playing, check average attributes values for the players in that league, calculated this new 'difference' attribute and then use coefficient to calculate player 'rating'.
To make it easier, and because we compare players that will play in the same league, we will just use original attribute values of players. This may introduce some additional error in the rating calculation, but since this makes calculation so much easier we accept this flaw of the method.
Btw how does this work with double attributes? Like Jum^. If the model uses adjusted values then squaring the full number seems odd and potentially overshadowing everything else.
Thanks for all the research!
Looks like jumping reach is most import attr for a striker, but how does this work when this thread (https://fm-arena.com/thread/14009-attribute-testing-football-manager-24/) indicates it's capped out at 17?
Yarema said: Btw how does this work with double attributes? Like Jum^. If the model uses adjusted values then squaring the full number seems odd and potentially overshadowing everything else.
In linear regression you basically check every single feature and select the one that fits to the data the most. Then you exclude that one and check the model with this first feature + the second one.
In polynomial regression you allow model to look for 'single' feature but also combination like feature1 x feature2 or feature1 x feature1. If it fits the data better than 'single' feature it will pick that feature. If 'single' feature fits the data better than 'combined features' it will keep the single one.
In this model I allowed searching for quadratic equations. And for some cases it turned out that at some point, like for the Forward, Jum x Jum fits the data better than any single feature that was left in the pool. But for example Jum x Jum gives worse fit that just Jum.
Pip said: Thanks for all the research!
Looks like jumping reach is most import attr for a striker, but how does this work when this thread (https://fm-arena.com/thread/14009-attribute-testing-football-manager-24/) indicates it's capped out at 17?
They used different methodology. Keep in mind that for a striker coefficient for Jumping Reach is ~0,024 while for Pace it's ~0,021 so high pace is 'almost' the same 'good' as high Jumping Reach but having both high is even better.
They use very different approach to the testing. If they proved that Jumping Reach above 17 doesn't give significant increase in the outcome so it be.
If you want to be precise in model results interpretation it basically says that every single point of Jumping Reach higher than the average for a certain league gives a striker +0,024768 for his average rating. So this +3 point difference (20 compared to 17) gives around +0,074 to average rating. If in reality it doesn't it's not much off. Especially if we consider that it represents features mostly in linear way.
I was really surprised that basically a linear regression model was able to be that accurate in this particular application.
Orion said: In linear regression you basically check every single feature and select the one that fits to the data the most. Then you exclude that one and check the model with this first feature + the second one.
In polynomial regression you allow model to look for 'single' feature but also combination like feature1 x feature2 or feature1 x feature1. If it fits the data better than 'single' feature it will pick that feature. If 'single' feature fits the data better than 'combined features' it will keep the single one.
In this model I allowed searching for quadratic equations. And for some cases it turned out that at some point, like for the Forward, Jum x Jum fits the data better than any single feature that was left in the pool. But for example Jum x Jum gives worse fit that just Jum.
Yes but the model was working with adjusted numbers. Let me give an example. Average JR in the league is 12, a striker has 17. The way you explained stuff in the original post and later on basically says you take 17-12=5 so you do the calculations with 5. And isn't it possible that quadratic fits well for adjusted number (5) but not the actual number 17? Squaring numbers like 5 is a lot different that 17. Above 8 it basically dominates every other attribute basically double dipping (or triple since it's also the best attribute for a striker anyway).
Edit: Or to put it another way. Comparing JR contribution in above scenario if striker has 17 or 18 JR. Taking numbers at face up value you'd get roughly 1,27 vs 1,40. So a difference of 0,13 which would roughly be the effect of +6 pace.
Using adjusted numbers of 5 and 6 makes the scores 0,49 and 0,55 respectively, a difference of 0,06 so half of the previous difference. Still equals +3 pace which feels a lot.
Pip said: Thanks for all the research!
Looks like jumping reach is most import attr for a striker, but how does this work when this thread (https://fm-arena.com/thread/14009-attribute-testing-football-manager-24/) indicates it's capped out at 17?
What made you think that the thread indicates Jumping Reach is capped in the sense that it’s not effective beyond 17? The explanation actually says they deliberately limited it to 17 in the test to avoid an unrealistic scenario, not that its effectiveness is capped in gameplay.
https://fm-arena.com/find-comment/40549/
CBP87 said: @Bafici can you share how you got those GS ratings please? It might just be me but I can't make sense of the Coefficient scoring on the opening thread
I match the highest number to 100 and scale the rest according to that.
Like this for example.
GK
Agi (0,014640/0,014640)*100 = 100
Ref (0,012837/0,014640)*100 = 88
Aer (0,011812/0,014640)*100 = 81
Thr (0,007465/0,014640)*100 = 51
Com (0,007436/0,014640)*100 = 51
Han (0,006255/0,014640)*100 = 43
Dec (0,005089/0,014640)*100 = 35
Numbers are rounded.
Yarema said: Yes but the model was working with adjusted numbers. Let me give an example. Average JR in the league is 12, a striker has 17. The way you explained stuff in the original post and later on basically says you take 17-12=5 so you do the calculations with 5. And isn't it possible that quadratic fits well for adjusted number (5) but not the actual number 17? Squaring numbers like 5 is a lot different that 17. Above 8 it basically dominates every other attribute basically double dipping (or triple since it's also the best attribute for a striker anyway).
Edit: Or to put it another way. Comparing JR contribution in above scenario if striker has 17 or 18 JR. Taking numbers at face up value you'd get roughly 1,27 vs 1,40. So a difference of 0,13 which would roughly be the effect of +6 pace.
Using adjusted numbers of 5 and 6 makes the scores 0,49 and 0,55 respectively, a difference of 0,06 so half of the previous difference. Still equals +3 pace which feels a lot.
Well, I guess this is derived from the idea of this experiment which is determining which attributes have the biggest impact on high match score. Hence, winning headers for ST has the biggest impact probably because striker has many oppurtinities to increase "key headers" stat during the match. Blind guess though.
So what do you guys think is the best genie ratings file to follow for the meta: this one from these findings or the findings from Zippo?
kvasir said: What made you think that the thread indicates Jumping Reach is capped in the sense that it’s not effective beyond 17? The explanation actually says they deliberately limited it to 17 in the test to avoid an unrealistic scenario, not that its effectiveness is capped in gameplay.
https://fm-arena.com/find-comment/40549/
Must have missed this - thanks for pointing out
Glennson said: So what do you guys think is the best genie ratings file to follow for the meta: this one from these findings or the findings from Zippo?
Would like to know this aswell guys, great work btw!
Yarema said: Yes but the model was working with adjusted numbers. Let me give an example. Average JR in the league is 12, a striker has 17. The way you explained stuff in the original post and later on basically says you take 17-12=5 so you do the calculations with 5. And isn't it possible that quadratic fits well for adjusted number (5) but not the actual number 17? Squaring numbers like 5 is a lot different that 17. Above 8 it basically dominates every other attribute basically double dipping (or triple since it's also the best attribute for a striker anyway).
Edit: Or to put it another way. Comparing JR contribution in above scenario if striker has 17 or 18 JR. Taking numbers at face up value you'd get roughly 1,27 vs 1,40. So a difference of 0,13 which would roughly be the effect of +6 pace.
Using adjusted numbers of 5 and 6 makes the scores 0,49 and 0,55 respectively, a difference of 0,06 so half of the previous difference. Still equals +3 pace which feels a lot.
Ok. I get your point. You are right. In such case you can exclude feautres that are quadratic. As said in the topic the results are shown related to the average attributes in a league and in perfect scenario they should be applie that way. My proposal was to simplyfy players comparison.
I'll try to figure something out for this scenario. I might even just run the model for those 2 positions again and allow only 'single' feautres since those are majority of the results anyway.
So I've been testing this on my journeyman save, previously I was a spreadsheet with calculations for attacking and defending based on attributes weighted by the FM Arena new attribute testing.
That got me three promotions and then three years title wins on the top league (Japan)
I'm currently in Korea and with the weighting based on Orions approach, I'm still winning at the same rate but the football seems better to watch and average player ratings are higher. I'm claiming more player & goals of the month.
In general, my players are slower than than what I would sign on my previous set up but using harvestgreen22 training, I could build up pace and acceleration over time if needed.
One thing that does feel off is the jumping reach ratings. I didn't understand the calculations at first so I'm obviously not an expert on any of this but my suspicions are you need one or two players with high ratings there to score from set pieces and it's those goals that boost the average rating rating - potentially any position could have a high attribute.
It might be along the same line of high free kicks/corners/penalty taking having an impact, but that's purely guesswork on my part
This was in my head for several weeks before this post reveal the answer. I sold my Youssef Chermiti who has 140CA with striker rating 79 and upgrading my Striker to Karim Konate 160CA with 84 striker rating. However, Konate never perform better than chermiti and i keep asking... how come? with new rating in this post, its clear that chermiti rating is 77%TS comparing to konate 73%TS.
RFC said: So I've been testing this on my journeyman save, previously I was a spreadsheet with calculations for attacking and defending based on attributes weighted by the FM Arena new attribute testing.
That got me three promotions and then three years title wins on the top league (Japan)
I'm currently in Korea and with the weighting based on Orions approach, I'm still winning at the same rate but the football seems better to watch and average player ratings are higher. I'm claiming more player & goals of the month.
In general, my players are slower than than what I would sign on my previous set up but using harvestgreen22 training, I could build up pace and acceleration over time if needed.
One thing that does feel off is the jumping reach ratings. I didn't understand the calculations at first so I'm obviously not an expert on any of this but my suspicions are you need one or two players with high ratings there to score from set pieces and it's those goals that boost the average rating rating - potentially any position could have a high attribute.
It might be along the same line of high free kicks/corners/penalty taking having an impact, but that's purely guesswork on my part
Good to hear that it works for you. I'm really looking forward to feedback since as said, the model is one thing, we know how it works and in theory I can calculate the average error for the model but knowing that it actually works in 'real life' is something different.
In terms of rating - FM Arena test is based on team Points. harvestgreen22 test is using goal difference. And my model is using players rating as a target variable. Hence my model will look for attributes that increase rating, but not team performance per se. We know that in general average rating will be somehow connected with team results. And I know about a flaw of this model that since it looks for players for high rating it will usually prioritise offensive player but that's due to FM rating system.
pek80 said: This was in my head for several weeks before this post reveal the answer. I sold my Youssef Chermiti who has 140CA with striker rating 79 and upgrading my Striker to Karim Konate 160PA with 84 striker rating. Konate never perform better than chermiti and i keep asking... how come? with new rating in this post, its clear that chermiti rating is 77%TS comparing to konate 73%TS.
Just to clarify. So at first you used GenieScout default 'rating' and then one from my model?
Orion said: Just to clarify. So at first you used GenieScout default 'rating' and then one from my model?
Yes bro, the one in your model is the only "rating model" that rate Chermiti higher than Konate since all other model value pace acceleration, and finishing but not weight much in jumping reach!
pek80 said: This was in my head for several weeks before this post reveal the answer. I sold my Youssef Chermiti who has 140CA with striker rating 79 and upgrading my Striker to Karim Konate 160CA with 84 striker rating. However, Konate never perform better than chermiti and i keep asking... how come? with new rating in this post, its clear that chermiti rating is 77%TS comparing to konate 73%TS.
Could Chermiti’s higher goal tally be because he was your team’s main aerial threat, with extra goals coming from set pieces?
These ratings have worked well for me too, though I find the high Jumping Reach requirement for full-backs a bit off-putting. Just finished my first season with Schalke and comfortably won the 2. Bundesliga—curious to see if I can win the top division as well. In a previous save using FM-Arena ratings, I nearly won La Liga with Alavés in the first season, finishing just one point behind a dominant Real Madrid, who only lost once.
pek80 said: Yes bro, the one in your model is the only "rating model" that rate Chermiti higher than Konate since all other model value pace acceleration, and finishing but not weight much in jumping reach!
Just out of curiosity can you post the screenshot of attributes for both player?
Orion said: Just out of curiosity can you post the screenshot of attributes for both player?
kvasir said: Could Chermiti’s higher goal tally be because he was your team’s main aerial threat, with extra goals coming from set pieces?
These ratings have worked well for me too, though I find the high Jumping Reach requirement for full-backs a bit off-putting. Just finished my first season with Schalke and comfortably won the 2. Bundesliga—curious to see if I can win the top division as well. In a previous save using FM-Arena ratings, I nearly won La Liga with Alavés in the first season, finishing just one point behind a dominant Real Madrid, who only lost once.
One of my wild guesses behind Jumping Reach being high rated attributed for other positions, beside obviously scoring a lot from set pieces, was that it has high correlation with Strength. So player with high jumping reach usually have also high height, high weight and high Strength. So maybe strength also plays part in it because player with high strength will be able to just push opposition player and get the ball back.
pek80 said:
Thanks. I really looks on the first glance that Konate should be far superior to the Chermiti. Like really the only thing that Chermiti has better than Konate is solely jumping reach and nothing else.
Yarema said: Yes but the model was working with adjusted numbers. Let me give an example. Average JR in the league is 12, a striker has 17. The way you explained stuff in the original post and later on basically says you take 17-12=5 so you do the calculations with 5. And isn't it possible that quadratic fits well for adjusted number (5) but not the actual number 17? Squaring numbers like 5 is a lot different that 17. Above 8 it basically dominates every other attribute basically double dipping (or triple since it's also the best attribute for a striker anyway).
Edit: Or to put it another way. Comparing JR contribution in above scenario if striker has 17 or 18 JR. Taking numbers at face up value you'd get roughly 1,27 vs 1,40. So a difference of 0,13 which would roughly be the effect of +6 pace.
Using adjusted numbers of 5 and 6 makes the scores 0,49 and 0,55 respectively, a difference of 0,06 so half of the previous difference. Still equals +3 pace which feels a lot.
For features using multiple attributes my suggestion is to multiply them and then take a square before multiplying by coefficient. So when you have for MLR Acc, Agi and coeff 0,003329 do:
SQR(Acc * Agi) * Coeff
This way you can balance the impact of this double attribute multiplication and you keep the 'partial score' properly scaled so for example when I tried said MLR using the same attribute value for all features, the 'partial score' for the double attribute feature stayed proportional to it's coefficient compared to other features.
FREVKY said: That's what I thought, so I created pretty simple excel spreadsheet that make mass player comparison possible and easy.
Here's how it works:
First, you need to import specified views. I created two sets of these: one for your team (for squad view) and one for scouting tab - both with CA & PA hidden - and another set with CA & PA visible for those who like to spoil the fun a little bit. Download and paste them into "views" folder in your Documents (C:\Users\your_name\Documents\Sports Interactive\Football Manager 2024\views is the default path).
When you load the view, you need to select every player, so click on one player and than ctrl+a to select everyone in the team or in the scouting range. Just bear in mind the more players you select, the more time it takes, so if you're about to select over 1000 players, give it a few seconds to work.
Then, press cltr+p to "print" the selection into HTML file. Save it wherever you want, name it whatever you want.
Then, you need the spreadsheet (MS Excel file). Open it and then in the Excel go to to File->Open and select the html file with your set of players. Copy it's whole contents (ctrl+a, then ctrl+c) and paste them into my spreadsheet in the blank sheet called "IMPORT" then switch the sheet to the one called "CALCULATION" and it should automatically calculate values for every player for each position using coefficients from this thread. Additionalli I added sections with CA, PA and difference between them (it will work only when you used views with PA and CA obviously).
Of course you can use whatever filters you want on the scouting section to narrow down the amount of players to whatever you really need.
At first glance it may sound complicated a bit but it's pretty easy to use. If you find any trouble using it, I'll try to help.
Spreadsheet link: https://www.mediafire.com/file/huj2qrmavoqnd6x/meta.xlsx/file
The spreadsheet is editable so do whatever you want with it, if you find any room for improvements, go for it.
Bafici said: I made a genie scout file. Biggest difference from fmarena ratings is this is bit more realistic. Great players in real life takes higher ratings.
Ratings
Just to keep you guys up to date I've decided to make a small adjustment for the features that use two attributes - so the last one for MLR and the last one for Striker. I've decided to take square of a product from multiplication those features (SQR(Attribute1 * Attribute2) * Coefficient). This way you can keep it proportional to the coefficient.
I don't know how it works with your files. I assume that for spreadsheet it's not an issue.
What about GenieScout? Does it support other mathematical operations? If not I'd suggest just to get rid of those 2 features (since they are last important for given positions) and adjust others to sum up to 100% just like you did for the other positions.
Just want to say how much I appreciate the work you put into testing and letting us see behind the curtain. I like the influx of new research regarding attributes and training and whatnot.
I just find it hard to wrap my mind around the fact that this is currently the best Left back I can have (playing as Malmö 3rd season)