Reading between the lines

Some people say the judging potential (JP) skill is a waste of time. The thinking being that some managers don't need it. They can tell a great player just by looking at their attributes.

The way that works is based on a couple of factors:

1. The older the player the better they should be. If you see a 5 star potential youngster with a bunch of single figure attributes at the age of 19 he is probably not going to become a world beater. If you see a 14 year old player with 14s and 15s he is probably going to become a star.


2. Key attributes in the right positions hint at the players potential - or more accurately hint at the player's current ability - and given the same level of development, they will judge the youngster to have potential.

Obviously any manager with JP has an advantage over anyone who doesn't because they have one more piece of information. If nothing else it's worth buying 5 star potentials because they sell for such high amounts - far more than 4 or 4.5 star youngers. If someone ever tells you one of your youngsters is a 5 star, remember that and adjust your expected price accordingly if you ever sell.

Anyway, using a data set from FM08 using only the initial players, no regens, I have attempted to see how accurate it is to use attribute scores to predict a players current ability (CA) rather than their potential ability. The test I ran was on almost 8,000 players who can only play in the striker role (STs). This prevents contamination of the sample set by hybrid players - who represent two or more positions.

My analysis was to determine how accurately I can predict current ability by reading a striker's attributes. Many of the input variables are highly correlated, so overfit is always going to be a problem. What this means is that generally, the better the player, the higher their attributes are. However, some variables are more important than others.

Firstly I stripped out the attributes that are not correlated with CA for strikers to remove some background noise from the model. These include marking, tackling and positioning for example (I first tested these variables to prove they had no effect on the model - so don't worry that I missed something). I loaded the data into SPSS, a stats package, and ran a linear regression model on the data. After a few iterations I found a model that explains 91% of variance in the data. The predicted value has a 0.951 correlation with current ability. So it's pretty accurate, and I think with some tweaking I could improve the fit much more (run separate models for young and old players; split strikers into target men and fast strikers).

So if you want to know what the approximate current ability of a player is, you have to multiply each attribute coefficient (numbers below) by the value the player currently has and then add them up and then add the constant (which is a negative number so it actually reduces the overall). The variables in desecnding order of importance are:

Pace 2.407
Acceleration 2.248
Strength 1.114
Finishing 1.057
Off The Ball 0.885
Decisions 0.779
Jumping 0.773
Heading 0.772
Balance 0.740
Concentration 0.719
Technique 0.664
Anticipation 0.627
First Touch 0.562
Composure 0.560
Passing 0.505
Dribbling 0.502
Long Shots 0.459
Stamina 0.411
Creativity 0.387
Agility 0.359
Crossing 0.205
Constant -92.09

So basically, Current ability = 2.407xPace + 2.248xAcceleration + 1.114xStrength + 1.057xFinishing... ...+ -92.09.

Below is a random sample of players with their predicted and actual current ability and the difference between them.

Name Pre CA Act CA Diff
Willock, Calum 63 62 1
Fleita, Juan Ramón 121 111 10
Crawford, Brian 36 21 15
Akinfenwa, Adebayo 86 81 5
Kornilenko, Syargey 101 120 -19
Solodukhin, Vladimir 47 41 6
Davydov, Sergey 59 64 -5
Mazilu, Ionut 135 135 0
Cadikovski, Dragan 115 110 5
Sturm, Jani 82 87 -5
Lavric, Klemen 122 124 -2
Andersson, John 40 35 5
Waldh, Daniel 27 30 -3
Arthuro 100 107 -7
Birchall, Adam 55 74 -19
Harris, Neil 82 93 -11
Demouge, Frank 124 110 14
Offiong, Richard 103 105 -2
Skjøth, Peter 73 73 0
Bymar, Jacob 50 54 -4
Viale, Julien 130 117 13
Manchev, Vladimir 118 118 0
Bibishkov, Krum 106 112 -6
Vittek, Robert 124 147 -23 (pictured right)
Nikolaou, Giorgos 102 100 2
Nenadic, Vladimir 70 77 -7
Vagner, Robert 73 89 -16


So not perfect, but about 91% accurate, which is more accurate than the JP lvl4 skill for judging potential! I've also built similar models for MCs and DCs - which obviously have different attributes at the top. If you want to read more about linear regression you can find the basics on Wiki.


I think a purer test would be to conduct the same analysis on regens only. Regens don't possess the biases of the SI researchers, and should be easier to build a model of. This article fits nicely with the guide I wrote on strikers last month - although that analysis included strikers with other positions (FC, AMC/FC etc). It's interesting that pace and acceleration come top of the variables. There is a high correlation between the two - in the normal scheme of thing you would only use one of them in a model because of colinearity. In game everyone talks about the importance of pace in strikers and maybe this model backs that up.

Thanks for the guys on the FM08 forum (tactics) who inspired this piece of analysis with their previous work on free attributes from training.