How are the card ratings calculated in FIFA Ultimate Team?
How EA determines FIFA player ratings has been a source of mystery to me since I started playing FIFA around three years ago. I’m not referring to how EA collects data from the real world and assigns attribute values to each player. I’m specifically referring to how an overall rating is determined for each player, based on each player’s individual sub-stats.
To clarify, each player is given an overall rating, shown on the top-left of each card. This rating must somehow be based on an aggregate of a player’s rating in each of 29 different categories (shown below for Marcos Acuna).
At first it may be tempting to assume that a player’s overall rating is based on a simple average of the 6 main category values (PAC, SHO, PAS, DRI, DEF, PHY). However if you were to take the average of these values for Acuna, and compare them to Pogba, you will find that Acuna has a higher average, even though Pogba has an overall rating 3-points higher:
Acuna: (76 + 74 + 82 + 86 + 78 + 82)/6 = 79.6 Pogba: (73 + 81 + 86 + 85 + 66 + 85)/6 = 79.3
The same is true if you took the average of all 29 substats. What gives? According to this article, EA uses positional coefficients to determine a player’s overall rating. I have scoured the internet for these positional coefficients, and have not found an updated source for all positions, for FIFA 21. So, naturally, I’ve decided to determine these coefficients myself. I web-scraped all the player data from futbin.com for all 20,760 players, and used a general linear model to determine the coefficients for each position. For each position, I performed a GLM with no intercept term, then I removed any variable with an insignificant p-value (PVAL>.01) or a negligible coefficient (COEF < .001). Finally I reran the GLM to get the coefficients shown in each table below. At the bottom of the table I provide an Average Error row, which represents the absolute average error between the model’s prediction for a given player rating and the actual EA player rating (all error values are well below 1.0, meaning that predictions using these coefficients are very close to the actual rating on each card).
As a fun follow-up analysis, I ran the data through a t-SNE dimensionality reduction method, using the 29 performance categories as features. For this analysis I combined the outside wing positions (LW + RW = OW), the outside fullback positions (LWB + RWB + LB + RB = OB), outside midfielders (LM + RM = OM), and strikers with center forwards (CF + ST = ST).
Another interesting thing to look at is how correlated each individual player stat is with the player’s cost on the FUT transfer market. I scraped the current market cost for each player, and determined the correlation between each player stat and log(Cost). This information is given in the table below. Interestingly, reactions and composure were the stats that were most correlated with player cost (I would have thought it’d be sprint speed or acceleration, since those are stats that benefit players at any position, and because FIFA gamers tend to fixate on pace).