-
Our hypothesis is that picking a player, using the easy "Overall" stats, is not the best investment for your game coins. In fact, the majority of people who play FIFA 19 tend to make that mistake, since the "Overall" Rating is the main stat that appears on the front of player cards which makes it an easy choice for most recreational players.
-
Our hypothesis is that if you dig deeper into the individual player skills and carefully curate the optimal skill combo for each position on the field, you will end up with a much stronger combination of players and better team chemistry.
-
Once we prove our hypothesis, we want to build a player recommendation system where you can input your budget and the position which you want to buy a player for and our selector will return a list of the top recommended players based on individual skills and not Overall rating.
-
Lastly, we would like to automate the process via an algorithm which can automatically build a team based on maximizing the specific FIFA stats that are highly important for that position and based on the team formation that you want to play with and whether you want to alocate more budget towards your defensive or offensive players.
Obtaining Data:
We found our data from a Kaggle dataset with detailed players stats for over 18,000 players in FIFA 19. After cleaning up the data, standardazing our features and creating our categorical variables, we ended up with 17907 players with 53 features
Outline of our process:
- Break down the DataFrame into data for the four main positions - GK, DF, MD, ST
- Run K-Means clustering to see what features the model will group players by
- Run four different regressions models which aim to predict a player's value based on 53 player attributes
- Isolate the top 15 features for each position by the weights that each model assigned
- Combined all the features and curate a list of the unique features
- Run our best model with those curated features on the four dataframes
- Apply the weights of our final model (XGBoost with 95% R2_score) to the important features for each position and use the resulting value as a “Selector” score, which we believe would be a "smarter" way of finding optimal players.
- Show Graphs that compare teams built using our Algorithm vs. teams built using Overall rating
The first step was to run a K-Means Unsupervised model and four regression models, which can help us isolate the top 15 player attributes for each position Ridge Regression, Random Forest, SVM, and XGBoost*. Below you can see a few samples of those features in barcharts:
As you can see from the graphs, Overall Rating is an important factor in some models, but not in all and there are a lot of other individual features taht were assigned higher weights than the Overall rating.
Next things we wanted to do was isolate all these important skills and run our top model - XGBoost to get weighted coffeicients for each attribute and multiply those together to create a final "Selector" score, which is a weighted combination of the most important attributes for each position - GK, DF, MD or ST.
Comapre Overall vs Selector Scores: Plot players' overall scores and price versus their combined slector score
The Graphs below demonstrate that for a good number of the players, there is a somewaht linear correlation between Overall Rating and our Selector score. However, we can also clearly isolate many instances where players with a high Overall Rating, which cost a lot of money, do not actually posses high ratings for the individual skills that are important.
In the example above, we can see that there is a "value separation area" right around 1.4-1.6 selector score, where you have a decent number of expensive players wiht high overall rating, that have a lower selector score, than some cheaper players towards the left corner of the graph, where you could be getting players with a much better all-around skillset for cheaper, just because their overall rating is not that high, so most recreational players are not drawn to bid up their price.
Writing Our Player Selector Algorithm:
To continue our investigation further, we decided to write an algorithm that takes in a budget and a player position that you want to spend that budget on and gives you a recommended list of players, optimizing for selector score. Below you can see the code for this simple algo:
def pick_player_by_selector(budget, position, count = 3):
count = count
empty = []
names = [item[0] for item in name_pos if item[1] == position]
STs = ["ST", "LF", "RF", "CF", "RS", "LS", "LW", "RW"]
MDs = ["RM", "LM", "CAM", "CDM", "LCM", "RCM", "RDM", "LDM", "RAM", "LAM"]
DFs = ["CB", "LB", "RB", "RCB", "LCB", "RWB", "LWB"]
if position == "GK":
keepers = sorted_players_selector[0]
for item in keepers:
if item['name'] in names and item['price'] <= budget and count > 0:
empty.append(item)
count -= 1
elif position in DFs:
defenders = sorted_players_selector[1]
for item in defenders:
if item['name'] in names and item['price'] <= budget and count > 0:
empty.append(item)
count -= 1
elif position in MDs:
mids = sorted_players_selector[2]
for item in mids:
if item['name'] in names and item['price'] <= budget and count > 0:
empty.append(item)
count -= 1
elif position in STs:
strikers = sorted_players_selector[3]
for item in strikers:
if item['name'] in names and item['price'] <= budget and count > 0:
empty.append(item)
count -= 1
return empty The Graph below was generated after running our player recommender for the Central Midfielder position - CM. This helps us identify some of those players with high Overall rating, but low Selector score and vice versa:
From the graph above you can see that there are a few players who cost over 20M, who have a low selector score. Here is the example in reverse below, where we're plotting the players that our algorithm recommends, rating them by Selector Score. You can see that not all of them have a high Overall Rating, but we know from our prior invesstigation on individual Model weights and insights that the players with the higher selector scores posses all the strong individual qualities to make a good defender.
Below you can see the code for the final and most exciting part of our project - using the insights from our investigation and our selector score ratings to program an automatic team builder, which takes in the following parameters:
- Total Budget you want to spent on your team
- What team formation you want to use (ex. 4-4-2 vs 3-5-2)
- What portions of your bugdet do you want to allocate for each position.
The default settings below, use the most popular soccer formation 4-4-2 and allocate a bigger portion of the budget on offensive player, which you'd do if you play a more offensive style.
def pick_team_by_selector(budget, DF = 4, MD = 4, ST = 2, GK_coef = 0.05, DF_coef = 0.2, MD_coef = 0.40, ST_coef = 0.35):
#split budget accordingly per position
GK_budget = budget*GK_coef
DF_budget = budget*DF_coef
MD_budget = budget*MD_coef
ST_budget = budget*ST_coef
#Create proper count for each position
GK_count = 1
DF_count = DF
ST_count = ST+1
MD_count = MD+1
#Kreate our list of players per position
keepers = sorted_players_selector[0]
defenders = sorted_players_selector[1]
mids = sorted_players_selector[2]
strikers = sorted_players_selector[3]
#Begin building final team
list_of_names = []
final_team = []
for item in keepers:
if GK_count > 0 and item['price'] <= GK_budget and item['name'] not in list_of_names:
item['position'] = "GK"
final_team.append(item)
list_of_names.append(item['name'])
GK_budget -= item['price']
GK_count -= 1
DF_budget += GK_budget
for item in defenders:
if DF_count > 0 and item['price'] <= DF_budget/DF_count and item['name'] not in list_of_names:
item['position'] = "DF"
final_team.append(item)
list_of_names.append(item['name'])
DF_budget -= item['price']
DF_count -= 1
for item in mids:
while MD_count == MD+1:
MD_budget = MD_budget + DF_budget
MD_count -= 1
if MD_count > 0 and item['price'] <= MD_budget/MD_count and item['name'] not in list_of_names:
item['position'] = "MD"
final_team.append(item)
list_of_names.append(item['name'])
MD_budget -= item['price']
MD_count -= 1
for item in strikers:
while ST_count == ST+1:
ST_budget = ST_budget + MD_budget
ST_count -= 1
if ST_count > 0 and item['price'] <= ST_budget/ST_count and item['name'] not in list_of_names:
item['position'] = "ST"
final_team.append(item)
list_of_names.append(item['name'])
ST_budget -= item['price']
ST_count -= 1
leftover_budget = GK_budget + DF_budget + MD_budget + ST_budget
rating_avg = round(sum([item['selector'] for item in final_team])/len(final_team), 3)
overall_avg = round(sum([item['overall'] for item in final_team])/len(final_team), 3)
return dict(team=final_team, budget_left = leftover_budget, AVG_selector=rating_avg, AVG_overall = overall_avg)Below you can see the result of setting a max budget and seeing what the most optimal team looks like:
As you can see all of our players above, have a very high Selector score, which is an indication that their individual stats are optimized for their specific position.
Now let's see what the team would look like if we spend our max budget onplayers, which have higher FIFA Overall Rating.
As you can see from the two tables above, almost every player in the first chart has a higher Selector Score than their counterpart in the second chart, which means that our Algorithm picked a better curated, more balanced team.
To test the validity of our findings further, we iterated our algorithm through 22 different budgets ranging from 15M to 125M and plotted the avg Selector score per team and budget leftover. The Blue Dots below represent our Algorithms picks, whi are consisntently appearing in that high-value area, meaning that for every given budget our algorithm gets you the combination of players that maximize the most important soccer skills for each position.
The Red Dots represent the average selector score for the algorithm that picks players solely on their Overall rating. As you can see, this data is a lot more incosistent and noisy and we get a few instances of teams with very low average Selector scores, which further supports our theory that a deeper dive into individual player stats and a more strategic allocation of your budget can result in a much stronger all-around team, which will crush your competition!
This was a very interestign deep dive, which gave us sufficient evidence that by quickly glancing at the Overall Rating score on a FIFA player's card and blindly spending your coins to get that player on your team, you're often not getting the best "bang for your buck".
We found that it is worth spending some extra time to click on that player and scroll through their individual stats, to make sure that they have higher values on all the important attributes for the respective position that they play.
Thank you.