Data Predicting European Top 5 League Soccer Team’s Points
The skills I demoed here can be learned through taking Data Science with Machine Learning bootcamp with NYC Data Science Academy.
Data Science Introduction
In professional sports, we constantly hear the saying that “every game counts” and players are expected to give their all from the beginning to end of a season. Those familiar with professional sports in America know that teams will often tank a season to be compensated with priority draft picks for the next season based on previous data.
However, the top flight leagues in England, Spain, Germany, Italy, and France have a drastically different setups. Teams that finish in the top 3-4 places are rewarded with Champions League qualification for the following season, while those fighting for 4th-7th are competing for other competitions such as the Europa League and the Europa Conference League. On the other side of the league table, teams that finish in the bottom 2-3 places of their respective league face relegation.
Whether a club’s ambition is to challenge for Champions League qualification or avoid relegation, it is important to understand that failure to meet these objectives can destabilize the finances of the organization. For example, Barcelona is viewed as a behemoth of the sport across the globe. In the pursuit of staying competitive domestically and on the European stage, the club started spending above their means. Inflated transfer fees, massive contracts, and a downturn of results on the pitch were only magnified by the crippling effects of COVID. As a result, the club was reported to be $1.5 billion in debt by the summer of 2021.
For this reason, it is essential for a club to set their short/long term goals and structure a payroll that compliments their vision. As seasons are concluded by point totals which are highly influenced by goals, I evaluated 4 models which predict the amount of points a team will garner based on their total shots, shot % on target, and touches in the penalty area.
To build this model, all data was sourced from Wyscout. The data is comprised of metrics from the top 5 European leagues (Premier League, Serie A, La Liga, Ligue 1, & Bundesliga) ranging from the 2015/16 season through the 2020/21 season. Data from Ligue 1 during the 2019/20 season was omitted since the season was halted without a resumption due to COVID.
Overall, the data frame is comprised of the following columns: Team, League, Year, Points, Total Shots, Shot % on Target, and Touches in the Penalty Area. Each row is a team’s data from a respective season. Before building the model, the data was randomly split into sub data set. The train data set was comprised of 75% of the data and was used to build the models, while the test data set was comprised of the remaining 25% and was tested against the model for accuracy.
Multiple Linear Regression - Training Dataset
For this project, 4 models were built to compare which can most accurately predict a club’s point total over the course of a season. The models are listed below:
- mod_1 – Using Total Shots, Shot % on Target, and Touches in Penalty Area to predict Points.
- mod_2 – Using Total Shots and Shot % on Target to predict Points.
- bc_mod_1 – Box Cox with the same x variables as mod_1
- bc_mod_2 – Box Cox with the same x variables as mod_2
While there were clear linear relationships between all three x variables and points, there was a slight skew when interpreting total shots and total touches in the penalty area. This is expected in soccer, because there are typically large gaps between top clubs and mid table clubs, as well as mid table clubs from relegation bound clubs. The Box Cox models were an attempt to normalize this.
Additional Data Exploration - Do All Leagues Share the Same Trend
In order to ensure all leagues share the same linear trend, ggplot scatter graphs were created with each league being separated via facet wrap. Images are pictured below:
Multiple Linear Regression – Test
Once all models were completed and confirmed significant by their p-value, we used the predict function in R to predict the point outcomes from our Test data set against each model. Upon completion, the predicted point totals were merged to 4 versions of the Test data set. Each new data set then had the predicted points generated from each model subtracted from the actual point metric, to generate a new column which displayed the point difference. Finally, we computed the mean point difference across the entire data set to find which was closest to the actual point metric, on average.
Below, we can see a snapshot of a summary table which shows the key takeaways from each model.
From this table, I conclude that mod_1 would be the strongest to utilize over the long term. At first glance, it seems as if the box cox models have stronger metrics. The RSE is drastically lower, both AIC/BIC are smaller than the non box cox models, and the R^2 adj is comparable to the non box cox models. However, when evaluating the average point differential, the standard models are far more accurate.
The average point difference, which is the most important factor when a club would be utilizing a model, is drastically more accurate on the non box cox models. For that reason, we will evaluate model 1 and 2. Model 1 has a lower RSE, higher R^2adj, and a smaller AIC/BIC. In addition to this, when we tested the model the average point difference was -0.2483004. This is closer than the average point difference of mod_2, which came out to -0.4957172.
Below, we can see some key takeaways listed from the summary of mod_1 and some plots:
- All coefficients are significant.
- The overall F-statistic is significant so the overall regression is significant.
- The RSE is 9.894, which is an estimate of the average deviation of the observations around the regression line.
- The R^2 adjusted is 0.6836, meaning 68.36% of the variation in points is accounted for by the variables in our model.
- The VIF for all variables are below 5.0, so we don't have an issue with multi-collinearity.
- Residuals vs Fitted Plot doesn't show a distinctive patter, indicating we don't have non-linear relationships.
- Normal Q-Q Plot shows the residuals follow a straight line, with some slight deviation at the ends.
- Scale-Location Plot shows that the residuals appear to be randomly spread, showing there will not be an issue with variance.
- Residuals vs Leverage Plot shows that we do not have any outliers that are influential against the regression line.
Call: lm(formula = Points ~ Total.Shots + Shot...on.Target + Total.Touches.in.Penalty.Area, data = training_set) Residuals: Min 1Q Median 3Q Max -30.910 -6.528 -0.367 6.348 34.309 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) -64.983879 5.562015 -11.684 < 2e-16 *** Total.Shots 0.091293 0.009888 9.233 < 2e-16 *** Shot...on.Target 1.607951 0.149640 10.745 < 2e-16 *** Total.Touches.in.Penalty.Area 0.030064 0.005400 5.567 4.59e-08 *** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Residual standard error: 9.894 on 427 degrees of freedom Multiple R-squared: 0.6858, Adjusted R-squared: 0.6836 F-statistic: 310.7 on 3 and 427 DF, p-value: < 2.2e-16 Total.Shots Shot...on.Target Total.Touches.in.Penalty.Area 2.895768 1.219490 3.245311
In conclusion, mod_1 would be an effective model to predict points based on total shots, shot % on target, and total touches in the penalty area. A club can estimate these totals based on the data from the players they have on their roster, compute team totals over the course of a 34-38 match season (depending on the league), and plug them into our trained model to find out how many points they expect their team to generate.
Based on the predicted point metric, they can use historical league data to find out what position in the table they will find themself in. This will allow them to structure the player payroll accordingly, or act if any changes are needed.
Bonus - Shiny App
In addition to the model, a Shiny App was created from the complete data set. The app displays a histogram which shows the total shots taken by each team in an individual season. The user can alter the number of bins displayed on the histogram by adjusting the slider on the right. Below the histogram, we find a summary of the data. Lastly, we can see a histogram that displays points on the y axis and total shots on the x axis. The link to the Shiny App is below: