A Serie A Analysis

I am a football (calcio in Italian) fan and Data Scientist and I have always looked around to find ways data is used in football. Possibly in a more scientific way than it is currently done in most newspapers and TV commentary.

There is a beautiful book from Sally and Anderson you can read. Apart from it I could not find anything else in published results. The most interesting analysis I read where taken from independent analysts, usually blogging or sharing their work on social media.

There is plenty of room for an original contribution to football analysis and I want to show here how a simple analytics framework can be set up given a free dataset and a couple of python libraries.

The Data

The most reliable, complete and above all free source of data I found is here. In this website you can find a number of statistics (in csv format) for many leagues across Europe. Start by downloading the files for the last couple of season.

``import pandas as pd ``
``df_cal = pd.read_csv("serieA_1516.csv")``

Here you justopen the serie A 2015/2016 file and in the dataframe there are all match infos for the season, results, shots on target, corners…
To have a detailed list of the info in the csv files you can look here.

The first step is to extract the name of the teams in the League, but of course one can also hard-code it as a list.

``teams_a = [] ``
``for i,cal in enumerate(df_cal.iterrows()):     teams_a.append(cal[1]["HomeTeam"])     teams_a.append(cal[1]["AwayTeam"])     #read only the first 10 matches     if i == 9:         break ``
``print teams_a``

As the rows in the dataframe represent a match, it’s enough to read the first 10 rows to get the names of all teams in serie A (20 teams). You can then loop on the rows of the dataframe, extract the Home and Away team and append them to our team list.

Analysis of correlations

From this point you can start showing all different types of correlations between, for example, the points scored by a team and the number of shots on target.

To extract the number of shots made one does the following:

``for team in teams:     team_home = df[df['HomeTeam']==team]     team_away = df[df['AwayTeam']==team]     #shots made     team_s = team_away["AS"].sum() + team_home["HS"].sum()``

Here, you read the dataframe `df` and extract the rows where the name of the home or away team is the team one is looking for. You then save these as two separate dataframes (one for the matches where the team played away, the other for the matches where the team played home). One then reads the column “AS” (away shots) in the case of away matches and the column “HS” in the case of home matches. Summing all values and one gets the number of shots made by that particular team in the season. Repeating this step for all teams you can build a simple dataframe that looks like this table.

It is interesting to add to this table a column showing the number of points made by each of these teams in the season and see if there is a correlation between the number of points and the shots made. It is really simple to do:

``for team in teams:     t = df[(df['HomeTeam']==team) | (df['AwayTeam']==team)]    team_home = df[df['HomeTeam']==team]     team_away = df[df['AwayTeam']==team]     team_h_win = len(team_home[team_home['FTHG']>team_home['FTAG']])     team_a_win = len(team_away[team_away['FTAG']>team_away['FTHG']])      team_draw = len(t[t['FTAG']==t['FTHG']])     team_points = 3*team_a_win + 3*team_h_win + team_draw``

The idea here is similar as before. Extract the rows with the team playing away or home, count the number of wins of the team by selecting the rows where the home team scores more goals then the away team and viceversa. Do not forget to account for draws. Finally sum the number of draws with the number of won matches multiplied by three. So we have a table like this now:

It turns out that there is quite a correlation between the number of points and the number of shots, at least for serie A 2015/2016 season.

Juventus was the team with the highest points in the season and the second highest number of shots. The correlation seems not to hold so well in the low end of the point spectrum (the two teams with the lowest points shot more than teams that staid up) hinting at other variables being important in predicting the number of points at the end of the season.

Build a Model to predict points

In order to build a model that can predict the number of points a team will make at the end of the season the first step is to extract as many predictive features as possible. Following the same procedure described above one can extract the followings:

2. Shots on target.
3. Shots conceded.
4. Shots on target conceded.
5. Corners.
6. Corners conceded.

The second step will be to feed these features to some kind of regression model. One can use part of the dataset to train the model and part to test the accuracy of it and then draw some conclusion.

For this exercise we can use a simple linear regression model.

``df = pd.DataFrame.from_csv("../data/serieA.csv") ``
``features=df[["shots","shots_ontarget","shots_conceded","shots_conceded_ontarget","corners","corners_conceded"]] ``
``target = df["points"]``

First step is to read the dataframe where all the features built for the model have been saved. Then one defines the target and feature vectors. The features vector has been discussed above, the target vector is made from the points of each team at the end of the season.

In this case many serie A seasons have been analysed and aggregated together in order to increase the statistics and improve the model precision.

Next step, create a Linear Regression Model and train it.

``from sklearn import datasets, ``
``linear_model regr=linear_model.LinearRegression(fit_intercept=False)``

In order to train the model one can use cross validation, there is a library in scikit-learn for it.

``from sklearn import cross_validation as cv ``
``scores = cv.cross_val_score(regr, features, target,cv=4) ``
``print("Regression scores", scores) print("Regression scores average %.2f" %np.mean(scores)) print("scores variance %.2f" %np.std(scores))``

Here one simply uses a 4-fold cross validation to train the model 4 times on 4 different subsets of the original data sample. At the end of each training one evaluates the score of the model and saves it in a vector called score. In the case of Linear regression the score is a simple R². All of it is done in just one line using the cross validation library provided by scikit-learn. In this case R²=0.70±0.07. Pretty good agreement.

Further developments

The analysis above can be improved in a number of ways. From analysing a bigger dataset to adding more features to applying a different model, like a regression tree for example. The general approach would be the same. Split the data sample in N-folds, use one fold to test the model and train the model on the remaining (N-1) folds. Repeat for each fold and average the result to get an idea of how good is your model. For an excellent introduction to the technique I recommend this video

Originally published at tropiano.github.io on September 21, 2016.

One clap, two clap, three clap, forty?

By clapping more or less, you can signal to us which stories really stand out.