The NFL Combine and Pro Bowl Cornerbacks: A Data Exploration
github repository for this project
Introduction
Every year the NFL invites the top college football players to participate in a feats of strengths challenge, the NFL Combine. The Combine helps NFL teams evaluate perspective talent and informs their decisions on which players to draft.
I’ve been interested for a while in exploring the NFL combine dataset and understanding how useful it is in predicting success in the NFL. What follows is a brief exploration of the data, specifically for prospective NFL cornerbacks.
Combine and Pro Bowl Basics
The NFL Combine is a week-long event where players hoping to enter the NFL are put through a battery of physical tests in order to evaluate their NFL potential. These tests include, among others, the 40-yard dash, the broad jump, and the bench press, as well as measures of physical attributes like height, weight, arm length, and hand size.
The Pro Bowl is an all-star game between the NFL’s top players. The 88 players who make the Pro Bowl are voted on by coaches, players, and fans and chosen from the pool of ~1700 NFL players. We assume here that a player making the Pro Bowl equates to individual success in the NFL.
Why Cornerbacks?
I narrowed this exploration to NFL cornerbacks. From a qualitative perspective, cornerbacks are often touted as the most athletic players and the position where physical athleticism is most important. Thus, if data from the Combine are useful in projecting performance, then the cornerbacks position will be the most promising candidate for modeling.
Core Question
Based on a cornerback’s performance in the NFL Combine, can we predict the likelihood of that cornerback making the NFL Pro Bowl in the future?
Data
Combine Data: The NFL combine results for each individual player can be found at nflcombineresults.com. I used the years 1987–2020 for my analysis. You can find also find the data in .csv format in the github repository for this project.
Pro Bowl Data: The list of NFL Pro Bowl players was pulled from Pro Football Reference. The data is also in .csv format in the github repository for this project.
Data Exploration and Visualization
First let’s look at the split of Pro Bowl versus non-Pro Bowl cornerbacks. As you can see, there is a big imbalance between non-Pro Bowl (94.6%, blue) and Pro Bowl cornerbacks (5.4%, orange).
~5% of cornerbacks in NFL Combine history make the Pro Bowl
Count of non-Pro Bowl cornerbacks v. Pro Bowl cornerbacks
non-Pro Bowler v. Pro Bowler visualizations
The following violin plots represent a selection of the Combine data broken out between distributions of non-Pro Bowlers (blue) versus Pro Bowlers (orange). Python code snippet follows.
import seaborn as sns
import matplotlib.pyplot as plt#data to plot
plots = ['college', 'height in', 'weight lbs', 'hand_size in',
'arm_length in', '40 yard', 'bench_press','vert_leap in',
'broad_jump in', 'shuttle', '3cone', '60yd_shuttle']#plot all the columns
for plot in plots:
sns.catplot(x="is_pro_bowl", y = plot, kind="violin", data=cornerbacks, height=8.5, aspect=.9)
plt.savefig(plot + '.png')
Pro Bowl Cornerbacks seem to be taller and heavier
Height (in)
Weight (lbs)
Pro Bowl Cornerbacks seems to be faster (lower 40yd dash time), but not necessarily stronger (similar bench press reps)
40 Yard Dash (s)
Bench Press (reps)
Pro Bowl Cornerbacks seem to be able to jump higher (vertical) and farther (broad jump)
Vertical Leap (in)
Broad Jump (in)
Pro Bowl Cornerbacks seem to be quicker with better acceleration when changing direction.*
*Shuttle run tests require an athlete to run while changing directions
Shuttle (s)
60yd Shuttle
Summary of Visual Observations
Based on these plots one gets the sense that Pro Bowl cornerbacks are bigger, faster, but not necessarily stronger (see bench press).
#lets see the means across columns for the two groups
cornerbacks.groupby('is_pro_bowl').mean()
Hypothesis Testing
In order to more rigorously evaluate whether there are significant differences between the two groups of cornerbacks I looked at each NFL Combine test and performed hypothesis tests between the groups of cornerbacks.
#============Hypothesis Testing======================
from scipy.stats import ttest_ind
for plot in plots:
#sns.catplot(x="is_pro_bowl", y = plot, kind="violin", data=cornerbacks, height=8.5, aspect=.9)
t1 = cornerbacks[cornerbacks['is_pro_bowl'] == 1][plot]
t2 = cornerbacks[cornerbacks['is_pro_bowl'] == 0][plot]
t1_mean = np.mean(t1)
t2_mean = np.mean(t2)
t1_std = np.std(t1)
t2_std = np.std(t2)
ttest,pval = ttest_ind(t1, t2)
if pval <0.05:
print('====== ', plot, " ======")
print("p-value", pval)
print("we reject null hypothesis [" + plot + "] IS sig different!")
print("Pro Bowl mean value:",t1_mean," stdev: ",t1_std)
print("Non- PB mean value:", t2_mean," stdev: ",t2_std)
print("==== END ", plot, " ====")
print(" ")The results show statistically significant differences between non-Pro Bowl and Pro Bowl cornerbacks for the NFL Combine tests which include: 40 yard dash, broad jump, shuttle, and 60 yd shuttle.
====== 40 yard ======
p-value 1.3708582248077843e-05
we reject null hypothesis [40 yard] IS sig different!
Pro Bowl mean value: 4.490860324171952 stdev: 0.09324748995664504
Non- PB mean value: 4.546474208961045 stdev: 0.10696579488451659
==== END 40 yard ====
====== broad_jump in ======
p-value 0.008383869956918569
we reject null hypothesis [broad_jump in] IS sig different!
Pro Bowl mean value: 121.37864864864866 stdev: 6.273889060387517
Non- PB mean value: 119.6110296617372 stdev: 5.516273484274304
==== END broad_jump in ====
====== shuttle ======
p-value 0.018817831255597058
we reject null hypothesis [shuttle] IS sig different!
Pro Bowl mean value: 4.148978537012703 stdev: 0.16950481445681398
Non- PB mean value: 4.187172800326174 stdev: 0.1321607800289746
==== END shuttle ====
====== 60yd_shuttle ======
p-value 0.005133910470464446
we reject null hypothesis [60yd_shuttle] IS sig different!
Pro Bowl mean value: 11.382047488584476 stdev: 0.23683360234089437
Non- PB mean value: 11.46302547313403 stdev: 0.24082481449027213
==== END 60yd_shuttle ====Machine Learning Model / Logistic Regression
There are significant differences between non-Pro Bowl and Pro Bowl cornerbacks across four of the NFL Combine tests. If one could use the Combine performance to predict whether a prospective cornerback is likely to make the Pro Bowl in the future, that would be interesting and potentially useful to NFL teams.
I’m not a machine learning expert (yet), but I know enough to hack a model together for my own entertainment. My process to make a model to predict NFL Pro Bowl cornerbacks follows.
Training/Test Split and Oversampling
First I split the dataset into two groups, a set of training data to train the model on, and a set of test data to test the model. I chose a 70/30% split given the small dataset.
Our data are also highly imbalanced with only around 5% of the data representing Pro Bowl cornerbacks and 95% representing non-Pro Bowlers. There may be too few instances of Pro Bowlers for our model to train on. Therefore, we use the synthetic minority oversampling technique (SMOTE) to create synthetic data points for the model training process.
#==========OVER SAMPLING USING SMOTE=================
X = cornerbacks.loc[:, cornerbacks.columns != 'is_pro_bowl']
y = cornerbacks.loc[:, cornerbacks.columns == 'is_pro_bowl']
from imblearn.over_sampling import SMOTE
os = SMOTE(random_state=0)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=0)
columns = X_train.columns
os_data_X, os_data_y = os.fit_sample(X_train, y_train) #oversample training data
os_data_X = pd.DataFrame(data=os_data_X,columns=columns)
os_data_y= pd.DataFrame(data=os_data_y,columns=['is_pro_bowl'])
# we can Check the numbers of our data
print("length of oversampled data is ",len(os_data_X))
print("Number of non-Pro Bowlers in oversampled data",len(os_data_y[os_data_y['is_pro_bowl']==0]))
print("Number of Pro Bowlers",len(os_data_y[os_data_y['is_pro_bowl']==1]))
print("Proportion of non-Pro Bowler data in oversampled data is ",len(os_data_y[os_data_y['is_pro_bowl']==0])/len(os_data_X))
print("Proportion of Pro Bowler data in oversampled data is ",len(os_data_y[os_data_y['is_pro_bowl']==1])/len(os_data_X))Modeling
With the training and test data sets in place, I use a logistic regression classifier to create the model in Python.
#==========LOGISTIC REGRESSION MODEL FIT=============
from sklearn.linear_model import LogisticRegressionX=os_data_X[cols]
y=os_data_y['is_pro_bowl']logreg = LogisticRegression(C=200, class_weight=None, dual=False,
fit_intercept=True, intercept_scaling=1,
max_iter=1000, multi_class='ovr', n_jobs=1, penalty='l2', random_state=None, solver='liblinear', tol=0.0001, verbose=0, warm_start=False)
logreg.fit(X, y)#predicting test set results and calculating accuracy
y_pred = logreg.predict(X_test)
print('Accuracy of logistic regression classifier on test set: {:.2f}'
.format(logreg.score(X_test, y_test)))
Is the Model Any Good?
The model has decent recall for Pro Bowlers (.71), however, the precision is low (.12), which leads to a low overall f-measure (.21). This tells us the model is not very useful for predicting Pro Bowl cornerbacks.
(0: non-Pro Bowl, 1: Pro Bowl)
Precision, Recall, and F-Measure
precision recall f1-score support
0 0.96 0.58 0.72 285
1 0.12 0.71 0.21 24
accuracy 0.59 309
macro avg 0.54 0.64 0.47 309Confusion Matrix
While we correctly identified 71% (17 out of 24) of the Pro Bowl cornerbacks in the test set, we ended up with a high false positive rate (120 false positives out of 137 predicted Pro Bowlers).
Conclusions
There are statistically significant differences between future Pro Bowl cornerbacks and their counterparts in the NFL Combine. Specifically, these differences show up in the 40 yard dash, broad jump, shuttle, and 60 yd shuttle.
However, the model to predict which Combine cornerbacks will actually make the Pro Bowl in the future produces too many false positives to be useful in a real world scenario. Future prediction models might benefit from exploring other non-binary success metrics such as career pass breakups, interceptions, or salary.
References: github repository for this project, NFL combine data, Pro Bowl Data, logistic regression process
