Exploring Performance Metrics: NFL Combine Analysis
Background
In the National Football League (NFL), athletes exhibit a remarkable range of physical attributes tailored to the demands of their positions. Every year, the NFL Combine collects physical and performance data from new athletes declaring for the NFL draft. Scouts use this data to further evaluate player prospects beyond their on-field performances.
In this post, I will investigate the effectiveness of K-means clustering in identifying player positions using NFL Combine data from 2000 to 2018. Additionally, I will conduct several tests to investigate annual trends in the Cone Drill.
Exploratory Analysis
The following code will be run using a Python 3 kernel in Jupyter Notebook. We’ll start by loading in the dataset and libraries below!
First, let’s take a look at the raw data.
We have 15 fields for 6218 players in our dataset, including information on their draft outcome. Some players have missing information, and several fields that are not of interest to us…
Let’s group similar player positons together (i.e., a linebacker LB and an inside linebacker ILB). This will be useful later in our cluster analysis.
Now let’s drop those pesky fields.
Upon investigating the data source, it is safe for us to assume players with NaN values for Team, Round, and Pick went undrafted. We will alias these players easier to make them easier to identify. As for the players with missing combine metrics, we will remove them altogether.
Let’s take one last look at a summary of the new dataframe.
We are left with 2885 player records to evaluate.
Variable Correlation
Next, we’ll investigate relationships between numeric variables in the dataset by creating a correlation heatmap.
Many of these correlations seem intuitive. We might expect a player’s weight to be positively correlated with their 40-Yard Dash time, which we see with a correlation coefficient r of 0.89 (i.e., heavier players often take longer to complete this drill). It’s also reasonable that a player’s standing jump would be negatively correlated with their Forty-Yard Dash time. Players who complete the dash in less time (faster) are more likely to jump higher than heavier, slower players. We will circle back to these correlations in a bit for further analysis.
Cluster Analysis
Let’s explore different player types within the dataset using K-means clustering. To decide on an optimal number of clusters, we’ll develop a function that visualizes inertia across different values of k. An elbow point on this plot will represent a significant decrease in inertia relative to increasing k-values, indicating the optimal number of clusters where adding more clusters does not significantly reduce inertia.
Now that we’ve created an inertia function, we can test it by first clustering on all numeric fields in the dataset. This process will effectively create k player clusters characterized by similar performance metrics among their members.
Based on the elbow plot above, we will select k=3 for our first K-means clustering on all numeric fields.
As specified, the 2885 player-records have been clustered on all numeric fields (Height, Weight, Forty, Bench…) and assigned to one of three groups. The plots above show the distribution of player position counts within each cluster.
Cluster 1 contains mostly offensive and defensive linemen, characterized by their substantial stature and weight. These players exhibit impressive power and explosiveness, making them formidable forces in the trenches.
Cluster 2 contains a majority of skilled players like recievers, defensive backs, and running backs. Known for their agility, speed, and explosive bursts, these athletes excel in making quick, decisive movements on the field.
Cluster 3 presents a diverse mix of player types, including linebackers, defensive ends, and tight ends. Combining elements from both Cluster 1 and Cluster 2, these players showcase a blend of strength, agility, and versatility, making them crucial assets in various facets of the game.
Rather than looking at all numeric variables, let’s narrow down our selection to two highly correlated variables, like Forty-Yard Dash and Weight (r=0.89). Similar to our last procedure, we’ll have to determine a new value of k to cluster on a new variable set.
Based on the elbow plot above, we will again select k=3 for clustering on Forty-Yard Dash (Forty) vs Weight.
The distribution of player positions within these new Forty-vs-Weight clusters is nearly identical to that of our clusters on all numeric fields. Now that we’ve narrowed our focus to two dimensions, we can generate a scatter plot comparing Forty-Yard Dash and Weight. We’ll color-code these points based on their respective clusters.
This plot provides a visual representation of our clusters in the context of Forty-Yard Dash times and Weight. As expected, the clusters are clearly delineated: Cluster 1 points (Power Positions) are predominantly positioned towards the top right on the plot, Cluster 2 points (Agile Players) occupy the bottom left, and Cluster 3 points (Hybrid Builds) are situated in the middle.
Cone Drill Regression Analysis
As promised, it’s time to shift our focus back to the correlation heatmap above. Considering annual trends, we observed that Year shared its strongest correlation with Cone Drill (r = -0.16). While this correlation is relatively modest, it does raise a question: Has the average time to complete the Cone Drill seen a significant decrease since 2000?
Before answering this question, we can investigate the average Cone Drill speed year-over-year with a line chart.
While there does appear to be a dip in the average cone drill speed over time, further analysis is necessary to confirm this trend. We’ll begin by fitting a linear regression model to the data. This model will test the following hypotheses:
𝐻0: There is no significant linear relationship between the year of the NFL Combine and the average speed at which players have completed the cone drill from 2000 to 2018.
𝐻𝑎: There is a significant negative linear relationship between the year of the NFL Combine and the average speed at which players have completed the cone drill from 2000 to 2018, indicating that the average speed has decreased during this period.
Across all players, the average time to complete the cone drill has been decreasing by about 0.0135 seconds per year. Further, approximately 51.95% of the variance in average cone drill speed can be explained by the year. Let’s examine the results of the hypothesis test based on this model.
The statistically significant t-statistic and small p-value suggest that the relationship between the year of the NFL Combine and the average cone drill speed is significant. Therefore, we reject the null hypothesis and conclude that there is a significant negative linear relationship between Year and average Cone drill speed in our data.
It’s important to consider the diversity of player positions within each year of the NFL Combine. A particular draft class might predominantly consist of fast players, resulting in a lower average cone drill time for that year. Conversely, the following draft year might be dominated by slower positions. This positional bias could give the impression that overall player speed changed significantly from year to year, when in reality, it may just reflect the composition of the draft class.
We can run a similar analysis to the one above for each cluster group from 2000 to 2018. In doing so, we mitigate some of the bias introduced by variations in positional subgroups. Let’s aggregate the data for each cluster by taking the average cone drill speed each year.
Cluster 1 Test
Cluster 1 Results:
LR Slope: -0.007284025968538165
R-squared: 0.27525153202504027
t-statistic: -2.5409474310335876
p-value: 0.021099052456581946
Similarly for clusters 2 and 3…
Cluster 2 Results:
LR Slope: -0.0077117141767901794
R-squared: 0.36953339540415586
t-statistic: -3.156605168807558
p-value: 0.00576068502698579
Cluster 3 Results:
LR Slope: -0.01193051964700224
R-squared: 0.48944498115367974
t-statistic: -4.036966314785228
p-value: 0.0008558581836797252
Cluster 1: The average cone drill time for Power Players has been decreasing by approximately 0.0073 seconds each year. The R-squared value of 0.2752 suggests that about 28% of the variance in the average cone drill speed for this cluster can be explained by the year. The negative t-statistic and p-value suggest that there is a significant negative linear relationship between the year of the NFL Combine and the average cone drill speed. Therefore, we reject the null hypothesis and conclude that the average cone drill speed for Power Players has been decreasing over the years.
Cluster 2: Agile Players have also seen cone drill performance improve with their average completion time decreasing by approximately 0.0077 seconds each year. An R-squared value of 0.3695 suggests that about 37% of the variance in the average cone drill speed for this cluster can be explained by the year. The negative t-statistic and p-value suggest that there is a significant negative linear relationship between the year of the NFL Combine and the average cone drill speed. Therefore, we reject the null hypothesis and conclude that the average cone drill speed for Agile Players has been decreasing over the years.
Cluster 3: Hybrid Players have seen the greatest improvement of all with their average cone drill time decreasing by 0.0119 seconds each year. The R-squared value of 0.4894 suggests that about 49% of the variance in the average cone drill speed for this cluster can be explained by the year. The negative t-statistic and p-value suggest that there is a significant negative linear relationship between the year of the NFL Combine and the average cone drill speed. We again reject the null hypothesis and conclude that the average cone drill speed for Hybrid Players has been decreasing over the years.
To investigate these results further, we will plot all three clusters alongside their regression models for comparison.
