Like yesterday’s post, this study examines the predictive value of some of the measurements taken at the NFL combine: height, weight, 40-yard dash, 20-yard shuttle, cone drill, bench press repetitions, vertical leap, and broad jump. I used these metrics to explain the variance in Career Approximate Value (CAV) per year of players drafted since 1999. As a whole, the combine metrics do a better job of predicting defensive performance than offensive production. However, these models are still very weak. They have highly heteroskedastic residuals, and at best explain 21% of the variance in production at a given position. Again, players are divided by the position they played in college, not in the NFL. I remain surprised that the combine statistics explain even that much. Take a look:
Defensive end: This model is one of the best that the combine data produced. It has a standard error of 1.59 and adjusted R2 of 0.19. The p values for each variable, in order, are 0.01, 0.01, 0.01, 0.00, 0.00, 0.01, which indicate that these terms are all highly significant.
CAV = 14.78*weight (pounds) – 0.06*weight2 + 0.00007*weight3 – 0.34*40-yard dash (seconds)2 – 0.09*cone drill (seconds)2 – 1290
There’s a lot going on in this model, but the basic interpretation is that heavier, faster, and more agile defensive ends are better than lighter, slower, and less agile players. This model is the first to include the cone drill, which is a much better predictor of defensive performance than offensive production.
Defensive tackle: Sticking with the defensive line, performance of defensive tackles is harder to project than for defensive ends. The regression output here has a MSE of 1.69 and an adjusted R2 of 0.11.
CAV = 0.03*weight – 2.45*shuttle (seconds) +0.05*bench press repetitions
The p values here are 0.01, 0.00, and 0.05, respectively. This model suggests that heavier, stronger, and more agile defensive tackles perform better than their lighter, weaker, less agile counterparts. These results make intuitive sense, but the variables not included in this model also stand out: height is not a factor, neither is “explosiveness” as measured by either of the jumping statistics.
Outside linebackers: Moving to linebackers, the outside linebacker model is one of the worst ones created from this combine data. It has the largest MSE (1.94) and one of the lowest adjusted R2 values (0.13) of any regression here. The p values for each variable are 0.00, 0.01, and 0.00 respectively.
CAV = -2.14*cone drill -3.48*40-yard dash +33
An example best demonstrates the limits of this of this model. DeMarcus Ware has a CAV/Year of just about 9; he is consistently rated as one of the most dominant defensive players in the game. Based on his combine numbers, he would be projected to have a CAV/Year of 2.5 – a marginal player at best. It should not be too much to ask of a decent model to separate an All-Pro player from a marginal one.
Inside linebackers: The model for inside linebackers is not much better than for outside linebackers, posting an MSE of 1.87 and an R2 of 0.15. These combine measurements simply do not do a good job of predicting performance for linebackers.
CAV = -6.12*40-yard dash +30
It is odd that 40-yard dash is the only significant variable here, but so it is with a p value of 0.00. Similar to the Ware example above, Patrick Willis would be projected as a marginal player based on this model.
Cornerbacks: Moving to the secondary, cornerbacks benefit from being faster, more agile, and heavy. This model has an MSE of 1.62 and an adjusted R2 of 0.13.
CAV = -1.18*cone drill +0.03*weight -5.15*40-yard dash + 27
This model may actually make intuitive sense. Cornerbacks generally need to be fast and agile enough to keep up with wide receivers, but also heavy enough to take them down. It is significant that neither of the leaping variables made it into this model, as it failed to do with wide receivers as well. Thus the notion of a receiver or corner who can “really use his vertical leap to go get after the ball” seems like a lot of empty breath.
Free Safety: The vertical leap does, however, come into play at the free safety position. In perhaps the worst model presented here, (MSE: 1.55, R2: 0.04), only the vertical leap is significant (p = 0.01).
CAV = 0.000027*vertical leap (inches)3
This model is horrendously bad. Really, all it demonstrates is the failure of the combine to predict the play of free safeties, and players as a whole.
Strong Safety: On the other hand, the strong safety model, while still bad, is one of the better models that one can make from this combine data (MSE: 1.28, adjusted R2: 0.18).
CAV = -4.36*40-yard dash – 0.05*weight + 32
Countering the LaRon Landry strategy of becoming freakishly bulked up, strong safeties lose value as they gain weight and gain value the faster they are. This model does not seem to jive with what strong safeties actually have to do in the NFL, which simply suggests, again, that the combine measurements just are not very predictive.
These results suggest that most of the measurements taken at the NFL combine are not predictive of NFL production. There are thousands of possible interaction variables that may be significant predictors of future performance. Further, there are a number of other measurements taken at the combine that were not included in this study that may be more predictive of value than those analyzed here. I also looked at a few broader categories, like “linebackers” and “offensive linemen”, but the models produced for individual positions proved to be much better, so I did not include them. Combined with the offensive models, only the broad jump did not predict future production in any position; most measurements only predicted performance in a few categories. I believe that these models should temper excitement over the results of the combine: because a player puts up a great 40-yard dash time or jumps really high does not say much about how they will perform in the NFL.