NFL Draft 15 Model
Li Liang
Friday, April 24, 2015
Intro
I wanted to see if I could create a model to predict if a candidate should be drafted in the 2015 NFL draft using historical combine data (1999-2014) and Madden 15 player ratings. I’m a big fan of the Madden games and I believe they do a good job at rating players on their roster. I downloaded the data from the following sites and assumed if they had a Madden rating then they should be drafted. Note: There were a few draft candidates with the same name as current NFL players. These were zeroed out.
Data overview
There are 5346 players total form 1999-2015 with 14 variables. At this
point, there appears to be many NA’s for Wonderlic, Bench Press, Shuttle and 3cone. The lack of Wonderlic scores was a surprise because I thought most candidates would have that published.
## 'data.frame': 5346 obs. of 15 variables:
## $ Year : int 2000 2011 2010 2010 2015 2011 2009 2005 2001 2013 ...
## $ Name : Factor w/ 5262 levels "A.C. Leonard",..: 4877 2861 298 1995 3664 1477 3664 746 3664 3267 ...
## $ College : Factor w/ 335 levels "Abilene Christian",..: 150 89 44 209 89 209 280 215 330 151 ...
## $ POS : Factor w/ 18 levels "C","CB","DE",..: 14 11 18 4 18 15 3 11 15 15 ...
## $ Height : int 77 75 70 76 74 73 76 75 69 73 ...
## $ Weight : int 211 270 186 295 202 213 274 236 207 230 ...
## $ Wonderlic : int 33 NA NA NA NA NA NA NA NA NA ...
## $ Bench.Press: int NA 30 13 23 NA 21 24 20 13 24 ...
## $ VertLeap : num 24.5 36.5 33.5 30.5 NA 34.5 31 45.5 39.5 31.5 ...
## $ BroadJump : int 99 125 105 114 NA 130 110 130 NA 118 ...
## $ Shuttle : num 4.38 4.37 4.18 4.48 NA 4.18 NA 4.13 NA 4.24 ...
## $ X3Cone : num 7.2 6.95 6.98 7.32 NA 7.28 NA 7.12 NA 6.75 ...
## $ X40.Yard : num 5.28 4.62 4.56 5.04 4.59 4.37 5 4.65 4.38 4.6 ...
## $ Rating : int 93 93 92 95 0 89 89 96 89 82 ...
## $ Draft : Factor w/ 2 levels "0","1": 2 2 2 2 1 2 2 2 2 2 ...## Year Name College
## Min. :1999 Anthony Davis : 3 Florida : 111
## 1st Qu.:2003 Brandon Williams: 3 Florida State : 110
## Median :2008 Chris Brown : 3 Georgia : 105
## Mean :2008 Chris Davis : 3 Alabama : 99
## 3rd Qu.:2013 Chris Jones : 3 Southern Californiaifornia: 93
## Max. :2015 Josh Davis : 3 Oklahoma : 92
## (Other) :5328 (Other) :4736
## POS Height Weight Wonderlic
## WR : 684 Min. :65.00 Min. :155.0 Min. : 4.00
## CB : 534 1st Qu.:73.00 1st Qu.:207.0 1st Qu.:20.00
## RB : 467 Median :74.00 Median :236.0 Median :25.00
## DE : 429 Mean :74.04 Mean :245.1 Mean :24.53
## OT : 418 3rd Qu.:76.00 3rd Qu.:287.8 3rd Qu.:30.00
## DT : 410 Max. :82.00 Max. :386.0 Max. :48.00
## (Other):2404 NA's :5061
## Bench.Press VertLeap BroadJump Shuttle
## Min. : 0.00 Min. : 0.00 Min. : 74.0 Min. :0.000
## 1st Qu.:17.00 1st Qu.:30.00 1st Qu.:108.0 1st Qu.:4.190
## Median :21.00 Median :33.00 Median :114.0 Median :4.340
## Mean :21.32 Mean :32.86 Mean :113.3 Mean :4.383
## 3rd Qu.:25.00 3rd Qu.:36.00 3rd Qu.:120.0 3rd Qu.:4.560
## Max. :51.00 Max. :46.00 Max. :147.0 Max. :5.560
## NA's :1702 NA's :1052 NA's :1183 NA's :1682
## X3Cone X40.Yard Rating Draft
## Min. :0.000 Min. :4.210 Min. : 0.00 0:3913
## 1st Qu.:6.990 1st Qu.:4.550 1st Qu.: 0.00 1:1433
## Median :7.220 Median :4.710 Median : 0.00
## Mean :7.304 Mean :4.795 Mean :20.16
## 3rd Qu.:7.570 3rd Qu.:5.000 3rd Qu.:64.00
## Max. :9.120 Max. :6.050 Max. :99.00
## NA's :1689 NA's :15I wanted to explore the distribution of the 7 combine metrics. Wonderlic, Bench Press, Vertical Leap, Broad Jump, Shuttle, 3Cone and 40 yd dash. Grey being the previous draft candidates and red being the 2015 candidates.
Modeling
I created a simple logistic classification model using past candidates combine performance metrics and Madden ratings to determine if the current candidates should be drafted or not.
Due to the poor data quality of Wonderlic metric, I’m setting all NA’s to be zero. I am also using complete observations to train the model. Luckily, there are 2048 complete observations, this should be a good representative sample. Looks like out of 2408 observations, 770 or 38% were drafted to the NFL.
## Year Name College POS
## Min. :1999 Aaron Williams : 2 Nebraska : 47 CB : 254
## 1st Qu.:2003 Larry Brown : 2 Iowa : 43 OT : 242
## Median :2007 Levi Brown : 2 Notre Dame: 43 DE : 227
## Mean :2007 Michael Johnson: 2 Georgia : 42 OLB : 218
## 3rd Qu.:2011 Mike Walker : 2 Oklahoma : 42 RB : 206
## Max. :2014 Ryan Grant : 2 Clemson : 40 DT : 202
## (Other) :2399 (Other) :2154 (Other):1062
## Height Weight Wonderlic Bench.Press
## Min. :65.00 Min. :166.0 Min. : 0.0000 Min. : 0.0
## 1st Qu.:73.00 1st Qu.:212.0 1st Qu.: 0.0000 1st Qu.:17.0
## Median :74.00 Median :248.0 Median : 0.0000 Median :21.0
## Mean :74.28 Mean :253.4 Mean : 0.4529 Mean :21.4
## 3rd Qu.:76.00 3rd Qu.:300.0 3rd Qu.: 0.0000 3rd Qu.:25.0
## Max. :80.00 Max. :386.0 Max. :34.0000 Max. :51.0
##
## VertLeap BroadJump Shuttle X3Cone
## Min. : 0.00 Min. : 82.0 Min. :0.000 Min. :0.000
## 1st Qu.:29.50 1st Qu.:106.0 1st Qu.:4.210 1st Qu.:7.000
## Median :33.00 Median :114.0 Median :4.370 Median :7.260
## Mean :32.55 Mean :112.5 Mean :4.404 Mean :7.329
## 3rd Qu.:35.50 3rd Qu.:119.0 3rd Qu.:4.590 3rd Qu.:7.620
## Max. :45.50 Max. :139.0 Max. :5.560 Max. :9.040
##
## X40.Yard Rating Draft
## Min. :4.210 Min. : 0.0 0:1640
## 1st Qu.:4.570 1st Qu.: 0.0 1: 771
## Median :4.750 Median : 0.0
## Mean :4.827 Mean :24.2
## 3rd Qu.:5.070 3rd Qu.:68.0
## Max. :6.000 Max. :99.0
## Training 1st trial
I fit the classification model at first using position, height, weight, wonderlic, bench press, vertical leap, broad jump, shuttle, 3cone and 40 yd dash as predictors of whether or not the candidate is drafted.
From the model output, it does not seem bench press is too significant of a predictor. I feed the input data back into the model to validate. From the confusion matrix, the model drafted 412 out of 2408 with 157 false positives and 515 false negatives. This gives a precision of 770 / (770 + 157) = 83% and recall of 770 / (770 + 515) = 60%.
##
## Call:
## glm(formula = Draft ~ POS + Height + Weight + Wonderlic + Bench.Press +
## VertLeap + BroadJump + Shuttle + X3Cone + X40.Yard, family = binomial,
## data = trainingdata)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -2.4119 -0.8505 -0.5596 1.0391 2.5522
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) 19.805418 3.899124 5.079 3.79e-07 ***
## POSCB 0.367359 0.549945 0.668 0.504139
## POSDE -0.791971 0.340890 -2.323 0.020166 *
## POSDT -0.639610 0.290455 -2.202 0.027659 *
## POSFB -1.059816 0.468788 -2.261 0.023774 *
## POSFS -0.302096 0.536582 -0.563 0.573434
## POSILB -0.670943 0.424751 -1.580 0.114195
## POSK 8.688904 882.745206 0.010 0.992147
## POSLS -23.844516 882.745651 -0.027 0.978450
## POSOG 0.007986 0.292633 0.027 0.978228
## POSOLB -0.464583 0.401690 -1.157 0.247448
## POSOT 0.104489 0.287851 0.363 0.716606
## POSP -21.523642 882.745225 -0.024 0.980547
## POSQB -14.405258 271.048630 -0.053 0.957615
## POSRB -0.718476 0.483170 -1.487 0.137014
## POSSS -0.211465 0.523777 -0.404 0.686411
## POSTE -0.568981 0.386264 -1.473 0.140741
## POSWR 0.482998 0.536366 0.901 0.367854
## Height -0.111030 0.035038 -3.169 0.001531 **
## Weight 0.031028 0.004994 6.213 5.20e-10 ***
## Wonderlic 0.101337 0.022027 4.601 4.21e-06 ***
## Bench.Press 0.025175 0.010693 2.354 0.018555 *
## VertLeap -0.074123 0.020599 -3.598 0.000320 ***
## BroadJump 0.043611 0.011557 3.773 0.000161 ***
## Shuttle 2.824002 0.373751 7.556 4.16e-14 ***
## X3Cone -2.865776 0.266635 -10.748 < 2e-16 ***
## X40.Yard -3.011864 0.455140 -6.617 3.65e-11 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 3022.0 on 2410 degrees of freedom
## Residual deviance: 2592.3 on 2384 degrees of freedom
## AIC: 2646.3
##
## Number of Fisher Scoring iterations: 13## lcmpred
## 0 1
## 0 1483 157
## 1 515 256Training 2nd trial
Looking at the predictive power and collinearity tables, I decided to take out bench press (not significant), wonderlic (poor data), broad jump (not significant), height and weight (positionally unique) for the next iteration. This time the model drafted 361 out of 2408 with 150 false positives and 559 false negatives. A slight bump in false negatives but not too big of a difference.
##
## Call:
## glm(formula = Draft ~ POS + VertLeap + Shuttle + X3Cone + X40.Yard,
## family = binomial, data = trainingdata)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -1.8869 -0.8738 -0.5990 1.1224 2.1954
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) 23.84178 2.66613 8.942 < 2e-16 ***
## POSCB -2.30202 0.37840 -6.084 1.18e-09 ***
## POSDE -1.62892 0.31371 -5.192 2.08e-07 ***
## POSDT -0.39844 0.28372 -1.404 0.1602
## POSFB -2.39270 0.42174 -5.673 1.40e-08 ***
## POSFS -2.66838 0.39703 -6.721 1.81e-11 ***
## POSILB -2.02460 0.36528 -5.543 2.98e-08 ***
## POSK 6.60800 882.74505 0.007 0.9940
## POSLS -22.60638 882.74549 -0.026 0.9796
## POSOG 0.19366 0.28589 0.677 0.4982
## POSOLB -1.83216 0.33986 -5.391 7.01e-08 ***
## POSOT 0.28289 0.27123 1.043 0.2969
## POSP -22.66549 882.74508 -0.026 0.9795
## POSQB -15.63707 302.53842 -0.052 0.9588
## POSRB -2.54794 0.37217 -6.846 7.58e-12 ***
## POSSS -2.41490 0.40622 -5.945 2.77e-09 ***
## POSTE -1.89693 0.33512 -5.660 1.51e-08 ***
## POSWR -2.01588 0.39129 -5.152 2.58e-07 ***
## VertLeap -0.03141 0.01750 -1.795 0.0726 .
## Shuttle 2.95647 0.36278 8.149 3.65e-16 ***
## X3Cone -2.74950 0.25454 -10.802 < 2e-16 ***
## X40.Yard -3.12900 0.41634 -7.516 5.67e-14 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 3022.0 on 2410 degrees of freedom
## Residual deviance: 2703.9 on 2389 degrees of freedom
## AIC: 2747.9
##
## Number of Fisher Scoring iterations: 13## lcm2pred
## 0 1
## 0 1490 150
## 1 559 212Results
I tested the second model with the 2015 draft class data and cross referenced with nfl.com rankings. http://www.nfl.com/top50
Results below:
| Name | College | POS | NFL.com Ranking |
|---|---|---|---|
| Dorial Green-Beckham | Oklahoma | WR | 26 |
| Phillip Dorsett | Miami | WR | 47 |
| Mario Alford | West Virginia | WR | |
| Kenny Bell | Nebraska | WR | |
| Justin Hardy | East Carolina | WR | |
| Ali Marpet | Hobart & William Smith | OG | |
| Trae Waynes | Michigan State | CB | 10 |
| Jalen Collins | Louisiana State | CB | 30 |
| Xavier Cooper | Washington State | DT | |
| Jake Fisher | Oregon | OT | 45 |
| Kevin White | West Virginia | WR | 2 |
| Amari Cooper | Alabama | WR | 3 |
| J.J. Nelson | UAB | WR | |
| Senquez Golson | Mississippi | CB | |
| Stefon Diggs | Maryland | WR | |
| Kaelin Clay | Utah | WR | |
| Cameron Clear | Texas AM | OT | |
| Chris Conley | Georgia | WR | |
| Deon Long | Maryland | WR | |
| Derrick Lott | Tennessee-Chattanooga | DT | |
| DeAndrew White | Alabama | WR | |
| Tyler Lockett | Kansas State | WR | |
| Troy Hill | Oregon | CB |