The Study: RB Combine Data

Thesis:

Use data collected every year at the NFL combine to predict NFL success based off of the combines of previous players who were determined to be successful.

 


Theory:

Determining success is a difficult process. For each debate, it could be the eye test, the game logs, the yearly numbers, averages, added value, or a million other systems. In this theory, it was decided that the way to test success would be a combination of multiple.

 

First, I wanted a back who could take over a game and cause the team to win a singular game. To attack this, I checked all of the players who had games where they rushed for over 120 yards from 2007 to 2017. Then, they were removed if they only had one game in this area, which weeded out players such as Jonas Gray, who were clearly not successful backs long term but, showed a bright flash.

 

Next, I tracked the added value (per sports reference) of all of these players per year. It is used as a general rating to gauge and re-confirm the “eye-test.” This per year measure excluded any year that they played less than 7 games because it wasn’t believed there was enough data in that season for said player to determine how they impacted the team’s entire season. For example, the highest AV/yr was Ladainian Tomlinson, which confirms the eye-test that he’s among the greatest RBs in this study’s era and Alfred Blue was among the lowest, confirming the eye test that he is not a talent worth tracking. From this data, I removed all of the players that scored below a 4.5 per season, weeding out players such as Joe McKnight, Michael Bush, or Ladell Betts.

 

Following that sort, I made two sorts that were there to check one another. For example, if one placed a player in the top-third, it was disregarded if they were bottom-third in the second. This was used to sort the players into 3 categories in the following table to get multiple averages/quartiles/points to measure the incoming rookies against. I wouldn’t want a back who did not put together a consistent or strong body of work over the course of a season. This was weeded out by checking what percent of their career seasons were over 1000 yards and if that was at or below one-third of the time. This removed players such as Jeremy Hill or Fred Jackson. After that, I looked to the YPC numbers because it’s undesirable for a back to not get “enough” yards per time their number is called. That number was determined to be 3.9 but, was considered the less important value when determining a player’s group in the following table. This removed players such as Cedric Benson, who was a solid RB but, not determined to be worth sorting players by.

 


Initial Strategy:

The groupings were created based off of the fourth paragraph (prior paragraph to this one) of the Theory section. The numbers found here were sorted into three separate groups. The first group was players who were top-third in both seasons above 1000 per season with more than 7 games played as well as top-third in career yards per carry. The second group was top-third in one of the categories and middle-third in the other. The final group was middle-third in both or top-third in one and bottom-third in the other. These categories don’t accurately represent the “eye-test” of the three tiers of players listed and the “group number” or “tier level” did not impact the weight of the results. They were used to provide a most-strict number and least-strict number for the incoming rookies to be tested against, ensuring that a, for example, 107 broad jump wasn’t treated the same as a 117 broad jump.

 


Data Collection and Analyzation:

The combine data collected was used in three different ways to create a multitude of checks to ensure that as many bases were covered as possible.

 

The first usage of the data used the 3 groups method mentioned above in the “Initial Strategy” section and took the minimum, first quartile, average, third quartile, maximum, median and any outliers. Naturally the quartile system caused some confusion early on as a third quartile 3 cone (a number that is prefered to be lower) would be require different treatment than a third quartile broad jump (a number prefered to be higher). During the sorting, this was double and triple checked to ensure that the numbers were in the proper order. Using the 7 collected points of data, they were put next to each other for each combine drill. From here they were sorted by which was the most-strict (lowest for high-number categories and highest for low-number categories) to the least-strict. From here, this was ran against the players combine drills and players began to fail the athletic tests. For example, 4.55 was the “most strict” and 4.6 was the “least strict”for 40 yard dash. If a player charted a “least strict” fail, they were immediately failed, regardless of the rest of their combine. In 2018’s combine, an example of this is Lavon Coleman of Washington who charted a 4.65 in the 40 but, passed the rest of his drills. Now, I’m not saying that there is absolutely no chance that this player succeeds. There are always going to be data points such as, LeGarrette Blount ran a 4.74 and is still a fine RB but, they likely lack the athletic ability to overcome the odds in the NFL. If we are looking at a dataset with with more than 70 data points, then it’s best to go by the means not the exceptions.

 

The second usage of data was primarily used to determine which drills were to be sorted by first and which drills meant nothing to success. This was done by running each of the scores of the players who made it to the group sort (almost 50 players) drills against their AV/yr. For example, this determined that vertical jump and broad jump had no correlation to AV/yr, displayed in their very high P-values. The p-value is the level of marginal significance within a statistical hypothesis test representing the probability of the occurrence of a given event. The hypothesis here is that as x-combine-drill gets a better score then, (the given event) AV/yr should go up. After a linear regression was done on each of the drills, they were sorted by importance. The order was 10 yard split(0.07), 40 yard dash(0.09), 20 yard shuttle(0.14), and then 3 cone(0.15).  Ideally, you would use the numbers that are below 0.05 but, since none were, I took the ones that were within 0.10. The events left out of the study were broad jump(0.37), vertical jump(0.46), and bench press(0.70) due to their high p-value. For pure entertainment value, I tested the round each player was drafted in against AV and that was the best value by far at 0.002.

 

The third usage of data was the “eye-test” theory. Here all of the data points (players) were plotted on a graph with the combine scores on one axis and the AV/yr on the other axis. The goal was to see if there were any visual lines were clearly apparent of the 4 events that were decided on in the previous paragraph. The 10 yard split data plot lead believe that if the player is over 6 AV/yr then they likely have a 1.6 or lower split. Only 4(Doug Martin, Mark Ingram, Devonta Freeman, and Ryan Grant) of the 32(12.5%) of the players over 6AV/yr had greater than a 1.6 split and 2 of those 4 were first round picks (which infers to me that they had clear talent to make it in the NFL as well as round being the most significant data column to AV/yr in these models). Another interesting form of data in 10yd split is that there were only 5 players below 6AV/yr(and none of them below 4.5) and below 1.55s while there were 17 in the above 6AV/yr category. The 40 yard dash displayed the 6 AV/yr line for scores with above 6AV/yr and with 4.6 or higher only having 3(9.3%) players (Mark Ingram, Ryan Grant, and Spencer Ware) but, 6AV/yr or lower and 4.6 or higher having nearly triple that at 8. The 20 yard shuttle cutoff was 4.27 with 5(15.6%) being above and double that amount being below the 6AV/yr mark. Interestingly, 4(Mark Ingram, Marshawn Lynch, Adrian Peterson, and Joseph Addai) of the 5 in the above 6AV/yr list were 1st round picks, and the other was Maurice Jones-Drew who was a 2nd round pick and (from reading old scouting reports online) would’ve likely been a first round pick if he was 2 inches taller. The 3 cone cutoff was 7.1 with 4(12.5%) of the 32 (Mark Ingram, Demarco Murray, Michael Turner and CJ Anderson) and there being double as many in the sub-6AV/yr section. In this part of the study is where it was discovered that the earlier cutoff AV/yr should have been 6 rather than 4.5.

 


Data Usage:

Saquon Barkely at the 2018 NFL Combine As mentioned in the “Initial Strategy” and second paragraph of the “Data Collection and Analyzation” sections, the “most-strict” and “least-strict” numbers from each of the 3 groupings were run over the 4 significant drills described in the third paragraph of the “Data Collection and Analyzation” section to get an initial list of players who failed number tests. Players were not penalized for skipping drills, although those who skipped drills will be listed in separate lists if they passed the initial sort. Here are the numbers for each of the drills in the format of (“most-strict”/”least-strict”): 10 yard split(1.64/1.66), 40 yard dash(4.55/4.6), 20 yard shuttle(4.4/4.55), and then 3 cone(7.28/7.5). The players who failed are (listed with school | drill or drills failed):

 

– Mark Walton | Miami | 40 yard Dash

– Dimitri Flowers(FB) | Oklahoma | All

– John Kelly | Tennessee | 20 yard Shuttle

– Roc Thomas | Jacksonville State | 40 yard Dash, 10 yard Split

– Darrel Williams | LSU | 40 yard Dash, 10 yard Split

– Donnie Ernsberger(FB) | Western Michigan | All

– Lavon Coleman | Washington | 40 yard Dash

– Demario Richard | Arizona State | 40 yard Dash, 10 yard Split

– Justin Crawford | WVU | 40 yard Dash

– Kamryn Pettway | Auburn | 40 yard Dash

 

Next, I turned to the data obtained in the fourth paragraph of “Data Collection and Analyzation” section to run over the data. To review said paragraph in short, the benchmarks were: 10 yard split(1.6, 1.55 would be considered an elite test), 40 yard dash(4.6), 20 yard shuttle(4.27), and then 3 cone(7.1). Players who passed the initial test but, raised flags in this test were(listed with school | drill of concern and | drills skipped, if necessary):

 

– Nick Chubb | Georgia | 10 yard Split

– Royce Freeman | Oregon | 10 yard Split

– Kalen Ballage | Arizona State | 20 yard Shuttle

– Ryan Nall | Oregon State | 10 yard Split

– Nyheim Hines | NC State | 3 Cone, 20 yard Shuttle

– Javarion Frankin | Western Michigan | 20 yard Shuttle

– Kerryon Johnson | Auburn | 3 Cone | 40 yard Dash (and 10 yard Split)

– Bo Scarbrough | Alabama | 20 yard Shuttle | 3 Cone

 

After that, I was left with a list of 5 players who tested with an elite combine. Those players were (listed with school | drills skipped):

 

– Nick Chubb | Georgia | 10 yard Split

– Royce Freeman | Oregon | 10 yard Split

– Kalen Ballage | Arizona State | 20 yard Shuttle

– Ryan Nall | Oregon State | 10 yard Split

– Nyheim Hines | NC State | 3 Cone, 20 yard Shuttle

– Javarion Franklin | Western Michigan | 20 yard Shuttle

– Justin Jackson | Northwestern

– Chase Edmonds | Fordham

 


Historical Data:

As you can tell from above, only Justin Jackson and Chase Edmonds completed all of the drills and did not skip any drills. Since 2015, only 5 other players have done this and they are David Johnson, Melvin Gordon, DeAndre Washington, Joe Mixon and Aaron Jones. 3 of them are full-time/workhorse backs and the other 2 are rotational pieces/occasional starters. Good company to be in. If there is 1 late-round RB I would be pounding the table for in a draft room is now Justin Jackson who, for me, passes both the eye test and flawlessly passes the numbers test.

 

Another interesting factoid I came across when working on this list is that, in the past 4 drafts, the fastest RB in the draft has been a sub-par/cut-level player at best with the highest AV/yr being Jeremy Langford with a 4 (only one with more than 15 rushes in a season) and currently being on the Jets practice squad, his third team in as many years. Sorry about that Nyheim Hines. His company is De’Anthony Thomas, Jeremy Langford, Keith Marshall, and TJ Logan.

 

Finally, here is where I show you that no data or set of standards comes without failures and that at the end of the day, they are only numbers. Kareem Hunt, Jordan Howard, Dalvin Cook, and Alex Collins all failed my test. 3 of them were rookie of the year contenders in their respective rookie seasons. This data set appears to side against the powerful one-cut runner, which salvages my hope for John Kelly.




Be the first to comment

Leave a Reply

Your email address will not be published.


*