Since 2021, NFL teams have spent more than $400 million on 1st round QBs. Considering QB hit rate in the 1st round is just 46% then ~$220 million will have been wasted on QBs who won’t be franchise QBs. That is a lot of money.
With that much money being invested (plus the scouting, coaching, etc), there must be a better way to predict NFL QB performance.
And, that is what I sought out to do – using the PFF premium stats that sort out QB statistics by play type, pressure vs no pressure, blitz vs no blitz, plays <2.5sec vs >2.5sec, etc, we can evaluate how college QBs performed in college and translate that to NFL QB performance.
There are over 900 different statistics for QBs in the PFF premium stats database. The PFF database has these college statistics from 2014 and on (there are some stats before 2014 but many are not broken out by all the different play types). I used a stepwise standard least squares modeling method that narrowed the most important statistics to 68 different statistics (still, a lot). However, developing the model with these statistics (many broken out by different play types), we can predict NFL QB performance using these college statistics with an accuracy of 86%.
You will not find any other model that is this predictive of NFL QB performance. I used NFL QBs who were drafted in 2015 and on and who have had at least 150 snaps in the NFL (~4-5 games worth). The model results are below and speak for themselves:
NFL EPA Per Play vs Predicted NFL EPA Per Play based on College-to-NFL QB Prediction Model (QBs with at least 150 snaps since 2015)

This model that uses 68 different statistics is incredibly predictive of a QB’s EPA per play in the NFL. The interesting thing about this model is how predictive it was for some of the most controversial QBs drafted – Josh Allen, Lamar Jackson, and even Brock Purdy scored high in this model. Zach Wilson, Will Levis, Anthony Richardson, and Josh Rosen scored low (Rosen very low). Nearly all QBs are within 0.1 NFL EPA per play of the predicted NFL EPA per play using this model.
I am not going to divulge the exact details of the model, but I will show some of the more impactful statistics and why I think they translate to the NFL better than others.
Pressure-to-Sack Rate on Non-Blitz Plays
In my previous post, I showed how impactful Pressure-to-Sack rate was for finding elite QBs, especially Pressure-to-Sack rate on plays without a blitz. Each sack has such a negative EPA (-1.87 EPA per sack or 3 sacks is equivalent to a TD) that avoiding sacks is critical for QB play.
And, when defenses only rush 4 (not blitzing), then not allowing a sack in those plays is incredibly important. So, QBs who have a low Pressure-to-Sack rate on non-blitz plays (hello Patrick Mahomes) in college translate very well to the NFL.
This is the single biggest variable within the College-to-NFL QB Prediction Model (among 68 different variables, of course).
NFL EPA Per Play vs Pressure-to-Sack Rate on Plays w/o a Blitz (in College) (QBs with at least 150 snaps since 2015)

Accuracy % on Intermediate Throws
As mentioned in the previous post, overall accuracy % in college does not translate well to NFL QB performance (hello Josh Allen). However, Accuracy % on intermediate throws (10-19 yards downfield) does have a good correlation to NFL QB performance. Breaking down accuracy to this area of the field does bring the “less accurate” passers, such as Josh Allen and Jordan Love, more in line with where the other elite QBs were in college.
NFL defenses are built to defend the run and prevent the big pass downfield. So, being able to attack the intermediate part of the field accurately is a very big part of being an NFL QB. So, being accurate in this region of the field (even if the QB is overall less accurate) is a critical indicator of how that QB will play in the NFL.
NFL EPA Per Play vs Accuracy % on Throws between 10-19 yards (in College) (QBs with at least 150 snaps since 2015)

BTT Rate on Plays >2.5 sec
One statistic that didn’t correlate well individually, but the model required higher weighting for further prediction accuracy, is BTT (big-time throw) rate on plays lasting longer than 2.5sec.
At first glance, the correlation by itself doesn’t seem too strong (R2 value of just 0.014). However, you do see the “bust” rate (QBs with <-0.05 NFL EPA per play) decreases significantly when the BTT rate on these plays is >9.9%, with only PJ Walker, Brett Rypien, Zach Wilson, and Desmond Ridder not performing well in the NFL.
So, although this doesn’t correlate well individually, it does help weed out some possible busts, although there are elite QBs on the low end of the metric, such as Brock Purdy and Lamar Jackson – and that’s why we need those 68 different statistics to accurately predict overall NFL QB play.
This metric does make some sense in that plays that break down (lasting longer than 2.5 sec) is now very dependent on QB play – the QB is making a play outside of the expected play structure. NFL offensive coordinators can’t be perfect in their play calling, so you do need QBs who are capable of making plays outside of the normal play structure. So, this is a critical skill – although one with just a slight correlation.
NFL EPA Per Play vs BTT Rate on Plays >2.5sec (in College) (QBs with at least 150 snaps since 2015)

College-to-NFL QB Prediction Model Summary
In summary, we need a better method of selecting college QBs in the draft, especially in the 1st round. The “eye test” and subjective scouting has not worked – and I don’t blame the scouts or GMs. It is difficult to evaluate human performance, especially in such a complicated sport such as football. It is nearly impossible for a single human to draft the “right” QB.
However, we have so much data now on QB performance that we can utilize this data to more accurately predict NFL QB performance based on how they perform in college. Now, no model, either analytical model or subjective scouting model, will ever be 100% perfect. That is impossible, but if I was an NFL owner, then I would like to have a little more confidence before paying hundreds of millions of dollars on that decision.
So, based on this model, who are the college QBs in the 2025 NFL Draft that should translate to high performing NFL QBs? Which of these draft hopefuls are in the Mahomes/Allen/Jackson/Daniels elite region (if any)? Which college QB draftees are in the Zach Wilson/Dwayne Haskins/Will Levis bust region?
That’s my next post – and it might surprise you.
NFL EPA Per Play vs Predicted NFL EPA Per Play based on College-to-NFL QB Prediction Model (QBs with at least 150 snaps since 2015)

One response to “The Most Accurate Predictor of NFL QB Performance: College-to-NFL QB Prediction Model”
[…] shown in the previous post, I have developed a prediction model based on college QB performance utilizing 68 different […]