2023 RB Model – Using Neural Networks

The Model

The goal of this model is to use college football statistics to predict fantasy football success for a college running back prospect to the NFL. To do this, I have created two models. One which predicts fantasy points per game (PPR) and one that predicts NFL draft position using a neural network. Last year, I used a linear regression model. Not every metric has a linear relationship to the input and output variables, which is where linear regression comes up short. Neural Networks are the computerized representation of the human brain that finds complex relationships between input and output nodes to create weights that help predict the desired outcome.

The Results 

The data was trained on 695 RBs in the classes of 2017-2022, and the prediction was on 328 RBs in the classes of 2023-2025. The sum of squared error for the fantasy points per game model is  0.4803, and the draft position model is 2.71474. The neural network is attempting to minimize the sum of squared error as an objective function within the model, which minimize variation in the prediction from the actual result. The input variables are: 

1. Weight: Recruit in 247
2. Power 5
3. College Rushing Fantasy Points Over Expectation Per Week for years 1-4 out of High School 
4. College Receiving Fantasy Points Over Expectation Per Week for years 1-4 out of High School 
5. Games Played: Years 1-4 out of High School 
6. College Rush Attempts  for years 1-4 out of High School 
7. College Receptions  for years 1-4 out of High School

The Prediction

Below I have identified the Top 60 RBs in the model (Sorted by Fantasy & Draft QB Index) for the class of 2017-2025, the Top 40 RBs in the model for the class of 2023-2025, and a separate Year 1 model comparing some current NFL players college ranking versus some college prospects from the 2024-2025 classes. 

All input variables are weighted, but some of the most influential variables for the fantasy points per game model are: 

1. Receptions in year 1
2. Rush attempts in year 3
3. College Rushing Fantasy Points Over Expectation Per Week in year 2
4. Games played in year 1
5. College Receiving Fantasy Points Over Expectation Per Week in year 2. 

All input variables are weighted, but some of the most influential variables for the draft position model are: 

1. Games played years 1-3
2. Rush attempts in year 1
3. Receptions in year 2
4. College Rushing Fantasy Points Over Expectation Per Week in year 1
5. College Receiving Fantasy Points Over Expectation Per Week in year 2-3.

A few takeaways: 

• 2023 Rookie Drafts (Keep Trade Cut Devy 2023 Rankings): 
  • I would take Bijan Robinson (RB1) as the 1.01 in rookie drafts, but Jahmyr Gibbs (RB2) is a great 1B option at RB. Bijan has the fifth-highest draft grade from the model behind Breece Hall, JK Dobbins, Christian McCaffrey, and Saquon Barkley. Gibbs has a top-three fantasy points grade and is positioned between Christian McCaffrey and Alvin Kamara.
• Zach Charbonnet (RB4), Israel Abanikanda (RB13), Tank Bigsby (RB6), and Devon Achane (RB8) are the next cluster of RBs. Both Zach Charbonnet and Israel Abanikanda provide better fantasy points grades while Tank Bigsby and Devon Achane provide better draft grades. I would be targeting Zach Charbonnet over Tank Bigsby and Devon Achane but Israel Abanikanda is the best value of the group, who I might wait to draft if I am forced to choose between Tank Bigsby and Devon Achane.
• Sean Tucker (RB3) and Zach Evans (RB5) fall significantly in the model. Sean Tucker still shows promise as a good fantasy RB. Zach Evans is someone I am still interested in cautiously against the wishes of my model. 
• Kendre Miller (RB7) falls off the chart due to his lack of receiving. 
• I will be drafting Deuce Vaughn at cost, which will probably be a fourth-round rookie pick. 

• 2024 Class  (Devy Rankings at campus2canton): 
  • The year 1 model had TreVeyon Henderson (RB1), Raheim Sanders (RB2), and Will Shipley (RB3) ranked in that order. All three players look like strong prospects. Donovan Edwards (RB5) is the clear fourth option for me. Considering the year 2 data, the model ranks this trio as Will Shipley, Rahiem Sanders, Donovan Edwards, then TreVeyon Henderson.  
• Braelon Allen (RB4) provides a descent draft grade but is being discounted by the model due to his lack of receiving. 
• Roman Hemby (RB9) and Jarquez Hunter (RB10) are two other names to have on the radar for the 2024 class over their current rankings. 

• 2025 Class  (Devy Rankings at campus2canton):
  • Nicholas Singleton (RB1) and Quinshon Judkins (RB2) are the two top-ranked RBs in the 2025 class but my model clearly likes Nicholas Singleton as the top player. 
• Trevor Etienne (RB8) is another interesting name who I would be moving up over TreVonte’ Citizen (RB4), Jamarion Miller (RB5), and Jaydon Blue (RB6) while considering him as RB3 in the class. 
• Ollie Gordon (RB22) and Damien Martinez (RB13) are two deeper names to have on the radar for the 2025 class over their current rankings. 
• I will be cautiously fading Kaytron Allen (RB6) at his current price. Although he pops in my model, I believe Nicholas Singleton is the premier player and Kaytron Allen relived him in more favorable situations. The average win probability when Singleton played was 70% and when Allen played it was 76%. Comparing another college football duo, Javonte Williams played when the win probability was 65% and Michael Carter was 69%. Although I do account for win probability as a situational element of my College Rushing Fantasy Points Over Expectation metric I do think Penn State will regress, which will impact the game script and the team will favor Singleton more. For reference the average win probability for Penn State in 2021 was 62% and in 2020 was 43%.