Modeling is the most powerful and underutilized tool in how we as a population analyze and understand football. Often I see people use metrics to try and understand/predict football. They use these metrics individually to understand small segments of a grand picture. For instance, one popular metric I see used is Market Share (MS), but what if instead of using MS alone, we used MS in combination with other variables to try and predict the response variables we truly care about. Let’s use a simple example here to understand why linear regression is better than metrics alone. Figure 1 is an “Simple Linear Model” with a singular variable (YPR), this would be no different than comparing them in Excel and fitting a trend line.

All the left is doing is using Yards per Rec (YPR) to predict yards with no other help. Figure 2 adds MS Targets to the model to create a multiple linear regression and it increases predictability. This is a very simple example of how in tandem metrics can be very powerful. The predictability increase is huge here since MS targets is so strong by itself, but we can tell both variables are significant since the Adjusted R^2 is so close to the General R^2. Little by little, we can understand the complete picture though. Every .01 of Adjusted R^2 gets us closer to 1.

There are a TON of different ways to model, many are listed to the left. Each modeling method has its strengths and weaknesses. No one modeling method is perfect, but some do apply very well to certain situations.

**Neural Network Steps**

One modeling type I want to try and explain is the Neural Network. Neural Networks are BY FAR the best model to use when all you care about is predicting an outcome. They are almost always EXTREMELY complex models that have very complicated formulas. I am going to try and explain the basics of a Neural Net, but this is a very basic intro to them.

- Creates the initial model based on random assumptions from your input variables
- Runs the model, which has an initial R^2. No model is perfect and will always have error.
- Model 1 has error, The Neural Net creates another model (with your variables) that tries to predict Model 1’s error. We shall dub this Model 2.
- Now Model 3 is a model comprised of Model 1 and Model 2. Model 3 has error, the Neural Net then creates a Model 4 that predicts Model 3’s error.
- These steps happen n times with diminishing returns per model created

That is a TL;DR and not complete way of explaining what a Neural Network is. This video has terrible sound quality but is very informative on how Neural Networks work.

https://www.jmp.com/en_ch/events/ondemand/mastering-jmp/neural-networks.html

I learned about Neural Networks while in an Applied Linear Regression class at the University of Georgia, and thought “Why aren’t these being used more?” The easy answer is that they are immensely hard to explain and the formulas are VERY ugly. They are what we in the statistics world call “Black Boxes” meaning when you throw the variables in, they just come out. The process in the middle isn’t that concerning.

I have used Neural Networks extensively (check my Twitter @**DFF_Koala**) to try and predict football. Topics include: the 2018 draft, 2018 rookie scrimmage yards, and predicting the 2018 season for Running Backs. Neural Networks are very versatile and can be used to do almost any type of prediction. As stated earlier though, they suffer in simplicity. In Neural Networks, you somewhat throw the kitchen sink at the model by including as many variables as you have. Since the Neural Network can omit useless variables, it isn’t that big of a deal. Here is an example of using a Neural Network using the SAME variables as before (MS Targets & YPR). Since we are only using 2 variables the improvement is minimal, but still nice improvement (.0154).

I would share the formula but it would take up 3 PAGES (I checked). In summary, Neural Nets are complicated, but powerful. If we decide to learn how they work, they could be the future of football as we know it. Below I included how much of an impact the Neural Network had on my 2018 RB draft model.

**Multiple Linear Regression:**

**Neural Network:**

Notice how much tighter, and better the predictions become when using a Neural Network. Below is how the Neural Network predicted the 2018 NFL Draft. Sorry for the messiness, wanted to let you see how each player was predicted. The Neural Network did this BEFORE the draft and achieved a .65 R^2 for 2018. To me, that is amazing. It had some exceptional hits like Rashaad Penny, Nick Chubb, and Saquon Barkley. It also had some bad misses like Josh Adams, Akrum Wadley, and Nyheim Hines. All models will have error though, and this model predicted better than almost EVERY single draft expert to my knowledge. Neural Networks are powerful, and we need to start using them. Neural Networks can be applied to fantasy football too, which is my next goal! Follow me @**DFF_Koala** to stay updated!

*Are you looking to get some action in on today’s games? Head over http://GTBets.eu for all of your sports betting needs with updated odds covering all major sports. The bottom line… GTBets.eu will have you placing your first bet within minutes with their simple to use registration process. So, register now and earn up to $500 in free money! Simply add FACTORY in the “referred by” field to get the highest possible welcome bonus! *

## Leave a Comment