Over a year ago I began my master's in Data Science. My dissertation was to create a Deep Learning Neural Network to compare the linear regression predictions compared to other machine learning algorithms in predicting weekly FanDuel points for NFL athletes.

Since graduating this past summer, I have had some nice results based on my projections. I have finished as high as 6th place in major tournament competitions. I have also had finishes at 8th, 16th, 35th, 68th, and numerous top 200.

While I won't share all my secrets, I thought it would be fun to at least share some of the top projections from my weekly model and we can see how they do.


Every week I run a statistical model I built to project weekly NFL player production. The model gets "trained" on the historical NFL data. It includes things like:

If it's home game or an away game for the player

Dome vs Outdoor (natural grass or turf)

What's the player's depth chart status (ie. is he WR1 or WR4)

What the player's average pass attempts/rush attempts/targets for the season going into the week

Who the opponent's starting cornerbacks are for the game

What's the team average points per game, total yards per game, total pass yards per game, total rush yards per game

As far as those averages go, where are they ranked in the NFL

What's the opponents points given up per game, total yards given up per game, passing yards given up per game, rush yards given up per game

The opponent's rankings for those averages

The weather conditions (snowing, raining, cold, etc.)

The average amount of FanDuel points the opponents give up against that player's position

The NFL week (ie. is it early in the season, late in the season?)

Vegas' over/under and spread of the game

And much more

It uses all that info to then correlate a relationship to all those features and how it influences an output (in this case Daily Fantasy Football points). Once it derives a non-linear model for that, it then validates how well it predicts by testing it's predictions against some more data, then readjusts to minimize the error.

After all is said and done, I can get a "score" of the model. Essentially, this gives us a confidence interval as seen below:

QBs Projections - 70% confident the actual points will be +- 6.9 from the predicted points

RBs Projections - 70% confident the actual points will be +- 5.7 from the predicted points

WRs Projections - 70% confident the actual points will be +- 5.6 from the predicted points

TEs Projections - 70% confident the actual points will be +- 4.7 from the predicted points

DEF Projections - 70% confident the actual points will be +- 6.4 from the predicted points

I'll try to post these every Thursday for the rest of the season.


Here are the top ratings for week 14. Sorry I didn't get this posted Thursday, as I would like to post this before Thursday's game. I was having issues and just got it fixed today. I'll also have to work on the images, as they get automatically resized and difficult to read.