Can we predict what sharps are going to do in DFS contests? Turns out, maybe we can!

While it does seem like real-life player analysis is the end all be all, it's not. What we should be doing is evaluating what sharps are doing and emulate that behavior.

With the help of @KoalatyStats , contest data from @FantasyCruncher , and inspirational thoughts from @blenderhd and @BalesFootball (yeah I am just trying to tag the world right now) I've built a model to try and learn what sharps are doing in the $150 power sweep

My process was to take a group of sharp players (top 20 RG rankings) exposures data and aggregate it into the wisdom of the crowds' fashion and train my model to take recent form data and try to learn how sharps are using the data we look at every week.

I know this chart looks a bit crammed, but we can actually see that sharps are using projected ownership (per FC) as huge reasoning as to why they have or do not have certain players in their lineups. It can explain 26% of the variance here

A few things might be able to explain this. Sharps could be...
1. Create their exposures to have a linear correlation with the field (maybe being a bit higher or underweight)

2. Using ownership projections as a starting point, and building from there

They are also using the same opportunity type stats such as WOPR, Rush share, and the like, to make their own decisions. They don't have any different data than we do!

QBs might be the most difficult to predict because our best stat here is CPOE along with Vegas data. Also, QB ownership is always very spread out. Even in a 3-max entry contest like the Power Sweep

You can make whatever claims you want, but these players were all very profitable in the power sweep. 1/3 winning the $100,000 top prize through what I would assume to be the ideal lineup strategy for this specific contest

For those interested in the more machine learning/python side of this, Optuna is the greatest parameter tuning package and it's not even close. This function pictured that I built made for a very fast hyper parameter tuning experience.

I also can't forget to thank @friscojosh for not blocking me when I cold dm'd him back when I was starting this.