Using Machine Learning and Wisdom of the Crowd: Week 1 DFS
Explaining my plays and breaking down mistakes
Credit: New York Post
In this space, I will review the week I had in DFS using a model I created using three sources of information:
Establish the run has some of the best projections in the business. From what I gather, they use a combination of machine learning, film analysis, and expert “gut” to arrive at numbers that have beaten the DFS market consistently. I was able to aggregate these floor and ceiling projections with Gridiron AI’s machine learning projections, which leverages millions of data points from the last 10-15 years of NFL data. For the final piece, I also averaged in the projections gathered by FFanalytics. the sources are here:
My theory is that inaccuracies that exist in the FFanalytics aggregation of expert consensus will be canceled out by a sharp group of experts at ETR and a noise-reducing machine learning model from Gridiron AI. It seems that I was correct after week 1 as my use of the model netted a 31% ROI in Week 1 DFS cash and small-field GPP tournaments. I was a few decisions away from having an incredible week. The spreadsheet put me on a lot of the correct fades and exposed me to awesome low-owned plays by simply “playing the players with the greenest boxes”. The simplicity of this should help keep consistent returns on the year as long as I stay disciplined. Either that or I just got lucky and the model sucks.
I probably left some money on the table this week because I didn’t listen to the model 100% of the time.
Variables that matter in the spreadsheet:
Price - obviously price matters, but it isn’t a huge priority to me. higher priced players don’t need to be super-low owned (<10%) for me to play them. “Middling” ownership (12-18%) is plenty low for some of these expensive players.
Sims Projection: the median projection for the player. This is what the machine learning model, experts, and ETR think is the most likely outcome for the player.
Sims Ceiling: This projection is a “what if everything goes right” projection. It’s less likely to happen, but given certain factors, some low-projected guys have high ceilings. This column is incredibly important.
DK Value: Calculated by ETR this is the value of the player given their projection and what players at their price usually score. Most places use a price/$ approach to value, but this favors cheap players too much.
DK Ownership percentage: This column tells us what percent of the field will play this player in the DK Millionaire Maker contest. This is a guage of what the public thinks about a player. Very useful.
Here is the cash lineup I entered. the late-minute news on Sermon and Moss being benched really threw me off. I panicked. Definitely a mistake, but my lineup did well enough to win more than it lost. So, we will take it. In short, I played the players with the most “value'“ while largely ignoring the ownership percentage. Play the best plays.
I invested around 240 dollars or so in this lineup, so despite some bad run-outs, it did really well.
For small-field tournaments (a few thousand entries at most), I started at the quarterback position. I used the QB’s with the highest projected ceiling and built different stacks with them because it makes sense to pair your high-scoring QB with their pass catchers. If it misses, so what. Here’s what the QB position looked like.
So, as you can see. There are a ton of what I call “3-Star” players here. These are simply guys with 3 green boxes which means they have a high ceiling, high value, and high projection. The median projection is less important in GPPs, so I ignore it. Basically, non of these QBs outside of Rogers hurt you, and the Sims caught that. Another note: ownership was pretty flat at QB( it usually is). Because there were plays at so many different price points, building out a bunch of stacks is a great way to diversify your lineup exposure without just throwing stuff against the wall. I faded Josh Allen because he stuck out like a sore thumb in ownership.
Here is a look at the highest value plays at other positions:
It starts to get more nuanced here on how I select players, but the clearest plays here are the ones that have a green box for ownership. I try to see what stacks reveal themselves (Panthers, Seahawks, Cardinals) and go from there.=Boyd, miss 5.3 PPR points
Derrick Henry, miss 7.7 PPR points, didn’t play him at all and I don’t have a good reason why. Dodged a bullet
Aaron Jones, meh, didn’t play him much because of price
Jerry Jeudy, 7 receptions and 13.30 points through a quarter and a half before getting injured:(
Tyler Lockett, 26 PPR points
Robby Anderson, 12.7 PPR points
Corey Davis, 26.7 PPR points
DJ Moore, 15.4 PPR points
Justin Jefferson, 12.5 PPR points
O'Dell Beckham Jr. - not happy (late scratch)
DK Metcalf, 16 PPR points
Jaylen Waddle, 16.1 PPR Points - I should have listened.
Michael Pittman, 14.6 PPR points
Laviska Shenault, 12.9 PPR points
Chase Edmonds, 14.6 PPR points
Using this method help me hit on a lot of low-owned stacks that did well( Panthers, Seahawks, Cardinals). Because a lot of the chalk failed (Pitts, Callaway, Mostert, Kamara, Elijah Moore), lower scores got it done in tournaments this week opening up a lot of opportunities for contrarian approaches like mine. My lineups did so well because I made some key leverage plays to make sure I benefitted the most when the field’s favorite plays were wrong. A few of those key pivots were:
Corey Davis > Elijah Moore
Chase Edmonds > Mostert at a similar price point
Prioritizing the high-priced receivers with middling ownership: Hopkins, Tyreek, AJ Brown
Laviska and DJ Chark > Marvin Jones. Had some sick teams with Chark on them.
You don’t have to get everything correct, but what you do get correct, you want it to be in spots where a lot of other people are wrong at the same time. This limits the number of things we need to get right to gain ground on the field. So, does stacking 2-3 players from the same team and running it back with one player from the other team, which I did a ton.
One final note before I show you guys some of the better lineups I landed on. Kenneth Gainwell had a projected 0% ownership and a 22-point ceiling. The only thing holding him back was his “value” projection because his baseline projection was so low. I ignored that when I should not have. These are the plays that win you tournaments. I think I may try to calculate a “ceiling” value and prioritize that in tournaments so I don’t miss on players like Gainwell.
Here is a quick breakdown of my single entry tournament lineup, which I put the most thought into and coincidentally was my best lineup in terms of placement. (174/681 in the single entry NFL Redzone contest)
The Cards stack with Edmonds and a Rondale Moore fade at high ownership was good process. The J.D. Mckissic play was cute, especially at 5100. I had a tough time with the late news, but Moster probably would have had a good game had he stayed healthy. Falcon’s D in 2000 was a lazy play. In a field this small, I didn’t have to get too crazy with the low-owned plays, but a large proportion of my plays are under 12%. I think I would lower this number to about 8 or 9% if I was trying to take down a bigger tournament.
I think I was one or two good decisions away from going from a “good” week to a “great week”. Either that or I got lucky and this model sucks. I’ll break down my player pool for everyone this weekend as these projections get updated and we get news about usage and injuries throughout the week.