I am sick and tired of losing at fantasy football every year. So this year, after losing in my regular season semifinal game, I took the decision-making power out of my hands and let my computer pick my team for the NFL playoffs. I’m hoping this objective approach will mean that I’ll fair better than the dismal 76th-out-of-131 that I managed last season.
So real-quick… how it works:
12 slots. 2 QBs, 3 RBs, 3 WRs, 2 TEs, 1 Defense, 1 Kicker
All 12 must be from a unique team.
At the bottom of this newsletter you’ll find my predicted fantasy playoff points for the top 100 skill-position players, plus their actual playoff fantasy points. But first, how we got there.
The Data
I got the data from nlfreadr, an R package. It included a player’s fantasy points (PPR and non-PPR) every week of the season, both regular and postseason. It also had info like the player’s position, team, how many passing yards/receiving yards, etc.
I decided to just use fantasy-related metrics because that’s what I’m trying to predict anyway.
In the end, the player-data I included looked like this:
Total fantasy points for the season
Average weekly points
Standard deviation of weekly points
Games played
Worst week points and best week points
25th, 50th, and 75th percentile of points for the season
I also included some data for each player about their team, because the model needs some inkling of how deep into the playoffs they will get. I didn’t have a lot of time to figure out each team’s seed and playoff win probabilities, so I settled for a few other data points that I hope are informative:
Wins
Average point differential
Standard deviation of point differential
Average points for
Average points against
Win rate
The Model
I took all this and put it into a model, where I trained on the regular season from 2010-2021 and predicted on that year’s postseason, and then measured the results.
I also only trained the model on players whose teams actually made it to the playoffs, since I will know this year which players I can choose from. There’s no point in feeding the model all the extra players who didn’t make it and will, as a result, score 0 points in the playoffs. (If I didn’t know who was going to make the playoffs at this point, then I would have to include all players in the training)
Results
The results were far from perfect. My predictions were off by an average of 11.5 points. But that’s still probably better than I could do picking based on my gut-feel, so I’ll take it!
Let’s take a look at who the model liked for this year’s playoffs.
As expected, a lot of the top players are QBs. However, what I’m also interested in is the righthand column where we see that a few players are just that much better than the next-best player at their position. Unsurprisingly, #1 in that category is Travis Kelce. Austin Ekeler is another one that comes up with a 10-point advantage over his next-closest counterpart, Saquon Barkley.
With that info, I decided to swap in Jalen Hurts for Patrick Mahomes, sacrificing a projected 3 points in order to gain a projected 15 points by having Kelce instead of Evan Engram, his replacement.
My team ended up looking like this:
I took a bit of a risk with Hurts since the Eagles had a bye, but I did that for the benefit of getting Kelce and the incremental points he brings. All the top submissions last year had both Allen and Mahomes, and they got the bulk of their points in the wildcard and divisional weekends.
Performance
So how did I do? Much better than last year. I’m currently 22nd out of 158 total, so clearly in the top 20% of all players. However, at this point, with only Jalen Hurts and Travis Kelce left on my team, I can’t improve much more in my final ranking from the Super Bowl because almost every team above me also has those two players.
The best lineup so far was only slightly different than mine in some key areas… mainly Jamar Chase over Justin Jefferson, and Christian Kirk over Tyreek Hill. Despite this, the top player still didn’t have an optimal entry. That would’ve required Burrow and Prescott at QB, Deebo Samuel at WR, and Dalton Schultz at TE2. Of course some of this could change depending on how Mahomes and others play in the Super Bowl.
Here are the top 100 players ranked by their predicted playoff fantasy points. On the far right you’ll see their actual points so far and the difference to the model predictions. Blue are over-acheivers and red are underacheivers.
Overall, the model performed pretty well, and was off by only 5.8 points on average in 2022—even better than we expected! Below is a visual of the predictions and actual points, with some key over- and underperformers highlighted.
This was a fun experiment and I think my computer certainly helped me to field a competitive team, but there’s always that element of randomness that leads to a guy like Christian Kirk outperforming Tyreek Hill and Justin Jefferson, or Brock Purdy outperforming Josh Allen. I think the model could be improved next year by giving it some idea of how many games a team is likely to play in the playoffs, including any potential bye week, and by factoring in a player’s previous-year playoff points.
Thanks for reading and let me know if you liked this content by replying to the email. Enjoy the Super Bowl and go birds!
— Kyle