2024 Fantasy Football Predictive Model

My output on this blog has been a little smaller than I was hoping it would be a year ago, but today I am back with version 1 of my fantasy football predictive model.

While this blog is coming out just after the season has started, I thankfully managed to get predictions for each position generated right before my two fantasy drafts, and that minimal research has resulted in a 5-3 start to the season for myself over 2 leagues (not that anyone cares about your fantasy teams).

The Model

I recently spent a session of my dedicated project time organizing the working directory, and I plan on having the folder uploaded to my GitHub shortly. The code for the visualizations I include will be uploaded to my blog repository as well for anyone interested on taking a deeper dive or offering any constructive feedback.

As of version one, the model is admittedly very simple. I used player stats data from nflFastR, and summarised to the mean and variation of each statistic for each player each season, with lags dating back the previous 3 years. In short, the model predicts next seasons total fantasy from each players statistics in the 3 previous seasons. I trained the model separately for each of the 4 skill positions (QB, RB, WR, TE), using xgboost cross-validation to determine the ideal parameters and to insure the xgboost model was working as expected.

The Predictions

Quarterbacks (draftjoshalleninfantasy.com)

Being the first model I ran I was glad to see the results passed the eye test... sort of. The top five match what most would expect. Dak at 3 may come as a surprise to some and while he likely could fall a bit short of that rank, this prediction is the little evidence I needed to believe that Dak's ADP of 9 among QB's (according to sleeper.com) is perfect value for those waiting on a QB.

The shortcomings of the QB model are very obvious sadly. From first look it seems that it has a tough time handling players who missed significant time the previous year. Aaron Rodgers falls on this list below players who have only seen action as backup QBs. While I did hope that players injury history would find its way into the model, it does not do so in a way that satisfies my expectations.

For example, it discounts Lamar significantly more for only playing 12 games in 2021 and 2022, despite playing 2023 entirely healthy and finishing as QB3 for weeks 1-17. Conversely, Kirk Cousins is projected one spot above Lamar, despite missing 8 games last year to an injury that creates more doubt in his future performance.

I tried training the model on only quarterbacks who met a certain games played threshhold, but diminishing the already relatively small sample size created more issues. I hope to improve this model (along with the other position models) in the future. Potentially forecasting the specific statistics that contribute to scoring individually, even perhaps on a week by week basis, could make the model more precise.

Additionally, the model seems to under predict the expected fantasy points (it does this for the other positions as well, but QB's are the worst offender). The QB1 in fantasy tends to finish anywhere in the high 300 to low 400 range, and the model has QB1 right at 300. This is less concerning of an issue, as a result of 400, while closer to what would be expected of a high end fantasy QB, is not seen very often in the training data. Hopefully, adjusting the predictor set and potentially predicting for individual statistics as mentioned before can help this. Another potential solution I hope to try out is a bayesian model, that would allow me to showcase the median outcome as well as high and low percentile outcomes. Adding depth chart data and team usage data to the predictor set should also help the model identify starting QB's and QB's on new teams and grade them accordingly.

Running Backs

While still underpredicting by a little bit, I am more confident in the output of all the non-QB predictions. With consenus top running backs, Christian McCaffrey and Breece Hall among others, at the top of the predictions I feel confident the model is working well enough.

The biggest outlier on this chart that pops out compared to consensus big boards is Alvin Kamara as RB2. He is definitlely of more value in PPR leagues (the format I am predicting for) with his high amount of targets, but is fifteen spots higher on my predictions than his Sleeper ADP among RB's (17). Devon Achane is also much higher here than his consensus here, RB3 compared to a sleeper ADP of RB10. If your like me and don't like taking running backs too high, particularly in the first round, these guys offer great value in the more middle RB range. James Conner, my models RB9 also fits in this category, as his ADP on sleeper is RB19.

Update: I am not a very quick writer. The season has already started as I am finishing up this report. And at this moment in time it does seem that Alvin Kamara is much closer to RB2 than RB17.

Wide Receivers

With the wide reciever predictions, we yet again see plenty of big names and top fantasy receivers at the top of the list here. With regards to the predictions, nothing stands out to me too much in terms of suprises in the rankings. However to me it does seem that the model is missing lots of information and I would like to examine what I will need to change to improve it.

Mainly, it seems that the model is having a tough time handling wide receivers who have new teams or new QB's. The model does not include this information, and its clear in how it has handled WR's like Stefon Diggs or Keenan Allen. Both these wide receivers are aging veterans on new teams with established young receivers, it would seem reasonable to expect their target share and opportunities to decrease. Not that I think they shouldn't be in the top 25 here, but it is something to note. Additionally, one would expect Garret Wilson to improve upon last year and place higher on this list, with new-ish QB Aaron Rodgers returning from injury (DISCLAIMER: no model can account for Rodgers inexplicable love of Allen Lazard).

Ideally, including an expected target share and/or yardage share metric would be the solution I would like to implement for the WR model.

Tight Ends

The tight end model is victim to the same flaws; TJ Hockenson is going to be missing most of the year with a knee injury from December.

In general, I've personally never been a big tight end guy when drafting my teams. My best seasons at the position have come when getting LaPorta or Andrews in the mid-late rounds back when they were big upside players.

Having Kelce and Kittle at the top here is definitely a good sign, although knowing what we know after 3 weeks into the season, I'm not sure the draft capital they demand is worth it for such a volatile position.

Conclusion

With plenty of time on my hands over the next year I hope I can return with a model that is much more sound, while the model might have a reasonable error, it is clear that there are issues, mainly stemming from the use of predictors. With more time to create a specefic set of predictors I feel confident that the model will improve. Currently, I'm working on a way to determine value of players in dynasty football taking into account future value. Hopefully that is successful and I'll update in a post on here.

And hopefully that will be in less than a year.