I am in a fantasy football league that has an interesting set up. This league started with 32 teams at the beginning of the year and each week the two lowest scoring teams are cut from the league and their players are put back into the free agent pool.

My question is how you could potentially model the probability of getting cut. Each week a team starts with projected points. Based on the projected points at the beginning of the week, can we estimate this probability?

I built a simple logistic regression model and it does okay, but one thing I think a good model should do is consider that only two teams will be cut. If two teams are projected to score 20 points and the rest of the league is projected for 120 points then the model should consider that most likely those two teams will be cut and everyone else should be relatively safe. It also should consider the number of teams. With 32 teams the probability should of being cut should be lower for most teams and when there are only 8 teams then it should be much higher for every team.


1 Answer 1


Personally, I think making a model to compute cut probability for each team, then using the two highest probabilities to predict who will be cut is your best bet. I have a few ideas, in no particular order, so I'll list them here:

  • I agree that contextualizing other teams is important. You may want to think of this problem as a classification problem, where your input vector includes features about the team being predicted upon, as well as features describing the other teams
  • Having a model be able to support variable numbers of things is hard. So I would either summarize all features for all other teams, or train different models for different numbers of teams. From a feature reduction perspective, the former seems more practical.
  • How exactly you set up your features, I think, will be your most difficult challenge. You can probably use almost any classification model, but how you choose to create your feature space (weather you're including temporal qualities, what features you're using, etc.) is your biggest issue. You can find resources to guide you in feature engineering, but, really, it will probably come down to a lot of "aha" moments and guess and check.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.