Building a machine learning model to predict the NBA MVP

Building a machine learning model to predict the NBA MVP(perthirtysix.com)

12 points by robmoore 5 years ago | 10 comments

mlex 5 years ago |

This is a great article and a great read, even as someone who knows very little about machine learning models but is a fan of the NBA.

One question I have is whether it’s possible to model voter fatigue. There was a point brought up that LeBron had a solid candidacy in the Rose MVP season but voters didn’t want to give it to him again, and I feel like this equally applies to Giannis this year. Though Giannis’ season isn’t as good as Jokic’s for sure, if they were closer I feel like Giannis would have to do WAY better than the competition to have a shot at the award, given general sentiment about him and his playoff results.

aparsons 5 years ago | |

I have not watched NBA basketball in years, but the NBA had the most vague definition of MVP going back decades.

What it certainly is not - since the 1980s or longer - is “best player”. Magic over Jordan in the late 90s set that precedent, followed by Karl Malone winning a few as a good but not great player.

This continued in the 2000s, with Shaquille O’Neal not winning as many as one might predict. I suspect that is still the case.

mlex 5 years ago | | |

Right, it’s cliché but it’s clear that MVP isn’t the “best player” award given how infrequently LeBron has won it. It’s more some nebulous combination of “how good you are”, “how much you contribute to your team’s success”, and “are you doing so without the support of another star”.

djitz 5 years ago | | |

I stopped reading at Karl Malone supposedly being a good but not great player.

ccmonnett 5 years ago |

Really enjoyed the article, and followed on twitter. Good luck with the launch.

Follow-up question: Do you see a correlation between your predictions and actual finishing order increasing over time? I ask because of some very weak seasons (Nique '86 and MJ '87) for 2nd-place finishers in the 80s. It made me wonder if the MVP is becoming more 'objective' now that the media is able to follow teams around the league more easily than before.

robmoore 5 years ago |

XGBoost with LambdaMART ranking performed the best across all models(neural nets, log reg also tested).

Adding pairwise ranking with LambdaMART helped to solve overfitting with only some 360 data points. Instead of looking at binary win/loss MVP, I looked at the ranking of all players receiving votes in each of the past 38 years.

sceadu 5 years ago | |

Would you be open to publishing the code for your model?