Twitter | Search | |
Kostya Medvedovsky
Antitrust lawyer at . Tweets about NBA stats sometimes. Not a Hipster, despite coining term "Hipster Antitrust".
6,545
Tweets
893
Following
922
Followers
Tweets
Kostya Medvedovsky Aug 17
Replying to @bbstats
I think about this like twice a day, trying to figure out how not to make this mistake. It seems incredibly hard to avoid in practice.
Reply Retweet Like
Kostya Medvedovsky Aug 15
Replying to @bbstats @ElGee35
Oh - didn't see #5 here. Thread opened weird.
Reply Retweet Like
Kostya Medvedovsky Aug 15
Replying to @bbstats @ElGee35
And vice versa, offensive boards happen when *your* team misses a shot, so you're selecting for "bad offense." Best win to win the offensive boards is to give the keys to your offense to like Dion Waiters.
Reply Retweet Like
Kostya Medvedovsky Aug 15
Replying to @bbstats @ElGee35
I'm not sure #2 is why defensive boards are more predictive of player value. Pretty sure the main reason is actually because defensive boards happen when the opposing team misses a shot, so you're selecting for "good defense."
Reply Retweet Like
Kostya Medvedovsky Aug 15
That's the theory, yes. That there's a strong selection bias to the initial drafting.
Reply Retweet Like
Kostya Medvedovsky Aug 15
's PIPM has splits, so possible to check there. Or once finish the rest of my HW, I can check using DRE or something.
Reply Retweet Like
Kostya Medvedovsky Aug 15
Haha. You're like a teacher giving me homework Seth!
Reply Retweet Like
Kostya Medvedovsky Aug 15
The negative intercept in particular is what is driving at here; I had been only doing regression to the mean, so a -1 player who changed teams would get a *bump*. With a negative intercept, that may no longer be the case, and the -1 player may likewise decline.
Reply Retweet Like
Kostya Medvedovsky Aug 15
I think the way to test this would be to have a separate coefficient and intercept for team changes. Under this theory, you'd get some regression to the mean effect (e.g., regression 20% to the mean), and then you'd just also have a negative intercept.
Reply Retweet Like
Kostya Medvedovsky Aug 15
That's right. once modeled this as an additional regression to the mean factor for players who change teams, so my RPM projections have always included dummies for team and coaching changes. 's point is that there may just be a *decline* effect as well
Reply Retweet Like
Kostya Medvedovsky Aug 15
Replying to @bbstats @DSMok1 and 5 others
Your SPM should be much better fit than mine given some of the work you've been doing there.
Reply Retweet Like
Kostya Medvedovsky Aug 15
Difference is the specific metrics in which they performed well/poorly in. Here's Ingram, cause he's always fun too. Big spikes!
Reply Retweet Like
Kostya Medvedovsky Aug 15
And lest it look like the point of the model is that "nothing matters", here's Trae Young. He started likewise with a terrible swoon (in orange), but ended with a spike that caused him to end the year at roughly a +2 projection (from a dip at -2 midseason).
Reply Retweet Like
Kostya Medvedovsky Aug 15
One more observation on Sexton...while he didn't get a huge bump during the tail-end of his season, he also didn't have a huge collapse in the first half when he was playing poorly. It depends on which specific stats he was performing well/poorly in.
Reply Retweet Like
Kostya Medvedovsky Aug 15
Yup. Want to take them for a spin, and no better way than to release.
Reply Retweet Like
Kostya Medvedovsky Aug 15
But I'm mostly going to leave the PIPM/SPM modeling to others like Kevin, and . Focus on my end is on the underlying box-score metrics projections, updating daily.
Reply Retweet Like
Kostya Medvedovsky Aug 15
I haven't done much modeling on the underlying SPM, so I'm using 's DRE here . I'm going to try and publish on a Google Sheet somewhere daily updating box-score projections for each player, which should be usable to create more SPMs (including something like PIPM).
Reply Retweet Like
Kostya Medvedovsky Aug 15
DaysAgo in the abstract should be better, as it contains more information, and has the benefit of adding an additional regression to the mean factor during the offseason (everything is decayed by roughly 200 days).
Reply Retweet Like
Kostya Medvedovsky Aug 15
So an established player who had the same season as Sexton would have a different projection, since their prior performance would have been in the model as well, although decayed by X^DaysAgo.
Reply Retweet Like
Kostya Medvedovsky Aug 15
There's a lot more to add here, including salient for Sexton, a dummy for rookie status or age, though they're both implicitly in the the model, since the X^DaysAgo framework initializes each component to a certain prior, and the projection is updated accordingly.
Reply Retweet Like