APBRmetrics

So I was wondering if anyone knows how the NBA Player projections are calculated in 538's RAPTOR system
(I believe the website owner(?) Nate Silver is sometimes commenting on ARBRmetrics so perhaps he can give some info)

https://projects.fivethirtyeight.com/20 ... ojections/

My understanding is that for a player projection the system looks at a number of similar players from the past. Then the top (say 10) matching players are combined to generate the projection.

My questions are:

1) How are matches made. i.e. what similarity measure are you using?

2) Once you have found the top 10 matching candidates how do you combine them to make the prediction? Simple weighted averaging or are you using something more advanced?

I am working on something similar and I would like to share some ideas.

It isn't laid out in extreme detail but there's a little more detail on this page https://fivethirtyeight.com/features/ho ... ric-works/ . Skim down to the PREDATOR/Projections section.

Hi, thanks for that.
I think the gist of 538's projection method is this:

1) Create a regression-based baseline projection of a player’s future RAPTOR ratings for the next seven seasons using his PREDATOR ratings from the past three seasons, plus his age and the other biographical variables

2) Identify similar or comparable players using a number of statistical and biographical categories

3) Tweak the player’s projection, and develop a range of uncertainty around the forecast, by seeing how the comparable players performed relative to their baselines when they were the same age as the player is now

So, Regress-Match-Tweak

I am actually using a slightly different approach: Match - Regress - Tweak
which makes the projections by first matching to other players and using the other player's future data to regress.

The reason being that I noticed that regressing from a single player his own future projections is very unreliable.
For example in step 1) above using 3 seasons to extrapolate 7 seems very poor in terms of performance. And it is.