APBRmetrics

[davis21wylie2121]

Hey, folks, it's pretty late, but I figured I'd post what I've been working on lately...

Team projections using ORtg/%Poss/DRtg and skill curves. It's pretty much like what I did here, except the projections are done using past performance only (+ an age adjustment -- think of it as our version of Tom Tango's Marcels). Once you've cranked out projections, the procedure for each team is basically the same as it was in the Celtics trial run linked above. Plug in playing time for every player, reconcile player-aggregate min to team min (team paces/minute projections were created using a best-fit regression on 3-year team data from 1974-2006... except Houston, whose pace I upped b/c I figured Adelman would run more), adjust ORtg upward/downward based on how many extra possessions each member of the team would have to use, and project team pts scored/allowed. After that, team efficiencies were scaled to fit the 2005-07 distribution using z-scores. Easy, huh?

Here are the projected standings:

[davis21wylie2121]

And the players...

[davis21wylie2121]

And the most valuable players, according to WARP:

[davis21wylie2121]

The best offenses:

[Mountain]

I may be forgetting some discussion from past but... on the skill curve part do you assume all players follow the the the slope of the next year of the general skill curve for the coming year (age or experience?) or do you in any way blend the shape of the individual's skill curve performance to date with the general skill curve? I have wondered the same thing with respect to John Hollinger's projections. I know each player is going to have their own level of performance that gets adjusted by the skill curve. If the shape of the individual's actual development is different in first 3-5 years than general skill curve how likely is it that the rest of their career will vary and in similar manner and amount? Have players been sorted into subgroups with sufficient similarity to say that instead of one skill curve there are a few and if a player looks like he is in one sub-group the probability is pretty good that he will be the rest of the way? Or does early development curve variance tend to diminish and support use of one general skill curve for next year performance projections?


Bulls as #1 in east on the stats makes sense to me. Skiles seems like a top level coach but also strikes me as a guy who could cost his team some wins under some circumstances. So while on paper I agree the Bulls may project to the most wins, it wouldn't suprise me if they slipped some on that and maybe Detroit passes them (or another but I could more immediately see Detroit outperforming). Central division title in particular is worth something but not sure how possessed by it each coach and team will be.

[supersub15]

Nice breakdown. Thanks.

Only concern is that the Eastern Conference has only five team above .500. How could that be?

[Mountain]

Were expected wins based off a generic schedule or a team's actual schedule which would capture east-west and other differences?

[davis21wylie2121]

Thanks for the quick feedback, guys!

Re: Mountain's 1st reply - I think you're confusing "skill curve" with "age adjustment". The first refers to how much a player's ORtg will be affected (good or bad) by a change in %Poss; for this, I basically used a sliding adjustment based on their initial projected %Poss (players who have high initial usages will experience small changes in ORtg with a usage change, while low-usage players will see huge swings if they're asked to take on a bigger/smaller role in the offense). The age adjustment was based on general trends, though, like the Marcels. It's a rough sketch, but I think it's cool that such a system actually passes the laugh test.

Re: Supersub, Mountain's 2nd reply - The W-L is based solely on pyth, with no regard to schedule. At some point in the future, I might take these pW%'s as a starting point and simulate every game of the season using log5, which would fix the East/West issues and generally make for more realistic projections. Although, I do have to note that the East had only 5 teams above .500 last year (2 were at .500), so even these results aren't all that far-fetched.

[deepak_e]

This is pretty cool.

How did you do the projections for rookies (e.g. Scola)? And do you have a listing of WARP for all the players?

[Mountain]

Right wherever I said "skill curve" substitute "age adjustment".

With that change, the commentary about ideally changing the age adjustment for the next season to reflect something more than general age adjustment and capturing at least part of the individual variance from that pattern in prior career years still stands for your projections and John Hollinger's if he doesn't do it already by identifying and using the next step in the development pattern of "similar players".

I'd also be interested to know if John's projections were based on real team schedules.

And if you went to the max this could mean accounting for home, road, back to back.

[davis21wylie2121]

For rookies, we'll just say it was an inexact science.

In a nutshell, I gathered all of the "NBA Comparisons" I could find for this year's rookies (from sites like NBADraft.net, etc.), and made a best guess based on how those players fared at the same age and in their first few seasons. We'll have to see how it works out as the season goes on.

As for the complete list, I'm going to upload the whole file to the Yahoo group, so I'll post the link when it's up.

[deepak_e]

I really like the BoP stats for describing past performance. But, I wonder, how effective is it really for doing forecasts?

Has there been any type of study on various methods for projecting standings at the beginning of a season? Has it been shown anywhere that statistical-based forecasts of wins is more accurate than the guesses of "experts"?

[davis21wylie2121]

I'm not sure if Dan and David covered that topic in their NESSIS presentation, but I think Dave Berri has commented on the subject (he found that stats beat expert predictions in a variety of fields). As for me, though, this is kind of an experiment in whether BoP stats can predict future team performance better than other methods. In fact, one reason I've sided squarely in the BoP camp is because I think Dean's metrics have a much, much, much better chance of predicting the performance of any given 5-man unit (or team) than PER/Wins Produced/etc. So I guess this is my way of putting my money where my mouth is...

[davis21wylie2121]

[John Beattie]

Remind me - what's BoP - is it Basketball on Paper