Similar Player Algorithms?

miabbett · Post by **miabbett** » Tue Jul 02, 2013 4:09 am

I was looking through Ken Pomeroy's site and noticed that for each player, he lists 5 past players that put up comparable statistics. He briefly mentioned the metrics used in his calculations, but does anyone know how exactly he formulates these comparisons? Also, are his player comparison methods better than those of Basketball-Reference shown here? http://www.basketball-reference.com/about/similar.html
My gut feeling is yes as these just use WS and don't factor in direct comparisons of each specific metric. Any thoughts or opinions would be greatly appreciated!

Crow · Post by **Crow** » Tue Jul 02, 2013 2:57 pm

Don't know what Ken Pom does.

I am not a fan of Basketball-Reference's similarity approach. I am not that interested in career shape and certainly not over style of play. I want a similarity rating based on current total impact and style of play.

If defensive winshares were better I might be more interested; but given how crude a measure of overall defensive impact it is, I am not. It mainly relies on the boxscore and then assumes for simplicity that everyone on the team contributes exactly the same to shot defense, all shot defense of all minutes whether on the court or not for particular good or bad results. Guys who get steals and blocks get big credit for these acts but do not pay any individualized price if other times these aggressive acts lead to easy points against.

Statman · Post by **Statman** » Tue Jul 02, 2013 7:19 pm

Crow wrote:Don't know what Ken Pom does.

I am not a fan of Basketball-Reference's similarity approach. I am not that interested in career shape and certainly not over style of play. I want a similarity rating based on current total impact and style of play.

If defensive winshares were better I might be more interested; but given how crude a measure of overall defensive impact it is, I am not. It mainly relies on the boxscore and then assumes for simplicity that everyone on the team contributes exactly the same to shot defense, all shot defense of all minutes whether on the court or not for particular good or bad results. Guys who get steals and blocks get big credit for these acts but do not pay any individualized price if other times these aggressive acts lead to easy points against.

Agreed.

I think similarity scores NEED to incorporate similar players based on their statistical profile & age. Overall impact similarity scores ignoring the differences when the stats are broken down to their minutiae doesn't tell me much.

As for Pomeroy - looked at couple examples - didn't like the results. Russ Smith's comps were, well, unimpressive. Russ Smith was Pomeroy's #1 rated player last season - his comps are: '07 Jerel McNeal (836), '12 Dundrecous Nelson (798), '10 Kim English (794), '07 Mike Mercer (789), '08 James Florence (781). So, his #1 ranked player IN THE NATION last season is compared to 5 guys - only one of which was maybe the best player on his own mediocrish team - and NONE of the 5 has their BEST season as a comp.

I know, the "similarity" is lowish for Russ' comps (sign of his "uniqueness") - BUT the impact of the player HAS to be incorporated better somehow into the comps. SoS and the player's impact in terms of his own team's quality seem to pretty much be ignored here. When the THIRD best season of a past player from Mercer is a top 5 comp for your #1 rated player in the nation - something needs to be reworked.

BTW - I LOVE Pomeroy's team stuff - but when he delves into the player stuff with his rankings or similarity scores, things get screwy. Case in point - Pomeroy's All SEC team before their conference tourney:

"March 13-16 at Nashville, Tenn.

All-kenpom: Patric Young, Florida (kPOY); Erik Murphy, Florida; Kenny Boynton, Florida; Mike Rosario, Florida; Scottie Wilbekin, Florida."

Yes, ALL 5 are Gators.

My Top 16 SEC (to get to Rosario of Florida):

1 Kentavious Caldwell-Pope Georgia
2 Erik Murphy Florida
3 Jarnell Stokes Tennessee
4 Alex Oriakhi Missouri
5 Trevor Releford Alabama
6 Murphy Holloway Mississippi
7 Nerlens Noel Kentucky
8 Patric Young Florida
9 Laurence Bowers Missouri
10 Casey Prather Florida
11 Marshall Henderson Mississippi
12 Marshawn Powell Arkansas
13 Scottie Wilbekin Florida
14 Kenny Boynton Florida
15 Reginald Buckner Mississippi
16 Mike Rosario Florida

So you know KCP was an EASY #1 by HnR - ranked 7th in the nation with a 176 rating - Murphy was 33rd in the nation with a 156 rating. In fact, the difference in rating between #1 & #2 in the SEC by HnR was slightly BIGGER than the difference between #2 and #14. Yet, no KCP in Pomeroy's top 5 - just all Gators.

BTW, Noel would have been #2 if the games missed due to injury were ignored - in case anyone wonders (still a distant 2nd to KCP).

So, somehow Pomeroy seems to way overvalue team strength when he attempts to do ratings at the player level (big reason why he has Russ Smith #1 in the nation - Russ was #12 in the nation by my calculations), which I believe he is fully aware and is why I think he doesn't show more than his top 10 nationally. However, his similarity scores seem to greatly ignore team strength. It's just odd.

But, I subscribe to Pomeroy ($20 a year I think) - I love many of the analytics he throws at us on the team pages. If you are a real college bball fan - it's worth the 20 bucks. Just know, his player rankings (kPOY top 10) and his player similarity scores aren't the reason to go there - but almost everything else is interesting imo.

miabbett · Post by **miabbett** » Tue Jul 02, 2013 7:36 pm

Awesome, I completely agree with both of you guys. I haven't come across a comparison system that accurately incorporates style of play. Has anyone found any better comparison methods for players or even 5 man lineups?

Crow · Post by **Crow** » Tue Jul 02, 2013 11:33 pm

APBRmetrics

Similar Player Algorithms?

Similar Player Algorithms?

Re: Similar Player Algorithms?

Re: Similar Player Algorithms?

Re: Similar Player Algorithms?

Re: Similar Player Algorithms?