Counting the Baskets

3-D Slashers, Take One

 

Like most of you, I have been thinking a lot about shot selection in the NBA recently.  

More and more three point shots are taken in the NBA each year as teams realize the value of that extra point and players entering the league get better and better at hitting those shots.  Guards who can hit these shots are currently being called 3 and D guys, even the ones that don't particularly play D.  Bigs or Wings that can hit threes are called Stretches.

However, if a player can get really really really close to the basket the percentage of shots that go in goes way up, more than compensating for the extra point on the long ball.  So, other guys are called Slashers, because of their ability to 'slash' to the basket.  Then there a big guys that never leave the basket area and couldn't find the 'elbow' with a map and a compass.

There aren't any really good definitions of these categories, so doing analysis of their effectiveness is difficult.  Luckily, thanks to play-by-play data we have counts of where every player shoots from on the court and how good they are at making shots from each location.  I grabbed the shot locations of every player in the NBA last year off of HoopData.

Then I did a couple of Cluster analyses on Bigs, Wings, and Guards based on their shot locations (without regard to actually hitting the shots).   Cluster analysis basically helps you group items or people together in categories based on their statistical similarity.

The analyses gave me back some fairly pleasing results in supporting the conventional wisdom and my prior beliefs, which is always nice.  Of the locations listed by HoopData, three point shots and shots at the rim were by far the most distinctive and important in grouping players in all positions. Mid-range shots are spread pretty randomly, except for Bigs where the ability to hit anything outside of arms length gets a separate category.

For Guards the shot location clusters look like this, all of the percentages of shot attempts:

	COUNT of Players	AVERAGE of Rim Shots	AVERAGE 3-9 Shots	AVERAGE 10-15 Shots	AVERAGE 16-23 Shots	AVERAGE Three Shots
	304	36.44%	13.41%	7.52%	23.18%	18.57%
Slashing Guard	78	33.18%	10.74%	8.46%	24.98%	23.00%
Three Point Specialist	90	18.61%	6.11%	5.03%	20.19%	49.00%
Grand Total	459	32.50%	11.58%	7.20%	22.90%	25.11%

 

The Three point specialists took nearly half their shots from three point land, while the slashers got to the rim for 33% of their shots.

Neat right? But wait there's more!

Next I joined the Clusters to my player projection data to see how Slashing Guards and Three Point Specialists compare as groups in actual effectiveness.  I filtered out the players playing less than 150 minutes last year or projected for this year.    The  preliminary results were pretty interesting.  

Values
Row Labels	Count	Ave TS%	Avg eFG%	Avg ORB%	Avg DRB%	Avg AST%	Avg TOV%
Slashing Guard	64	51.3%	46.8%	2.69	9.94	23.25	15.76
Three Point Specialist	74	53.9%	50.6%	2.05	9.54	17.97	13.49
Grand Total	134	52.7%	48.8%	2.35	9.72	20.42	14.54

 

The three point shooters score more efficiently, with a better effective field goal percentage and a better slightly true shooting percentage.  But they're lacking in other areas like offensive rebounding because they're too far from the basket most of the time.

 

	Values
Row Labels	Count	Ave TS%	Avg Proj xRAPM	Avg Wghtd 2 Yr ASPM	Avg WS/48	Avg WP 3 yrs
Slashing Guard	64	51.3%	-0.51	-0.41	0.081	0.080
Three Point Specialist	74	53.9%	-1.09	-0.67	0.089	0.086
Grand Total	134	52.7%	-0.82	-0.55	0.085	0.083

 

As before, I am using Arturo Galleti's projection of Wins Produced, Nathan Walker's xRAPM projections, and a two year weighted average of ASPM and Win Shares with an aging factor.  The advanced metrics actually liked the Slashers and Three Point Specialists about the same, with the Plus/Minus based metrics a little in favor of the Slashers.

Since I lumped the Points Guards and Shooting Guards together, it's fair to ask whether the cluster just renamed the traditonal position.  The answer is not really, though there is a bias for Point Guards to be categorized as Slashers with Shooting Guards tilting a bit less heavily to Three Point specialists. See below:

 

	SG	PG
Slashing Guard	33%	67%
Three Point Specialist	57%	43%

 

Some of the same patterns came out between Slashers and Three Point Shooters even when only comparing Point Guards or Shooting guards.  The Three Pointer shooters had higher eFG's and slightly higher true shooting percentages while the Slashers rebounded better, particularly on offense.

New Team	(All)
Pos	PG
AvgOfMP Projection	(Multiple Items)
	Values
Row Labels	Count	Ave TS%	Avg eFG%	Avg ORB%	Avg DRB%	Avg AST%	Avg TOV%
Slashing Guard	43	51.1%	46.5%	2.39	9.50	27.48	17.33
Three Point Specialist	32	53.1%	49.6%	1.91	8.53	24.12	16.05
Grand Total	75	52.0%	47.8%	2.19	9.09	26.05	16.79

 

New Team	(All)
Pos	SG
AvgOfMP Projection	(Multiple Items)
	Values
Row Labels	Count	Ave TS%	Avg eFG%	Avg ORB%	Avg DRB%	Avg AST%	Avg TOV%
Slashing Guard	21	51.6%	47.3%	3.31	10.84	14.59	12.54
Three Point Specialist	42	54.5%	51.3%	2.15	10.30	13.28	11.54
Grand Total	63	53.5%	49.9%	2.54	10.48	13.72	11.87

 

I will follow this up next time with the wings.

 

Great, and Not-So-Great, Expectations

Instead of projecting the Celtics record, which I plan to do, I thought I would start by evaluating the Celtics roster overall.  Projecting wins means projecting minutes, which is particularly tough on a team with turnover and a new coach.

The approach I took was to compare how the advanced metrics view the C's roster compared to a replacement player in the NBA. This gives a partial idea of how good the players look over the entire roster and gives a range of their expected value. Take a look at the brief explanation of the metrics used  in the extended Grantland style footnote**, or go here for Evan Zamir's primer.

I used 'replacement player' rather than 'average player' because when I looked at the stats I found that most players have below average production. That's because the best players on most teams get more minutes, so the production numbers are skewed upward. Replacement player also gives one an idea of the value you could expect with a vet minimum type player. The level I chose is also at about the rating of players in the 25th percentile in the NBA.

All that said, any player with less than 4 wins over replacement has below average production.

So, below is each of the metrics ratings comparing playing that player playing starter level minutes, 36 minutes for every game, to a replacement level player win terms of added or subtracted wins and losses.

 

Name	xRAPM Wins Compared to Replacement Player per 36	ASPM Wins Compared to Replacement Player per 36	Win Shares Compared to Replacement per 36	Wins Produced Compared to Replacement per 36	EZPM Wins Compared to Replacement player per 36	Average Wins Over Replmnt Player	Range
Avery Bradley	2.7	0.6	2.6	0.2	-2.8	0.7	5.5
Brandon Bass	0.8	2.7	5.7	3.8	-4.2	1.8	9.8
Courtney Lee	1.9	3.7	3.8	5.2	3.8	3.7	3.3
Gerald Wallace	8.0	6.7	4.1	6.6	5.4	6.2	3.9
Jared Sullinger	4.4	5.2	9.0	7.2	6.0	6.4	4.7
Jeff Green	2.1	3.1	4.9	2.0	6.4	3.7	4.4
Jordan Crawford	-0.2	-1.0	1.8	-2.0	-7.0	-1.7	8.8
Keith Bogans	-1.5	0.0	-0.2	3.8	-4.8	-0.6	8.6
Kelly Olynyk	3.6	4.0	2.2	1.6	4.0	3.1	2.4
Kris Humphries	3.6	0.4	5.8	10.5	2.8	4.6	10.1
Marshon Brooks	-1.9	0.4	2.2	2.6	6.0	1.8	7.9
Phil Pressey	-0.3	-0.3	-1.0	-1.5	0.0	-0.6	1.5
Rajon Rondo	11.0	9.0	5.8	12.6	10.5	9.8	6.8
Vitor Faverani	1.1	1.1	1.3	2.1	1.1	1.4	0.9

 

A couple things come out here:

• Avery Bradley- The metrics don't like his game too much, in part because he didn't play very well last year on offense. xRAPM, which as a plus/minus stat, is probably the best at accounting for defense, likes him the best.
• Gerald Wallace- All the metrics like him. I think fans are under valuing him because he is over payed and his contract is too long. Over payed does not equal can't play. If he bounces back, he could be valuable to a contender at the deadline.
• Jared Sullinger- The metrics think he earned his starting minutes. He could be the most fun player to watch this year if the back is indeed healthy.
• Jeff Green doesn't do great or too bad. The metrics don't know which Green will show up any more than anyone else does.
• Crawford and Bogans- None of the metrics like them particularly, and they're rated below replacement level overall.
• Kris Humphries is the most controversial according to the metrics with a range of 10 wins. Win's Produced, which is a rebound heavy metric, likes Humphries a lot.
• Rajon Rondo- Not surprisingly, Rondo is the key to the season. Coach and locker room drama aside, Rondo's production is the key to being competitive for this team at all with a 10 win swing from Rondo to a replacement level point guard, which about what the team has as options.
• Faverani, Pressey, and Olynyk are based on a combo of the Wages of Wins projections based on their Euro or NCAA numbers and my selection of rookie 'comps' for other metrics.  In other words, semi-educated guesses.

So, that's it there're talent on the team, but at this point one star and some potential.

** xRAPM- This is an advanced adjusted plus minus stat. The version here comes from Nathan Walker at Basketball Distribution. He used three years of data and an aging factor.

ASPM- This a box score based stat developed by Daniel Meyers. I like this one, in part because it did the best in an independent study in terms of both consistency and predicting, though only narrowly. It's also pretty intuitive. I used two years of ASPM ratings weighted by minutes and giving more weight to last year than the prior year. I also added an aging factor.

Win Shares- Box score stat developed by Justin Kubatko based Dean Oliver's 'four factors'. I used two years of ratings and an aging factor.

Wins Produced- Box score stat developed by Dave Berri that loves rebounds. I used a projection by Arturo Galleti at Wages of Wins, using three years of ratings weighted by minutes and an aging factor.

EZPM- Evan Zamir's Play by Play data based stat. This one is sort of complicated, but it's the only one using play by play data. I only had one year of data and applied an aging adjustment.