The following article is a post from guest CelticsHub contributor Romy Nehme. Romy’s work can also be found at 2 girls 1 ball
KG continued his milestone one-upping ways in the Celtics gory drudge-fest against the Pacers, a win that was as much keyed by Garnett’s defense as it was by his ability to scan every player’s minute movement on the court and thread the defense with two pin-point accurate passes to Bradley and Green in the last 90 seconds of the game.
Something about Wednesday night’s go-ahead basket got me thinking about the anatomy of an offensive play: Pierce’s sticky back pick was the one that initially freed up Jeff Green on his curling dash to the basket, but there was more to that KG pass than what met the eye:
I saw Jeff Green open at the basket, and then I saw help come across the baseline. I thought I had a little bit extra on it, but he was open off the initial [move]. I just had to make sure he took the corner and I hit him with the pass. The first time I did it in the first quarter, I noticed, obviously Hibbert is a very long guy. … I noted to myself to wait until Jeff turned the corner and the guy’s hands go down. Second time, I just did that. I pump-faked and his hands went up, I saw his adjustment … and threw it right over the top of his hands to Jeff. Jeff made a tough, difficult layup for the game.
And what about Jeff Green making the acrobatic catch, gathering himself and laying it in on the other side? For all of the debates about qualitative vs. quantitative measurement, basketball remains as much craftsmanship as it is artistry.
Now, much ink has been spilled over stats geekery and the difficulty in properly weighting the impact of any one player’s defense and how it correlates to a team’s ability to keep opponents starved for baskets. Brendan recently wrote about the fresh insights we can extract from a new set of interior defense analytics — using SportVu data as a complement to more traditional metrics.
But are the evaluation models that spit out numbers quantifying offensive impact really that much more developed than those available on defense?
A CASE FOR A MULTI-TOUCH ATTRIBUTION ASSIST MODEL IN THE NBA
A multi what, you say?? Basketball is not entirely different from the world of online marketing. For years, marketers have been scratching their heads wondering how they could possibly quantify all of their expensive marketing efforts (email marketing, social marketing, paid search, etc.) in a way that would commensurately assign credit for a consumer’s purchase — and allocate future advertising monies — to those touchpoints.
Say you received an email from Delta about cheap flights to Boston, liked the promotion on their FB fan page (bear with me here), tossed the offer onto your mental backburner and eventually went on to purchase the discounted flight by googling ”Delta cheap flight to Boston” a day later. The old model would only take into consideration that last action, i.e. the Google search preceding the purchase — a clearly inadequate view of the events incrementally nudging you all the way to the purchase.
In basketball, that’s the equivalent of giving the assist to the player who makes the pass right before the basket. Which is what we currently do according to the NBA’s statistics manual as long as the “player scoring the goal responds by demonstrating immediate reaction to the basket”.
What’s so wrong about a system that’s entirely weighted in favor of the last pass?
It turns out the answer is “a lot”, actually. Mainly, it’s an overly simplistic view that fails to capture the multiple actions in a single possession and some of the crucial interactions that precede that last, and for now, only assist that we keep track of. Let’s look at a few other things we might want to know about how a play unfolds:
1) The quality of the opportunity created by the assist: once the last pass is made, how hard was it for the player to finish off the play? Further complicating this subtlety is the fact that assists are subject to the whims of humans no more perfect than Marc Davis. Chilling thought. This is something we came across as Rondo’s assist streak was being alternately threatened and aided by any number of scorekeepers. HoopData recently published a study gauging the “lift” that teams benefit from just by virtue of playing at home. It was found that the difference in assist percentage from 2006-2012 between playing at home vs. playing away was on average a 4.9% increase in recorded assists, courtesy of generous scorekeepers. That means that some “assists” are not only of the Hail Mary variety but also of the dubious kind where the basket is less about the pass and more about the scorer’s creativity.
Two plays from the Golden State game that illustrate the stark contrast in the quality of KG’s “assists”:
Someone was clearly eating his Legal Sea Foods shrimp platter and not paying attention to the game in the latter instance.
2) Assists negated by fouls: going by #1, not all assists are created equal. As a matter of fact, a lot of the higher quality assists that lead to bunny shots within 5ft are often never realized because of fouls.
3) The hockey assist: Tom Haberstroh wrote about how the Heat are one of the teams that track hockey assists, only later discovering that they tend to be more fruitful than regular assists. How so? Of the 26 hockey assists the Heat had amassed up until that point, 16 of them led to 3pt shots (an average of 2.6 points) whereas a regular assist only yielded 2.2 points.
4) The all-important yet mostly unheralded screener: Van Gundy had the following to say a few weeks ago as part of his daily digest of on-air tirades, an idea that Brett Koremenos further explored at HoopSpeak:
A screen is just a different form of an assist. We keep track of passes that lead to an assist, but screens are the same thing. It’s giving yourself up to free a teammate or get a teammate a shot.
Good point. Here it is in action:
Either way, what we end up with when utilizing a more “linear multi-touch attribution system” (copyright that mouthful!) is a much less myopic, hence more instructive and deconstructive view of the offense. One that takes into account who initiates the offense, how players get open looks, and how sometimes, one of the intermediary passes or parallel off-the-ball plays can be just as important as the dazzling one at the end that unfairly gets all the credit.
If you still don’t believe me that an assist is in the eye of the beholder, take a look at the variability in the following four charts resulting from different modelling approaches (taken from an online marketing campaign). The takeaway: a different pair of glasses can really Shallow Hal-skew the landscape.
Earlier this season, I wrote about how aging players now laugh in the face of Father Time’s prescriptions of temperance given recent technological and medical advances. And that in some extreme cases, they’re not just slowing down the effects of the inverted U curve of NBA productivity, but flat out reversing them.
A hypothesis I wanted to explore was that, over the last 16 games, more touches for the sprightly KG would produce more situations where he would be involved in high lows, post ups, cuts, and ultimately fertile assist situations.
The numbers bear this out:
KG Pre-Rondo Injury (43 games) KG Post-Rondo Injury (16 games)
Usage rate 24.5% 26.4%
Assist percentage 13.2% 16.8%
Essentially, in the first 43 games, KG tallied 91 assists, or 2.11 assists/game, whereas he’s notched 44 in the last 16 games (tied for 3rd on the team), i.e. 2.75 assists/game. The main beneficiaries of KG’s playmaking ability have been Pierce (8 assists), Jet (7 assists), Green (7 assists) and Bass (7 assists).
Below is the KG to Pierce assist chart. It turns out that many of KG’s assist distribution patterns resemble the same paint-by-number, as 38% assists to those four players are in that oh so sweet spot hugging the rim. Just the type of quality assists that we saw KG coolly deliver to Bradley and Green in Indiana.
The Celtics are also averaging 0.5 assists more per game over that stretch, despite the team’s assist percentage dropping from 62.8% to 62% (the increased pace and possessions make up for this).
UP FOR FURTHER EXPLORATION …
Posts ups and cuts — two areas that are greatly facilitated by KG — just so happen to be the most fruitful plays for the Celtics (according to Synergy). The Celtics average 0.88 points per possession for the former, good for 5th in the league, while averaging 1.24 PPP for the latter, which puts them 6th in the rankings. For now, these are little used in the Celtics’ offensive system (Boston uses post ups 9.8% of the time and cuts only constitute 6.5% of the offense).
Since Rondo’s injury, however, the Celtics have only gone to the post up 8.55% of the time in spite of the fact that PP and KG are elite when it comes to squeezing out points when operating out of the post.
I may still be partially brainwashed by all the Cartesian coordinate systems we encountered at the MIT conference last weekend, but I think it would be interesting to start tabulating and weighting a larger subset of “assists”. Not because it would produce subjective measurements that angry drunkards could debate over a lager — like whether a 2-pt comeback win on the road was a bigger triumph than the 19-pt trouncing the Celtics dealt the Pacers at home earlier this year — but because the next wave of analytics is most likely going to enlighten us on matters of synergies between complementary skill sets and playing styles: things we know precious little about.
Perhaps that’s a bone we could throw to the stats community as they brainstorm what research papers to pitch to the gatekeepers of next year’s Sloan conference. In the meantime, the Celtics should explore running more of their offense through KG, who age be damned, is the gift that keeps on giving.
Romy Nehme is a Canadian hoops junkie who grew up worshipping the Boston Celtics and is a regular guest contributor for CelticsHub. Romy can also be found at 2 girls 1 ball, which is not nearly as salacious a site as the name might imply. — Ed.]