Tick-Tock Shot Clock: Optimal Stopping in the NBA - UC San Diego ...
Tick-Tock Shot Clock: Optimal Stopping in the NBA - UC San Diego ...
Tick-Tock Shot Clock: Optimal Stopping in the NBA - UC San Diego ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
<strong>Tick</strong>-<strong>Tock</strong> <strong>Shot</strong> <strong>Clock</strong>: <strong>Optimal</strong> <strong>Stopp<strong>in</strong>g</strong> <strong>in</strong> <strong>the</strong> <strong>NBA</strong> ∗<br />
Matt Goldman 1 Just<strong>in</strong> M. Rao 2<br />
1 Department of Economics, University of California, <strong>San</strong> <strong>Diego</strong><br />
2 Yahoo! Research, <strong>San</strong>ta Clara, CA<br />
Fall 2011<br />
Abstract<br />
The paper exam<strong>in</strong>es <strong>the</strong> optimality of <strong>the</strong> shoot<strong>in</strong>g decisions of National Basketball<br />
Association (<strong>NBA</strong>) players us<strong>in</strong>g a rich dataset of 1.4 million offensive possessions. The<br />
decision to shoot is a complex problem that <strong>in</strong>volves weigh<strong>in</strong>g <strong>the</strong> cont<strong>in</strong>uation value of<br />
<strong>the</strong> possession and <strong>the</strong> outside option of a teammate shoot<strong>in</strong>g. We model this as a dynamic<br />
mixed-strategy equilibrium. At each second of <strong>the</strong> shot clock, dynamic efficiency<br />
requires that marg<strong>in</strong>al shot value exceeds <strong>the</strong> cont<strong>in</strong>uation value of <strong>the</strong> possession. Allocative<br />
efficiency is <strong>the</strong> additional requirement that at that “moment”, each player <strong>in</strong><br />
<strong>the</strong> l<strong>in</strong>e-up has equal marg<strong>in</strong>al efficiency. To apply our abstract model to <strong>the</strong> data we<br />
make assumptions about <strong>the</strong> distribution of potential shots. We first assume noth<strong>in</strong>g<br />
about <strong>the</strong> opportunity distribution and establish a strict necessary condition for optimality,<br />
which nearly all players/teams pass. Add<strong>in</strong>g distributional assumptions, we<br />
establish sufficient conditions for optimality. In l<strong>in</strong>e with dynamic efficiency, we f<strong>in</strong>d<br />
that <strong>the</strong> “cut threshold” decl<strong>in</strong>es monotonically with time rema<strong>in</strong><strong>in</strong>g on <strong>the</strong> shot clock<br />
at approximately <strong>the</strong> correct rate. Most l<strong>in</strong>e-ups show strong adherence to allocative<br />
efficiency. We l<strong>in</strong>k departures <strong>in</strong> optimality to l<strong>in</strong>e-up experience, player salary and<br />
overall ability.<br />
∗ We would like to thank James Andreoni, David Eil, Pedro Rey Biel and Charlie Sprenger for extensive<br />
notes. Nageeb Ali, Gordon Dahl, Uri Gneezy, Craig McKenzie, David Miller, William Peterman, Joel Sobel<br />
and Joel Watson provided helpful comments. This paper is a merger of Goldman’s “Optimiz<strong>in</strong>g <strong>in</strong> <strong>the</strong><br />
Half-Court” and ideas presented <strong>in</strong> Rao’s “He Got Game Theory: <strong>Optimal</strong> Decision Mak<strong>in</strong>g and <strong>the</strong> <strong>NBA</strong>”.<br />
1
1 Introduction<br />
Basketball is hard. Indeed one has to be tall, athletic and coord<strong>in</strong>ated. In this paper we<br />
focus on ano<strong>the</strong>r way basketball is hard. It is a strategic <strong>in</strong>teraction, <strong>in</strong>volv<strong>in</strong>g a complex<br />
optimization problem, <strong>in</strong> which good decision-mak<strong>in</strong>g is paramount. Professional basketball<br />
is played <strong>in</strong> a series of 24-second (or less) “possessions.” In each possession <strong>the</strong> offensive<br />
team has to maximize <strong>the</strong> number of po<strong>in</strong>ts scored (save end-of-game situations). In order<br />
to optimize, players must effectively employ two workhorse concepts from game and decision<br />
<strong>the</strong>ory. First, s<strong>in</strong>ce a shot has to be taken <strong>in</strong> <strong>the</strong> 24-second time limit, <strong>the</strong> team has to solve<br />
an optimal stopp<strong>in</strong>g problem, which <strong>in</strong>volves sett<strong>in</strong>g thresholds, based on <strong>the</strong> time rema<strong>in</strong><strong>in</strong>g<br />
on <strong>the</strong> shot clock, that determ<strong>in</strong>e if a shot opportunity is realized or passed up <strong>in</strong> favor of<br />
cont<strong>in</strong>u<strong>in</strong>g <strong>the</strong> possession. Second, <strong>the</strong>se thresholds ought to be uniform across players. Or<br />
<strong>in</strong> o<strong>the</strong>r words, <strong>the</strong> team must share <strong>the</strong> ball so that marg<strong>in</strong>al productivity at each moment<br />
is equal across players.<br />
Optimiz<strong>in</strong>g <strong>in</strong> basketball thus requires adher<strong>in</strong>g to optimality conditions that are literally<br />
“textbook.” Solv<strong>in</strong>g a stopp<strong>in</strong>g problem entails both accurately estimat<strong>in</strong>g a cont<strong>in</strong>uation<br />
value and compar<strong>in</strong>g it to <strong>the</strong> currently observed outcome, conditional on do<strong>in</strong>g <strong>the</strong> same<br />
procedure <strong>in</strong> <strong>the</strong> next period. This type of reason<strong>in</strong>g underlies <strong>the</strong> equilibrium concepts of<br />
nearly all repeated games. <strong>Optimal</strong> allocation requires distribut<strong>in</strong>g a scare resource (shots)<br />
to many factors of production with differ<strong>in</strong>g technology (players) <strong>in</strong> order to maximize output<br />
(po<strong>in</strong>ts). Shar<strong>in</strong>g <strong>the</strong> ball is basic producer <strong>the</strong>ory.<br />
For <strong>the</strong>se reasons, we th<strong>in</strong>k basketball is an <strong>in</strong>tensely <strong>in</strong>terest<strong>in</strong>g test<strong>in</strong>g ground of fa-<br />
miliar standards of optimality. Moreover, optimiz<strong>in</strong>g requires <strong>the</strong> use of reason<strong>in</strong>g concepts<br />
that laboratory subjects are notoriously bad at employ<strong>in</strong>g, mean<strong>in</strong>g we may have doubts<br />
<strong>in</strong> people’s ability to solve this type of decision problem <strong>in</strong> <strong>the</strong> way predicted by <strong>the</strong>ory. A<br />
fur<strong>the</strong>r layer of complexity is that appropriately analyz<strong>in</strong>g optimal play <strong>in</strong> basketball also<br />
presents <strong>in</strong>terest<strong>in</strong>g challenges. <strong>NBA</strong> players exhibit <strong>in</strong>credible heterogeneity <strong>in</strong> both <strong>the</strong><br />
ability to score and <strong>the</strong> chosen technique to accomplish this goal. In order credibly address<br />
shot selection, one must first flexibly model <strong>the</strong> shot opportunity distribution of each player.<br />
We exam<strong>in</strong>e <strong>NBA</strong> player decision-mak<strong>in</strong>g us<strong>in</strong>g rich data set compris<strong>in</strong>g all shots taken<br />
(over 1.3 million possessions) <strong>in</strong> <strong>the</strong> <strong>NBA</strong> from 2006–2010. Our model<strong>in</strong>g approach is to<br />
embed allocation with<strong>in</strong> <strong>the</strong> stopp<strong>in</strong>g problem. At each po<strong>in</strong>t on <strong>the</strong> shot clock, <strong>the</strong> team<br />
must randomize over potential shooters <strong>in</strong> order to atta<strong>in</strong> equal marg<strong>in</strong>al productivity across<br />
<strong>the</strong> 5-man l<strong>in</strong>eup currently on <strong>the</strong> court. We call this allocative efficiency. 1 At each po<strong>in</strong>t<br />
1 The vectors of shoot<strong>in</strong>g frequency should maximize po<strong>in</strong>ts per possession. One might th<strong>in</strong>k <strong>the</strong> players<br />
are not consciously randomiz<strong>in</strong>g, but <strong>in</strong>stead respond<strong>in</strong>g to slight changes <strong>in</strong> <strong>the</strong> game. In this case, each<br />
2
<strong>in</strong> <strong>the</strong> shot clock, <strong>the</strong> potential shooter must decide if he should realize <strong>the</strong> current shoot<strong>in</strong>g<br />
opportunity, or pass up <strong>the</strong> shot. This requires sett<strong>in</strong>g <strong>the</strong> appropriate threshold value for<br />
cont<strong>in</strong>u<strong>in</strong>g <strong>the</strong> possession versus realiz<strong>in</strong>g a shot opportunity. If <strong>the</strong> shot opportunity is<br />
perceived to exceed this value, <strong>the</strong> shot is taken. We call this dynamic efficiency.<br />
Naturally it is challeng<strong>in</strong>g to fit a complex game such as basketball <strong>in</strong>to a tractable<br />
model, but we show that we can establish necessary and sufficient conditions for optimality<br />
with m<strong>in</strong>imal assumptions. Like all stopp<strong>in</strong>g problems <strong>in</strong> <strong>the</strong> field (see Rust (1984) for a<br />
nice discussion), we have to compute <strong>the</strong> “marg<strong>in</strong>al counterfactual;” based on <strong>the</strong> shots we<br />
actually observe a player tak<strong>in</strong>g, we have to estimate <strong>the</strong> shot opportunity distribution he<br />
faces, so we can evaluate if he appears to be pass<strong>in</strong>g up shots that he should take (to test<br />
dynamic efficiency) and if he should shoot more or less (to test allocative efficiency). Our<br />
identification scheme, which we describe <strong>in</strong> detail <strong>in</strong> Section 3, simultaneously illum<strong>in</strong>ates a<br />
player’s ability, by model<strong>in</strong>g <strong>the</strong> shot distribution which <strong>the</strong>y draw shoot<strong>in</strong>g opportunities<br />
from, and <strong>the</strong>ir chosen rule for realiz<strong>in</strong>g an opportunities at each po<strong>in</strong>t of <strong>the</strong> shot clock,<br />
depend<strong>in</strong>g on l<strong>in</strong>e-up.<br />
Our key f<strong>in</strong>d<strong>in</strong>g is that <strong>NBA</strong> players are superb optimizers. The average <strong>NBA</strong> player is<br />
shown to adopt reservation shot values almost exactly equal to <strong>the</strong> cont<strong>in</strong>uation value of his<br />
team’s possession throughout <strong>the</strong> entire range of <strong>the</strong> shot clock. This “cut threshold” decl<strong>in</strong>es<br />
monotonically with time rema<strong>in</strong><strong>in</strong>g on <strong>the</strong> shot clock at almost <strong>the</strong> exact rate implied by<br />
dynamic efficiency. Most teams core l<strong>in</strong>e-ups show impressive allocative efficiency, with <strong>the</strong><br />
spread of marg<strong>in</strong>al efficiencies with<strong>in</strong> a l<strong>in</strong>e-up quite small. Very few players can be shown<br />
to <strong>in</strong>dividually overshoot from <strong>the</strong> dynamic (shoot too soon) or allocative (shoot too much<br />
overall) perspective. Undershoot<strong>in</strong>g is more common and seems to occur primarily <strong>in</strong> players<br />
easily recognized as amongst <strong>the</strong> <strong>NBA</strong> elite. 2<br />
The f<strong>in</strong>d<strong>in</strong>g that <strong>NBA</strong> players can reliably solve an optimal stopp<strong>in</strong>g problem is <strong>the</strong> first<br />
field (or lab, for that matter) evidence of its k<strong>in</strong>d that we are aware of, with <strong>the</strong> excep-<br />
tion of large firm-level decisions such as patent renewal (Pakes, 1986) and tree harvest<strong>in</strong>g<br />
(Provencher, 1997). The evidence on optimal stopp<strong>in</strong>g <strong>in</strong> <strong>the</strong> field, especially with regards to<br />
<strong>in</strong>dividual decision mak<strong>in</strong>g, is th<strong>in</strong> because <strong>the</strong> requirements for reliable estimation are stiff.<br />
One must correctly specify preferences with respect to risk and time, and possess rich enough<br />
data to <strong>in</strong>fer <strong>the</strong> opportunities an agent has observed, but did not execute. For example,<br />
to understand if job hunters term<strong>in</strong>ate employment search us<strong>in</strong>g <strong>the</strong>oretically implied wage<br />
vector implies a “threshold equilibrium” of play, <strong>in</strong> <strong>the</strong> spirit of a purified game (Harsanyi, 1973).<br />
2 Such behavior is suboptimal <strong>in</strong> our model, but is easily rationalized if we allow for such players to conserve<br />
<strong>the</strong>ir energy and health for <strong>the</strong> long-haul of <strong>the</strong> season at <strong>the</strong> slight expense of <strong>the</strong>ir team’s immediate<br />
performance.<br />
3
thresholds, <strong>the</strong> econometrician would need f<strong>in</strong>ancial balances, risk/time preferences, and a<br />
record of all offers. Such data are hard to come by given current collection technology and<br />
privacy legislation. In contrast, we show that risk and time neutrality can be established<br />
for <strong>the</strong> majority of basketball game-play (with <strong>the</strong> rema<strong>in</strong>der elim<strong>in</strong>ated) and we can model<br />
unrealized opportunities based on <strong>the</strong> detailed data we have on taken shots. We believe <strong>the</strong><br />
study of basketball, while <strong>in</strong>terest<strong>in</strong>g to some readers <strong>in</strong> its own right, can also lend <strong>in</strong>sight<br />
to <strong>the</strong> human capacity to solve this sort of problem more generally.<br />
Our analysis also reveals an additional twist of optimal play. In our ma<strong>in</strong> estimation,<br />
we only use possessions <strong>in</strong> which <strong>the</strong> team ought to focus only on po<strong>in</strong>ts, <strong>in</strong> a risk-neutral<br />
fashion. In contrast, towards <strong>the</strong> end of games, <strong>the</strong> trail<strong>in</strong>g team does better by lower<strong>in</strong>g<br />
<strong>the</strong> threshold function, so that <strong>the</strong>y take a worse shots on average, but save time by do<strong>in</strong>g<br />
so. The trail<strong>in</strong>g team also should <strong>in</strong>duce variance, by tak<strong>in</strong>g more 3-po<strong>in</strong>ters. Conversely,<br />
<strong>the</strong> lead<strong>in</strong>g time ought to raise <strong>the</strong> cut threshold, pass<strong>in</strong>g up good shots <strong>in</strong> order to use time<br />
and reduce variance by tak<strong>in</strong>g more 2-po<strong>in</strong>ters. We f<strong>in</strong>d that <strong>NBA</strong> players recognize and<br />
respond to this trade-off. The cut-thresholds significantly change <strong>in</strong> <strong>the</strong> correct direction<br />
when time becomes <strong>in</strong>tr<strong>in</strong>sically valuable and <strong>the</strong> trail<strong>in</strong>g team reliably <strong>in</strong>duces variance<br />
by tak<strong>in</strong>g more 3-po<strong>in</strong>ters. There is evidence, however, that <strong>the</strong> cut-threshold shift is too<br />
conservative, especially for <strong>the</strong> trail<strong>in</strong>g team.<br />
Our f<strong>in</strong>al piece of analysis seeks to understand which factors of a l<strong>in</strong>eup are associated<br />
with optimal play. We l<strong>in</strong>k <strong>the</strong> sufficient statistics from our tests of optimality to l<strong>in</strong>e-up<br />
experience (how often <strong>the</strong> players share <strong>the</strong> court), player tenure <strong>in</strong> <strong>the</strong> league, player salary<br />
and salary <strong>in</strong>equality us<strong>in</strong>g OLS regression. We f<strong>in</strong>d that adherence to allocative efficiency<br />
significantly positively related to l<strong>in</strong>e-up output and salary level. Increased variance <strong>in</strong><br />
l<strong>in</strong>e-up salary leads to more optimal play. On <strong>the</strong> <strong>in</strong>dividual level, star players are more<br />
likely to undershoot, perhaps to conserve energy/health of <strong>the</strong> course of <strong>the</strong> season. For<br />
both measures, player tenure <strong>in</strong> <strong>the</strong> league does not appear to have a mean<strong>in</strong>gful impact<br />
controll<strong>in</strong>g for <strong>the</strong>se o<strong>the</strong>r state measures.<br />
The rema<strong>in</strong>der of <strong>the</strong> paper proceeds as follows. In <strong>the</strong> follow<strong>in</strong>g two subsections we<br />
review <strong>the</strong> related literature and provide a basketball primer, <strong>in</strong> order to <strong>in</strong>troduce <strong>the</strong> key<br />
game elements necessary to understand our model of basketball decision mak<strong>in</strong>g. Section<br />
2 describes <strong>the</strong> data, Section 3 presents <strong>the</strong> model <strong>in</strong> detail and discusses our identify<strong>in</strong>g<br />
assumptions <strong>in</strong> detail, <strong>the</strong> key results on optimal play are presented <strong>in</strong> Section 4, additional<br />
co-variate analysis and a discussion follow <strong>in</strong> Section 5 and Section 6 concludes.<br />
4
1.1 Related Work<br />
Our work is most closely related to two different literature. First, s<strong>in</strong>ce optimal stopp<strong>in</strong>g<br />
<strong>in</strong>volves <strong>the</strong> important game <strong>the</strong>oretic concept cont<strong>in</strong>uation value, <strong>the</strong>re is a deep literature<br />
of laboratory experiments <strong>in</strong>vestigat<strong>in</strong>g people’s ability to effectively employ such reason<strong>in</strong>g. 3<br />
Early work <strong>in</strong> this literature used <strong>the</strong> “classical secretary problem” (CSP). In <strong>the</strong> CSP, a<br />
subject observes a series of “applicants” and w<strong>in</strong>s a prize if she “hires” <strong>the</strong> best one. She<br />
can only hire <strong>the</strong> current applicant and is given only rank order <strong>in</strong>formation on <strong>the</strong> applicant<br />
(how does <strong>the</strong> applicant compare to those who have come before). A robust f<strong>in</strong>d<strong>in</strong>g is that<br />
subjects “under-search” by hir<strong>in</strong>g a candidate to quickly (Kahan et al., 1967; Rapoport and<br />
Tversky, 1970). Recent work that has relaxed many assumptions of <strong>the</strong> CSP has also found<br />
early stopp<strong>in</strong>g (Seale and Rapoport, 2000; Zwick et al., 2003; Bearden et al., 2005; Bearden<br />
et al., 2006). In <strong>the</strong> full <strong>in</strong>formation CSP payments are simply <strong>the</strong> chosen draw (like a taken<br />
shot). Just like basketball, <strong>the</strong> solution requires a monotonically decl<strong>in</strong><strong>in</strong>g cut threshold<br />
with draws rema<strong>in</strong><strong>in</strong>g; yet <strong>the</strong> majority of subjects use flat thresholds (Lee et al., 2004),<br />
even when <strong>the</strong> time horizon is short (Lee, 2006). 4 Overall, across a wide variety of stopp<strong>in</strong>g<br />
problems, lab subjects typically stop <strong>the</strong> process early and use fixed, <strong>in</strong>stead of decl<strong>in</strong><strong>in</strong>g<br />
thresholds. In stark contrast, our results <strong>in</strong>dicate <strong>NBA</strong> players do not stop <strong>the</strong> early and<br />
use a decl<strong>in</strong><strong>in</strong>g threshold <strong>in</strong> l<strong>in</strong>e with <strong>the</strong> <strong>the</strong>oretical optimum.<br />
The second closely related literature exam<strong>in</strong>es <strong>the</strong> predictive power of game <strong>the</strong>oretic<br />
equilibrium concepts <strong>in</strong> professional sports. 5 Mixed-strategy Nash Equilibrium (MSNE) has<br />
been tested <strong>in</strong> <strong>the</strong> field us<strong>in</strong>g 2×2 simultaneous move games analogous to match<strong>in</strong>g pennies.<br />
In soccer penalty kicks (Chiappori et al., 2002; Palacios-Huerta, 2003) and tennis serves<br />
(Walker and Wooders, 2001; Hsu et al., 2007), players typically randomize across strategies<br />
and over time consistent with m<strong>in</strong>imax play. MSNE is exam<strong>in</strong>ed <strong>in</strong> baseball and American<br />
football <strong>in</strong> Kovash and Levitt (2009), with <strong>the</strong> authors assert<strong>in</strong>g that baseball pitchers do not<br />
3 Past work us<strong>in</strong>g experts has studied firm decisions <strong>in</strong> dynamic sett<strong>in</strong>gs such as as harvest<strong>in</strong>g tree stands<br />
(Provencher, 1997), renew<strong>in</strong>g patents (Pakes, 1986), replac<strong>in</strong>g bus eng<strong>in</strong>es (Rust, 1987) and cont<strong>in</strong>u<strong>in</strong>g drives<br />
<strong>in</strong> professional football (Romer, 2006). These problems are similar to <strong>the</strong> stylized lab version of <strong>the</strong> problem<br />
<strong>in</strong> one respect — <strong>the</strong> decision is based on <strong>the</strong> comparison of expected value of act<strong>in</strong>g versus cont<strong>in</strong>u<strong>in</strong>g —<br />
but are dissimilar <strong>in</strong> that <strong>the</strong>y do not <strong>in</strong>volve random arrivals of opportunities (for <strong>in</strong>stance, one can replace<br />
a bus eng<strong>in</strong>e at any time). <strong>NBA</strong> shoot<strong>in</strong>g presents a situation that is strik<strong>in</strong>gly similar to <strong>the</strong> classic version<br />
of <strong>the</strong> stopp<strong>in</strong>g problem.<br />
4 With a very long time horizon, a high constant threshold is not optimal, but does pretty well s<strong>in</strong>ce with<br />
high likelihood one will get a good draw with this strategy. The short horizon games are a better analog to<br />
basketball.<br />
5 There is of course also a deep literature of test<strong>in</strong>g game <strong>the</strong>ory <strong>in</strong> laboratories. Three recent papers<br />
present a nice discussion about <strong>the</strong> <strong>in</strong>terplay between lab and field <strong>in</strong> MSNE games, and <strong>the</strong> reader is<br />
po<strong>in</strong>ted <strong>the</strong>re for a deeper dive (Palacios-Huerta and Volij, 2009; Levitt et al., 2010; Palacios-Huerta and<br />
Volij, 2008).<br />
5
andomize over pitch selection properly and football coaches do not pass enough. However,<br />
<strong>the</strong> analysis is hampered by <strong>the</strong> difficulty com<strong>in</strong>g up with <strong>the</strong> correct dependent variables <strong>in</strong><br />
<strong>the</strong>se sports. 6<br />
The only work we are aware of that applies an optimal stopp<strong>in</strong>g model to professional<br />
sports is Romer (2006), which exam<strong>in</strong>es football coaches’ decisions to play versus punt on 4th<br />
down on <strong>the</strong> NFL. Romer f<strong>in</strong>ds that coaches are too conservative on average—<strong>the</strong>y punt too<br />
often. The key difference between this study and ours is that Romer studies a deliberative<br />
decision made about 10 times per season by <strong>the</strong> coach, not <strong>the</strong> players. The low number<br />
of occurrences is <strong>in</strong> contrast to <strong>the</strong> roughly 8,000 possessions played each year by an <strong>NBA</strong><br />
team. These differences, as well as o<strong>the</strong>r obvious features of <strong>the</strong> decision problem <strong>in</strong> <strong>the</strong><br />
two sett<strong>in</strong>gs, could expla<strong>in</strong> <strong>the</strong> lack of consonance <strong>in</strong> <strong>the</strong> f<strong>in</strong>d<strong>in</strong>gs of his paper and ours.<br />
Empirical work on firm-level stopp<strong>in</strong>g decisions has typically found that studied firms have<br />
dynamically managed <strong>the</strong> asset <strong>in</strong> question <strong>in</strong> close accordance to <strong>the</strong>oretical predictions<br />
(Pakes, 1986; Provencher, 1997; Rust, 1987).<br />
2 A Basketball Primer and Key Strategic Elements<br />
In this section we give an overview of <strong>the</strong> strategic elements <strong>in</strong> <strong>the</strong> game of basketball as it<br />
is played professionally <strong>the</strong> <strong>NBA</strong>. The purpose is to prevent <strong>the</strong> relevant background before<br />
jump<strong>in</strong>g <strong>in</strong>to <strong>the</strong> nitty gritty of <strong>the</strong> model.<br />
2.1 Basketball basics<br />
In <strong>NBA</strong> play, two twelve-man teams face each o<strong>the</strong>r <strong>in</strong> a 48-m<strong>in</strong>ute game. Each team has<br />
five players on <strong>the</strong> court. A five player comb<strong>in</strong>ation is referred to as a “l<strong>in</strong>eup” and <strong>the</strong><br />
l<strong>in</strong>eup that typically plays <strong>the</strong> most consists of <strong>the</strong> players who beg<strong>in</strong> <strong>the</strong> game for <strong>the</strong> team,<br />
<strong>the</strong> “starters.” The team who scores <strong>the</strong> most po<strong>in</strong>ts <strong>in</strong> <strong>the</strong> game w<strong>in</strong>s. Po<strong>in</strong>ts are scored by<br />
shoot<strong>in</strong>g <strong>the</strong> ball <strong>in</strong> <strong>the</strong> basket. Long-range shots earn three po<strong>in</strong>ts, normal shots earn two<br />
po<strong>in</strong>ts and shots taken after a foul (free throws) earn one po<strong>in</strong>t each (a player is awarded<br />
6 For baseball, <strong>the</strong> authors use a metric of batt<strong>in</strong>g performance (slugg<strong>in</strong>g percentage + on base percentage)<br />
that does not map directly to <strong>the</strong> probability of w<strong>in</strong>n<strong>in</strong>g (conditional on game conditions). Indeed any bias <strong>in</strong><br />
<strong>the</strong> metric as it maps to w<strong>in</strong>n<strong>in</strong>g percentage biases <strong>the</strong> results, and <strong>the</strong>re is good reason to believe such bias<br />
exists. The metric gives equal weight to slugg<strong>in</strong>g percentage and on base percentage, but on base percentage<br />
is a better predictor of runs scored. S<strong>in</strong>ce this measure overvalues “big hits” and undervalues walks, it is not<br />
surpris<strong>in</strong>g <strong>the</strong> authors f<strong>in</strong>d that pitchers use too many pitches that limit walks but risk big hits (fastballs)<br />
and not enough that limit big hits but risk walks (curve balls). For football, <strong>the</strong> game is not as clearly a<br />
simultaneous move game, because <strong>the</strong> offense can change <strong>the</strong> play (“audible”) after see<strong>in</strong>g <strong>the</strong> players <strong>the</strong><br />
defense has on <strong>the</strong> field and it is easier (and more common) to audible to a run.<br />
6
<strong>the</strong> number of free throws correspond<strong>in</strong>g to <strong>the</strong> value of <strong>the</strong> shot he was fouled on). In our<br />
sample period, teams scored on average 92.5 po<strong>in</strong>ts per game.<br />
2.2 The central limit <strong>the</strong>orem for basketball<br />
Although basketball game-play is fast-paced and (potentially) cont<strong>in</strong>uous, 7 it is effectively<br />
broken up <strong>in</strong>to a series of non-overlapp<strong>in</strong>g “offensive possessions.” A possession starts when<br />
one teams gets <strong>the</strong> ball; at that moment <strong>the</strong> 24-second “shot clock” starts. The offensive<br />
team must shoot <strong>the</strong> ball with<strong>in</strong> this 24-second <strong>in</strong>terval, fail<strong>in</strong>g to do so results <strong>in</strong> a turnover<br />
(0 po<strong>in</strong>ts). 8 Each team has about 100 possessions per game—a critical element that allows<br />
us to model <strong>the</strong> game as a large number of well-def<strong>in</strong>ed pieces, which allows for a tractable,<br />
quantitative analysis of game play. Crucially we show here (and more formally <strong>in</strong> <strong>the</strong> Ap-<br />
pendix) that <strong>the</strong> large number of possessions implies teams are best off behav<strong>in</strong>g like a<br />
risk-neutral po<strong>in</strong>t maximizer <strong>in</strong> each possession. The two key aspects of risk neutrally max-<br />
imiz<strong>in</strong>g po<strong>in</strong>ts are: 1) at a given time <strong>in</strong> <strong>the</strong> shot clock <strong>the</strong> team only cares about expected<br />
po<strong>in</strong>ts (“risk neutrality”) 2) <strong>the</strong> team does not value time itself, that is, for two shots of equal<br />
expected value, <strong>the</strong> team is <strong>in</strong>different between when <strong>the</strong>y are taken (“time-<strong>in</strong>variance”). In<br />
what follows, we exam<strong>in</strong>e <strong>the</strong>se two conditions through a simple model of <strong>NBA</strong> game play.<br />
The goal of <strong>the</strong> team is to score more po<strong>in</strong>ts than <strong>the</strong> opponent <strong>in</strong> <strong>the</strong> entire game.<br />
Consider <strong>the</strong> two teams, home (h) and away (a). Let Sh,N and Sa,N denote <strong>the</strong> current<br />
scores for <strong>the</strong> home and away team with N offensive possessions (for each team) rema<strong>in</strong><strong>in</strong>g<br />
<strong>in</strong> <strong>the</strong> game respectively. Let Ph,i and (Pa,i) denote <strong>the</strong> number of po<strong>in</strong>ts scored by <strong>the</strong> home<br />
(away) possession on <strong>the</strong> i th possession from <strong>the</strong> end of <strong>the</strong> game. The home team w<strong>in</strong>s if<br />
<strong>the</strong>y have more po<strong>in</strong>ts at <strong>the</strong> end of <strong>the</strong> game. This equal to <strong>the</strong> current score, plus <strong>the</strong><br />
po<strong>in</strong>ts scored <strong>in</strong> subsequent possessions, as given by:<br />
Sh,0 > Sa,0 ⇐⇒ Sh,N +<br />
N<br />
Ph,i > Sa,N +<br />
i=1<br />
N<br />
Pa,i ⇐⇒<br />
i=1<br />
N<br />
Ph,i − Pa,i > Sa,N − Sh,N (1)<br />
To model how po<strong>in</strong>ts are generated, let {µh, σ 2 h } and {µa, σ 2 a} represent <strong>the</strong> mean and<br />
variance of po<strong>in</strong>ts per possession that each team is able to achieve <strong>in</strong> <strong>the</strong> match-up. If <strong>the</strong><br />
number of rema<strong>in</strong><strong>in</strong>g possessions, N, is large, <strong>the</strong> central limit <strong>the</strong>orem gives probability of<br />
<strong>the</strong> home team w<strong>in</strong>n<strong>in</strong>g as:<br />
7 The game stops for out of bounds, fouls and time-outs.<br />
8 The clock is prom<strong>in</strong>ently displayed right above <strong>the</strong> basket. <strong>Shot</strong>s must hit <strong>the</strong> rim to count as a shot<br />
attempt, so you cannot fake a shot.<br />
7<br />
i=1
P (Home W<strong>in</strong>) = P (Sh,0 > Sa,0) = P<br />
N<br />
= Φ<br />
i=1<br />
(Ph,i − Pa,i) > Sa,N − Sh,N<br />
<br />
<br />
Sh,N − Sa,N + N(µh − µa)<br />
<br />
2 N(σh + σ2 <br />
a)<br />
where Φ is <strong>the</strong> CDF of <strong>the</strong> standard normal distribution. Exam<strong>in</strong><strong>in</strong>g this expression,<br />
we see that if <strong>the</strong> score is currently tied and <strong>the</strong> teams are of equal quality, <strong>the</strong> game is a<br />
co<strong>in</strong> toss. Hav<strong>in</strong>g an ability advantage (µ higher than opponent) matters proportional to<br />
<strong>the</strong> rema<strong>in</strong><strong>in</strong>g possessions, which is <strong>in</strong>tuitive. If you are <strong>the</strong> better team, it is more likely<br />
to be reflected <strong>in</strong> <strong>the</strong> game outcome when <strong>the</strong>re are many periods rema<strong>in</strong><strong>in</strong>g. The marg<strong>in</strong>al<br />
impact on w<strong>in</strong>n<strong>in</strong>g for each factor can be easily obta<strong>in</strong>ed by differentiat<strong>in</strong>g equation (3).<br />
The follow<strong>in</strong>g expression gives <strong>the</strong> impact on w<strong>in</strong> probability of a po<strong>in</strong>t scored for <strong>the</strong> home<br />
team:<br />
dP (Home W<strong>in</strong>)<br />
dSh,N<br />
= φ<br />
<br />
(Sh,N − Sa,N) + (N(µh − µa))<br />
<br />
2 N(σh + σ2 <br />
1<br />
<br />
2<br />
a)<br />
N(σh + σ2 a)<br />
2.2 (5)<br />
Expression shows that po<strong>in</strong>ts become <strong>in</strong>creas<strong>in</strong>gly impactful on <strong>the</strong> game outcome when<br />
<strong>the</strong> current score is close (Sh,N − Sa,N small) and few possessions rema<strong>in</strong>, while <strong>the</strong> impact<br />
becomes exceed<strong>in</strong>gly small when <strong>the</strong> score marg<strong>in</strong> is high; <strong>in</strong> <strong>the</strong> appendix we show that<br />
when a po<strong>in</strong>t is worth less than .001 w<strong>in</strong>s, teams tend to give up and pull <strong>the</strong>ir starters out<br />
of <strong>the</strong> game. We have no reason to believe basketball players cont<strong>in</strong>ue to pursue optimal<br />
play after this po<strong>in</strong>t and we exclude <strong>the</strong> correspond<strong>in</strong>g data from our ma<strong>in</strong> analysis.<br />
For <strong>the</strong> risk-neutral and time-<strong>in</strong>variance conditions to hold it must be <strong>the</strong> case that<br />
teams care much more about maximiz<strong>in</strong>g <strong>the</strong> efficiency of <strong>the</strong>ir possessions <strong>in</strong>stead of try<strong>in</strong>g<br />
to play fast (slow) or have high (low) variance possessions (ex. shoot<strong>in</strong>g more 3-po<strong>in</strong>ters).<br />
By compar<strong>in</strong>g ratios of marg<strong>in</strong>al w<strong>in</strong> effects we can determ<strong>in</strong>e <strong>the</strong> appropriate <strong>in</strong>difference<br />
ratios. For ease of demonstration, we assume <strong>the</strong> teams are of approximately equal quality<br />
(µh = µa), <strong>the</strong> basic <strong>in</strong>tuition is unaffected by this simplification, because <strong>the</strong> impact of score<br />
difference and ability difference (weighted by possessions rema<strong>in</strong><strong>in</strong>g), is very similar.<br />
The ratio of <strong>the</strong> value of possessions (extend<strong>in</strong>g or shorten<strong>in</strong>g game by one possession for<br />
each team) to <strong>the</strong> value of a po<strong>in</strong>t is given by:<br />
8<br />
(2)<br />
(3)<br />
(4)
dP (Home W<strong>in</strong>)<br />
dN<br />
dP (Home W<strong>in</strong>)<br />
dSh,N<br />
= 1<br />
2<br />
<br />
(Sa,N − Sh,N)<br />
To understand this equation, let’s exam<strong>in</strong>e <strong>the</strong> case when <strong>the</strong> home team is trail<strong>in</strong>g with<br />
N possessions rema<strong>in</strong><strong>in</strong>g. S<strong>in</strong>ce Sa,N > Sh,N, <strong>the</strong> numerator is positive. Differentiat<strong>in</strong>g<br />
equation (3) with respect to N, one can show that <strong>the</strong> team values possessions <strong>in</strong>tr<strong>in</strong>sically<br />
(it needs to catch-up). The relative values of po<strong>in</strong>ts to possessions is given by equation 6.<br />
Indeed, as <strong>the</strong> number of possessions rema<strong>in</strong><strong>in</strong>g grows, this ratio goes to zero at rate N.<br />
Given that an average possession cycle (one for each team) takes about 30 seconds, <strong>the</strong><br />
number of rema<strong>in</strong><strong>in</strong>g possessions need not greatly exceed <strong>the</strong> score difference before this<br />
quickly becomes a very unattractive tradeoff. Sacrific<strong>in</strong>g a mean<strong>in</strong>gful amount of po<strong>in</strong>ts (on<br />
<strong>the</strong> order of 0.005) to save a few seconds is a los<strong>in</strong>g proposition. In our empirical analysis we<br />
elim<strong>in</strong>ate observations for which <strong>the</strong> terms of this tradeoff imply that seconds are relatively<br />
valuable as compared po<strong>in</strong>ts. The threshold we use entails that if one team is trail<strong>in</strong>g by 10<br />
po<strong>in</strong>ts, we elim<strong>in</strong>ate all observations when less than 33 possessions rema<strong>in</strong>.<br />
The ratio between mean efficiency and variance presents a similar tradeoff:<br />
dP (Home W<strong>in</strong>)<br />
dσ 2 h<br />
dP (Home W<strong>in</strong>)<br />
dµh<br />
=<br />
1<br />
σ 2 h + σ2 a<br />
N<br />
<br />
(Sa,N − Sh,N)<br />
N<br />
≈ 1<br />
2<br />
<br />
(Sa,N − Sh,N)<br />
In terms of variance, <strong>the</strong> choice is essentially between 2-po<strong>in</strong>ters, which offer a variance<br />
of about 1 and 3-po<strong>in</strong>ters, which give variance about 2. The trail<strong>in</strong>g team would like to<br />
<strong>in</strong>crease variance, 9 what this equation says is that switch<strong>in</strong>g <strong>in</strong>to 3-po<strong>in</strong>ters, which would<br />
lead to a drop <strong>in</strong> efficiency because of defensive response and equilibrium conditions, is only<br />
rational when <strong>the</strong> number of rema<strong>in</strong><strong>in</strong>g possessions is quite low. Aga<strong>in</strong> <strong>in</strong> our ma<strong>in</strong> empirical<br />
analysis, we elim<strong>in</strong>ate cases when this ratio diverges mean<strong>in</strong>gfully from zero.<br />
In both cases, when <strong>the</strong> number of possessions rema<strong>in</strong><strong>in</strong>g is large, <strong>the</strong> goal of risk neutrally<br />
maximiz<strong>in</strong>g po<strong>in</strong>ts b<strong>in</strong>ds more tightly. The <strong>in</strong>tuition is that at any given time, <strong>the</strong> team<br />
currently expected to lose would like to <strong>in</strong>duce variance <strong>in</strong> <strong>the</strong> outcome of <strong>the</strong> game. However,<br />
with a large number of possessions rema<strong>in</strong><strong>in</strong>g, <strong>the</strong> central limit <strong>the</strong>orem makes <strong>the</strong> terms<br />
of this tradeoff very unattractive, because per-possession variance is killed at rate root-N,<br />
and extend<strong>in</strong>g/shorten<strong>in</strong>g <strong>the</strong> game comes at a large cost (sacrific<strong>in</strong>g efficiency on many<br />
possessions) for a small benefit. Although our ma<strong>in</strong> analysis excludes cases <strong>in</strong> which risk<br />
neutrality and time <strong>in</strong>variance do not hold <strong>in</strong> a tight approximation, we also exam<strong>in</strong>e <strong>the</strong><br />
9 If teams are of unequal ability, <strong>the</strong> trail<strong>in</strong>g teams <strong>in</strong>centive to <strong>in</strong>duce variance depends on <strong>the</strong>ir relative<br />
ability. If <strong>the</strong>y are much better and a decent number of possessions rema<strong>in</strong>s, this <strong>in</strong>centive goes away, as<br />
<strong>the</strong>y are still favored to w<strong>in</strong>.<br />
9<br />
N<br />
(6)<br />
(7)
excluded cases to see if teams do <strong>in</strong> fact become risk (or time) lov<strong>in</strong>g (or averse) when do<strong>in</strong>g<br />
so <strong>in</strong>creases <strong>the</strong> chances of w<strong>in</strong>n<strong>in</strong>g <strong>the</strong> game. 10<br />
2.3 Anatomy of a possession<br />
In <strong>the</strong> previous subsection we showed that a basketball can be broken down <strong>in</strong>to a series<br />
of nodes, possessions, and modeled as a game of risk-neutral po<strong>in</strong>t maximization at every<br />
node. To help build <strong>in</strong>tuition for <strong>the</strong> reader, <strong>in</strong> this section we expla<strong>in</strong> <strong>the</strong> basic features<br />
of a possession. We def<strong>in</strong>e <strong>the</strong> offensive output of a possession as <strong>the</strong> total po<strong>in</strong>ts scored<br />
before <strong>the</strong> opponent starts its next possession. For <strong>in</strong>stance, if a shot is taken, but missed,<br />
and <strong>the</strong> offense rebounds <strong>the</strong> ball, this does not count as a new possession. We make this<br />
important dist<strong>in</strong>ction because <strong>the</strong> po<strong>in</strong>ts we assign to a specific shot are not necessarily <strong>the</strong><br />
outcome of that shot, ra<strong>the</strong>r is a forward look<strong>in</strong>g metric that captures how well <strong>the</strong> team<br />
does conditional upon <strong>the</strong> player us<strong>in</strong>g <strong>the</strong> possession versus wait<strong>in</strong>g.<br />
Table 1: Ways Possessions Can Be Used<br />
Outcome Total <strong>in</strong> Sample Average per-game Po<strong>in</strong>ts Per Possession<br />
Made 2-po<strong>in</strong>ter 390763 26.3495 2.0579<br />
Made 3-po<strong>in</strong>ter 88009 5.9345 3.0226<br />
Turnover 190443 12.8417 0<br />
Fouled <strong>in</strong> <strong>the</strong> Act of Shoot<strong>in</strong>g 141737 9.5575 1.5707<br />
Missed 2 po<strong>in</strong>ter 438231 29.5503 0.3547<br />
Missed 3 po<strong>in</strong>ter 134777 9.0881 0.2833<br />
All (possessions) 1,376,893 92.8451 1.0624<br />
This br<strong>in</strong>gs up an important concept of “us<strong>in</strong>g” <strong>the</strong> possession. A player is said to use<br />
<strong>the</strong> possession if he is <strong>the</strong> first player from his team to shoot or turn <strong>the</strong> ball over (which<br />
we model as similar to a missed shot with no rebound). Table 2.3 shows <strong>the</strong> empirical<br />
frequencies for <strong>the</strong> ways <strong>in</strong> which a possession can be used and <strong>the</strong> immediate outcome.<br />
Column (1) gives <strong>the</strong> total sample N and Column (2) shows <strong>the</strong> averages on a game level, to<br />
provide an idea of general scor<strong>in</strong>g levels and possession frequency. The table gives <strong>the</strong> first<br />
usage, which is not necessarily how <strong>the</strong> possession ends. For <strong>in</strong>stance, a missed 2-po<strong>in</strong>ter<br />
10 In <strong>the</strong> Appendix we estimate <strong>the</strong> probability of w<strong>in</strong>n<strong>in</strong>g equation, (3), via probit (due to <strong>the</strong> Guassian<br />
error structure implied by <strong>the</strong> central limit <strong>the</strong>orem). This allows us to predict <strong>the</strong> relevant quantities of<br />
<strong>in</strong>terest per possession, and appropriately elim<strong>in</strong>ate cases when <strong>the</strong> marg<strong>in</strong>al impact of time and risk exceeds<br />
<strong>the</strong> tolerance levels we set for our analysis.<br />
10
could be rebounded by <strong>the</strong> offense, and <strong>the</strong>n lost out of bounds. The key actions are shoot<strong>in</strong>g<br />
(ei<strong>the</strong>r 2 or 3-po<strong>in</strong>ter), turnover and draw<strong>in</strong>g a foul. As a rem<strong>in</strong>der, when a player is fouled<br />
<strong>in</strong> <strong>the</strong> act of shoot<strong>in</strong>g he is awarded ei<strong>the</strong>r 2 or 3 unguarded 14-foot shots (depend<strong>in</strong>g on<br />
<strong>the</strong> value of shot he was fouled attempt<strong>in</strong>g) worth 1 po<strong>in</strong>t each. Free-throws are valuable; a<br />
good player makes <strong>the</strong>m about 80% of <strong>the</strong> time. In comparison, 50% is considered good for<br />
a 2-po<strong>in</strong>ter taken <strong>in</strong> <strong>the</strong> course of <strong>the</strong> game and 33% for a 3-po<strong>in</strong>ter, roughly speak<strong>in</strong>g.<br />
2.4 Basketball as a stopp<strong>in</strong>g problem<br />
We are now <strong>in</strong> a position to describe our model of a possession. Given that a team must<br />
use a possession before <strong>the</strong> 24-second shot clock expires (or face an automatic turnover), it<br />
is natural to model a possession as a stopp<strong>in</strong>g problem: which shots should be taken versus<br />
cont<strong>in</strong>u<strong>in</strong>g <strong>the</strong> possession? The added twist is that team must also ensure everyone gets<br />
<strong>the</strong> ball so as to maximize production across <strong>the</strong> l<strong>in</strong>eup. We formally present <strong>the</strong> model <strong>in</strong><br />
Section 3, but give <strong>the</strong> basic <strong>in</strong>tuition here to build <strong>in</strong>tuition, just as <strong>the</strong>ory papers often<br />
present a toy model before jump<strong>in</strong>g <strong>in</strong>to abstract formalism.<br />
We follow <strong>the</strong> standard tools for model<strong>in</strong>g a stopp<strong>in</strong>g problem: opportunities are stochas-<br />
tically generated by some process and <strong>the</strong> decision maker must decide whe<strong>the</strong>r to realize an<br />
opportunity or wait for a (potentially better) opportunity down <strong>the</strong> l<strong>in</strong>e. Here <strong>the</strong> relevant<br />
quantity is a ‘shot opportunity.” We model shot opportunities as draws from a player-l<strong>in</strong>eup-<br />
defense specific distribution. That is, our most general specification allows for not only player<br />
ability heterogeneity, but also l<strong>in</strong>eup composition effects and opponent effects. In each sec-<br />
ond of <strong>the</strong> shot clock (which we observe <strong>in</strong> discrete <strong>in</strong>crements), we assume that <strong>the</strong> player<br />
with <strong>the</strong> ball observes a shot opportunity with mean-zero noise (bias <strong>in</strong> observation is per-<br />
fectly co-l<strong>in</strong>ear decision error, so this assumption is necessary). The player must quickly<br />
decide whe<strong>the</strong>r to shoot or not. While it might not be literally true that a player observes<br />
an opportunity every second, we easily <strong>in</strong>corporate <strong>the</strong>se “non-opportunities” as draws that<br />
have very low value.<br />
To complete <strong>the</strong> model, we assume that <strong>the</strong> defense must allocate a f<strong>in</strong>ite amount of defen-<br />
sive attention to m<strong>in</strong>imize offensive output. We solve for equilibrium us<strong>in</strong>g familiar dynamic<br />
programm<strong>in</strong>g techniques. Equilibrium requires that offensive players adopt monotonically<br />
decl<strong>in</strong><strong>in</strong>g (with time rema<strong>in</strong><strong>in</strong>g on <strong>the</strong> shot clock) “cut-thresholds.” A cut-threshold deter-<br />
m<strong>in</strong>es which opportunities should be realized and which should be passed up—if a realization<br />
is observed to be above <strong>the</strong> threshold, <strong>the</strong> shot is taken. We call this condition dynamic effi-<br />
ciency. It ensures that po<strong>in</strong>ts are maximized given <strong>the</strong> time-constra<strong>in</strong>t. Allocative efficiency<br />
is our concept of optimally shar<strong>in</strong>g <strong>the</strong> ball. It requires that all players <strong>in</strong> a given l<strong>in</strong>eup<br />
11
Figure 1: Shoot<strong>in</strong>g hazard by how <strong>the</strong> possession orig<strong>in</strong>ates as a function of time rema<strong>in</strong><strong>in</strong>g<br />
on <strong>the</strong> shot clock.<br />
have <strong>the</strong> same marg<strong>in</strong>al efficiencies each po<strong>in</strong>t of <strong>the</strong> shot clock. If this condition fails, one<br />
player should get passed <strong>the</strong> ball more (or less), because output would <strong>in</strong>crease by do<strong>in</strong>g<br />
so. If both conditions hold, all players <strong>in</strong> <strong>the</strong> l<strong>in</strong>e-up use <strong>the</strong> same cut-thresholds—<strong>the</strong>se<br />
thresholds maximize total output, conditional on <strong>the</strong> opportunity generat<strong>in</strong>g function. 11<br />
Like all stopp<strong>in</strong>g problems, <strong>the</strong> key challenge is model<strong>in</strong>g shots observed by <strong>the</strong> player,<br />
but not <strong>the</strong> econometrician. S<strong>in</strong>ce our equilibrium concept says <strong>the</strong> marg<strong>in</strong>al shot should<br />
exactly equal <strong>the</strong> cont<strong>in</strong>uation value we must answer <strong>the</strong> question, “What would happen if<br />
player X shot more with t seconds rema<strong>in</strong><strong>in</strong>g on <strong>the</strong> shot clock?” We describe our approach<br />
<strong>in</strong> Section 4. The <strong>in</strong>tuition is that we use <strong>the</strong> fact that later <strong>in</strong> <strong>the</strong> shot clock, usage rates<br />
(probability of shoot<strong>in</strong>g conditional on reach<strong>in</strong>g that period, “shoot<strong>in</strong>g hazard”) <strong>in</strong>crease for<br />
all players. This gives us an <strong>in</strong>strument to identify marg<strong>in</strong>al shoot<strong>in</strong>g opportunities, and thus<br />
<strong>the</strong> distribution of shot opportunities fac<strong>in</strong>g <strong>the</strong> player. Figure 1 shows <strong>the</strong> comb<strong>in</strong>ed hazard<br />
for 5-man l<strong>in</strong>eups presented by how <strong>the</strong> possession starts. Note that <strong>the</strong> hazard <strong>in</strong>creases as<br />
time w<strong>in</strong>ds down. This is <strong>in</strong>tuitive, as <strong>the</strong> end node approaches, <strong>the</strong> cut-thresholds drop, and<br />
thus shots are taken at higher frequency (conditional on reach<strong>in</strong>g that po<strong>in</strong>t). The <strong>the</strong>oretical<br />
optimum l<strong>in</strong>e is expla<strong>in</strong>ed <strong>in</strong> detail later—it is derived through <strong>the</strong> assumption of player-<br />
specific uniformly distributed shot opportunities. The figure also shows that possessions<br />
orig<strong>in</strong>at<strong>in</strong>g as steals tend to have a “surpris<strong>in</strong>g” number of shots taken early <strong>in</strong> <strong>the</strong> shot<br />
clock. These are what are known as “fast-breaks,” <strong>in</strong> which <strong>the</strong> team rushes down <strong>the</strong> court<br />
to take an unguarded shot. Possessions that start with <strong>the</strong> ball be<strong>in</strong>g <strong>in</strong>-bounded from under<br />
11 It is of course possible that <strong>the</strong> team could run a better overall offense, that is one that generates better<br />
opportunities for all players, what this paper studies is a question of constra<strong>in</strong>ed optimization, where <strong>the</strong><br />
constra<strong>in</strong>t is <strong>the</strong> offense, <strong>the</strong> coach, player ability, etc.<br />
12
<strong>the</strong> opponent’s basket (made field goal, made free throw, dead ball turnover) tend to “start<br />
later,” <strong>in</strong> that <strong>the</strong> hazard is very low until about 17 seconds rema<strong>in</strong><strong>in</strong>g, <strong>in</strong>dicat<strong>in</strong>g <strong>the</strong> ball<br />
had not been brought <strong>in</strong>to <strong>the</strong> offensive range yet. The hazards converge at 12 seconds. S<strong>in</strong>ce<br />
this gives us an equal play<strong>in</strong>g field (we are confident <strong>the</strong> team has <strong>the</strong> ball <strong>in</strong> <strong>the</strong> offensive<br />
half court and are not gett<strong>in</strong>g easy shots based on steals, etc), we only <strong>in</strong>clude <strong>the</strong> f<strong>in</strong>al 12<br />
seconds <strong>in</strong> our ma<strong>in</strong> analysis.<br />
In <strong>the</strong> next section we present <strong>the</strong> general model <strong>in</strong> full detail. In Section 4 we present<br />
our identification strategy. The purpose of <strong>the</strong> basketball primer is to make <strong>the</strong>se sections<br />
parsable by readers heretofore unfamiliar with basketball.<br />
3 Data<br />
This analysis is entirely based on play-by-play data for all <strong>NBA</strong> games from 2006-2010 (four<br />
seasons). The game logs detail all <strong>the</strong> players on <strong>the</strong> court <strong>the</strong> outcome of every play. All<br />
variables discussed here<strong>in</strong> are constructed (via some fairly extensive cod<strong>in</strong>g) from <strong>the</strong> raw<br />
game logs. Approximately 100 games (out of <strong>the</strong> 4,920 played dur<strong>in</strong>g this time period) are<br />
miss<strong>in</strong>g from this data set. We have no reason to believe <strong>the</strong>ir omission is anyth<strong>in</strong>g o<strong>the</strong>r<br />
than random.<br />
Table 2: Data Overview and Description<br />
Event/Action Description<br />
Offensive/defensive l<strong>in</strong>e-up Players on court at given time<br />
Game-time M<strong>in</strong>utes and second of each event<br />
Game day Date of game<br />
Shooter Player/time of <strong>the</strong> action<br />
Rebound/assist Player/time of <strong>the</strong> action<br />
Foul Shoot<strong>in</strong>g, non-shoot<strong>in</strong>g, flagrant, illegal defense<br />
x,y coord<strong>in</strong>ate of shot Physical location of shot<br />
Turnover Broken down by bad pass, dribbl<strong>in</strong>g error, charge, lost ball<br />
We performed data clean<strong>in</strong>g to elim<strong>in</strong>ate cases where <strong>the</strong> assumptions of our model are<br />
clearly violated. First, we remove possessions where risk-neutrality is violated, such as end of<br />
quarter and end of game situations. S<strong>in</strong>ce our model applies only to half-court sets and not<br />
to “fast-breaks” (actions taken immediately after steals or offensive rebounds, which <strong>in</strong>volve<br />
fairy trivial decision mak<strong>in</strong>g). We def<strong>in</strong>e a fast break as any possession used with<strong>in</strong> <strong>the</strong> first<br />
seven seconds of <strong>the</strong> offense and have purged all such possessions from <strong>the</strong> data. F<strong>in</strong>ally,<br />
13
Table 3: Overview of Possession Data<br />
Method of Possession Orig<strong>in</strong> N Ē ˆσ 2 E<br />
SC ¯<br />
¯<br />
SCerror ˆσ 2 SCerror<br />
All 1376893 1.0624 1.2255 9.8089 -1.5575 13.1961<br />
Opponent Made Field Goal 537811 1.0316 1.2277 7.0405 -2.3012 4.3343<br />
Defensive Rebound 477576 1.0727 1.2285 12.2681 0.3059 4.1053<br />
Steal (game play does not stop) 106397 1.2555 1.1550 15.3404 -0.1231 0.7034<br />
Opponent Turnover (game play does stop) 104394 1.0227 1.2225 9.0007 -0.4148 0.5411<br />
Opponent Made Free Throw 150715 1.0308 1.2284 8.5500 -0.4940 0.3936<br />
we remove situations when one team is lead<strong>in</strong>g by a po<strong>in</strong>t marg<strong>in</strong> so wide that <strong>the</strong> game<br />
outcome is no longer <strong>in</strong> doubt as our model only applies to situations where <strong>the</strong> outcome of<br />
<strong>the</strong> game can reasonably be effected by chosen offensive strategy.<br />
4 Model<br />
In this section, we first motivate our general model of basketball as one of po<strong>in</strong>t maximization<br />
per possession. Importantly this requires an assumption of risk-neutrality, which we defend<br />
with data and simulations <strong>in</strong> <strong>the</strong> Appendix. The basic <strong>in</strong>tuition is that a game <strong>in</strong>volves a<br />
large number of possessions, so sacrific<strong>in</strong>g mean efficiency to <strong>in</strong>crease or decrease variance<br />
has a first order effect on <strong>the</strong> mean but only a second order impact on <strong>the</strong> variance, so it<br />
is almost surely a los<strong>in</strong>g strategy. It is trivial to rule out <strong>the</strong> rare end of quarter situations<br />
where this logic collapses. The general model <strong>in</strong>corporates <strong>the</strong> <strong>in</strong>tertemporal trade-off of<br />
realiz<strong>in</strong>g an opportunity versus cont<strong>in</strong>u<strong>in</strong>g <strong>the</strong> possession and <strong>the</strong> <strong>in</strong>terpersonal trade-off<br />
of how frequently each member of <strong>the</strong> team should shoot at a given time <strong>in</strong>terval. The<br />
latter will be <strong>in</strong>terpreted here as cut-po<strong>in</strong>ts for each player’s worst selected shot <strong>in</strong> a pure<br />
strategy equilibrium. But, could also be <strong>in</strong>terpreted as a vector of shoot<strong>in</strong>g frequencies <strong>in</strong><br />
<strong>the</strong> correspond<strong>in</strong>g Harsanyi purified mixed strategy equilibrium. To fit an abstract model<br />
to observed play we must make assumptions that sufficiently simplify <strong>the</strong> complex game of<br />
basketball. The more restrictive <strong>the</strong> assumptions, <strong>the</strong> stricter <strong>the</strong> test of optimality, but <strong>the</strong><br />
greater chance of misspecification. As such, we adopt a telescop<strong>in</strong>g approach <strong>in</strong> which we<br />
start with a m<strong>in</strong>imal set of assumptions that provide a necessary condition of optimality. We<br />
<strong>the</strong>n add assumptions to <strong>in</strong>crease <strong>the</strong> power our optimality tests and use robustness checks<br />
(overidentification tests) to ensure <strong>the</strong>ir validity.<br />
14
4.1 Def<strong>in</strong>ition of Variables<br />
There are 820 dist<strong>in</strong>ct players <strong>in</strong> our six year sample, denote each one by i ∈ {1, ..., 820}.<br />
These players are observed to play 1376893 different possessions - all equilibrium concepts<br />
will be possession specific. Each possession will be <strong>in</strong>dexed by p ∈ {1, ..., 1376893}. In any<br />
possession p <strong>the</strong>re are exactly five offensive and defensive players.<br />
Op ≡ {i: player i is on offense on possession t}<br />
Dp ≡ {i: player i is on defense on possession t}<br />
and <strong>the</strong> follow<strong>in</strong>g <strong>in</strong>dicator variables.<br />
The vectors Op and Dp are sorted by position such that Op,1 denotes <strong>the</strong> current Po<strong>in</strong>t<br />
Guard for <strong>the</strong> offensive team and Dp,5 denotes <strong>the</strong> Center for <strong>the</strong> defensive team. Additionally<br />
<strong>the</strong>re are 30 different teams <strong>in</strong> our data play<strong>in</strong>g over 6 seasons. We will consider each team-<br />
season a separate unit and denote <strong>the</strong>ir presence on offense (defense) dur<strong>in</strong>g possession p by<br />
<strong>the</strong> notation T Op ∈ {1, ..., 180} (T Dp ∈ {1, ..., 180}).<br />
The game of half-court offense is def<strong>in</strong>ed to occur only when <strong>the</strong> current value of <strong>the</strong> shot<br />
clock is weakly below 12 (t ∈ T 12 ≡ {0, ..., 12}). More generally, convex sets of shot clock<br />
periods may be denoted:<br />
T t = {s ∈ N : 0 ≤ s ≤ t}<br />
Tt = {s ∈ N : t ≤ s ≤ 24}<br />
T t′<br />
t = {s ∈ N : t ≤ s ≤ t ′ }.<br />
In every second of half-court offense, it is possible that an <strong>in</strong>dividual offensive player might<br />
use <strong>the</strong> possession by attempt<strong>in</strong>g a shot, draw<strong>in</strong>g a shoot<strong>in</strong>g foul, or committ<strong>in</strong>g a turnover.<br />
Once a possession is used, half-court offense ends and <strong>the</strong> offensive team is not allowed to<br />
’use’ ano<strong>the</strong>r possession until <strong>the</strong> o<strong>the</strong>r team has had a turn with <strong>the</strong> ball. Possession use<br />
events are denoted as below:<br />
15
Up,i,t = 1{player i used possession p <strong>in</strong> shot clock period t}<br />
Up,i = <br />
Up,i,t = 1{player i used possession p}<br />
t∈T<br />
Ui,t = <br />
Up,i,t = # of possessions used by player i <strong>in</strong> shot clock period t<br />
p<br />
Ui = <br />
Ui,t = <br />
Up,i = # of possessions used by player i<br />
t<br />
p<br />
Similarly, we may be <strong>in</strong>terested <strong>in</strong> <strong>the</strong> total number of opportunities an offensive player<br />
has to use a possession.<br />
Np,i,t = 1{i ∈ Op ∩ {∩i∈Op ∩s∈t∗ (1 − Up,i,s)}}<br />
The above expression should be understood to mean that Np,i,t = 1 if and only if player i<br />
is <strong>in</strong> <strong>the</strong> offensive l<strong>in</strong>eup for possession p and his team has not yet used <strong>the</strong> possession with<br />
t seconds rema<strong>in</strong><strong>in</strong>g on <strong>the</strong> shot clock.<br />
F<strong>in</strong>ally, any time Up,i,t = 1 we track <strong>the</strong> total number of po<strong>in</strong>ts scored before <strong>the</strong> o<strong>the</strong>r<br />
team beg<strong>in</strong>s its possession and store it as Pp,i,t 12 . It is not necessary that player i score all<br />
(or any) of <strong>the</strong>se po<strong>in</strong>ts, merely that <strong>the</strong>y are an end result preceded by his <strong>in</strong>itial decision<br />
to ’use’ <strong>the</strong> possession.<br />
{P, N, T }p,i,t will comprise <strong>the</strong> observed features of our model, but we will often refer to<br />
two <strong>in</strong>tuitive, reduced form expressions <strong>in</strong>stead. Usage rates are def<strong>in</strong>ed as <strong>the</strong> probability<br />
that player i will use a particular possession conditional on hav<strong>in</strong>g <strong>the</strong> opportunity to do<br />
so. Traditional usage rates 13 refer to <strong>the</strong> aggregate probability of player i us<strong>in</strong>g a possession<br />
conditional on be<strong>in</strong>g on <strong>the</strong> court.<br />
up,i ≡ P (Up,i = 1|i ∈ Op)<br />
In this paper we will be more <strong>in</strong>terested <strong>in</strong> hazard rates of possession use at a given value<br />
of <strong>the</strong> shot clock. Informally, we want <strong>the</strong> probability that player i uses a possession with<br />
exactly t seconds rema<strong>in</strong><strong>in</strong>g on <strong>the</strong> shot clock conditional on <strong>the</strong> possession be<strong>in</strong>g unused up<br />
12 Up,i,t = 0 =⇒ Pp,i,t = 0<br />
13 www.basketball-reference.com presents one version<br />
16
to that po<strong>in</strong>t.<br />
up,i,t ≡ P (Up,i,t = 1|Np,i,t = 1, p ∈ Op)<br />
Also, conditional on player i select<strong>in</strong>g to use a possession, we are <strong>in</strong>terested <strong>in</strong> <strong>the</strong> average<br />
return of po<strong>in</strong>ts to that possession. Informally, we are be <strong>in</strong>terested <strong>in</strong> <strong>the</strong> efficiency of player<br />
i’s possession use.<br />
ep,i,t ≡ E(Pp,i,t|Up,i,t = 1)<br />
For purposes of estimation, we will be <strong>in</strong>terested <strong>in</strong> empirical analogs of <strong>the</strong>se quantities.<br />
Even though our equilibrium concepts will be possession-specific, we will have to average over<br />
subsets of possessions <strong>in</strong> order for this to be mean<strong>in</strong>gful. We will denote this by replac<strong>in</strong>g<br />
<strong>the</strong> p subscript with notation <strong>in</strong>dicat<strong>in</strong>g a given subset of possessions. As one example, we<br />
may wish to evaluate player i’s performance only aga<strong>in</strong>st oppos<strong>in</strong>g team x. Then:<br />
A ≡ {p : T Dp = x}<br />
uA,i,t = P (Up,i,t = 1|Np,i,t = 1 ∩ p ∈ A)<br />
ûA,i,t = UA,i,t<br />
NA,i,t<br />
eA,i,t = E(Pp,i,t|Up,i,t = 1 ∩ p ∈ A)<br />
êA,i,t = PA,i,t<br />
UA,i,t<br />
.<br />
The empirical analogs (denoted by hats) are simple sample averages and thus obey stan-<br />
dard Central Limit Theorems.<br />
4.2 General Model<br />
In this section we will write down a general model of a half-court offense game between an<br />
offensive and a defensive player. This game will give economics significance to <strong>the</strong> observable<br />
quantities <strong>in</strong>troduced <strong>in</strong> <strong>the</strong> previous section.<br />
On each possession <strong>the</strong> Offensive and Defensive l<strong>in</strong>eups compete <strong>in</strong> a zero-sum game<br />
over <strong>the</strong> number of po<strong>in</strong>ts result<strong>in</strong>g on <strong>the</strong> possession. Offenses do this by runn<strong>in</strong>g plays to<br />
17
create scor<strong>in</strong>g opportunities and <strong>the</strong>n mak<strong>in</strong>g quick decisions about when an opportunity is<br />
valuable enough to warrant its exploitation. Defenses respond by allocat<strong>in</strong>g scarce defensive<br />
attention to <strong>in</strong>terfere with <strong>the</strong> Offense’s ability to realize efficient opportunities.<br />
Formally, defensive strategy on each possession is summarized by <strong>the</strong> static selection<br />
of five choice variables dp ≡ {dp,j}j∈{1,...,5} which represent <strong>the</strong> average level of attention<br />
devoted to each offensive player dur<strong>in</strong>g <strong>the</strong> possession. These variables are constra<strong>in</strong>ed by<br />
<strong>the</strong> abilities of <strong>the</strong> defenders accord<strong>in</strong>g to (8).<br />
<br />
j∈{1,...,5}<br />
αT Dp,j dp,j ≤ 1 (8)<br />
Here <strong>the</strong> team specific parameters (αx ≡ {αx,j}j∈{1,...,5}) allow arbitrary heterogeneity <strong>in</strong><br />
<strong>the</strong> budget constra<strong>in</strong>ts faced by different defensive teams. Intuitively, if team x has a very<br />
good defensive Shoot<strong>in</strong>g Guard, it might face a relatively low cost toward defend<strong>in</strong>g oppos<strong>in</strong>g<br />
Shoot<strong>in</strong>g Guards (αx,2 relatively small). This is not equivalent to, but will generally result <strong>in</strong>,<br />
a higher level of equilibrium defensive pressure exerted aga<strong>in</strong>st oppos<strong>in</strong>g Shoot<strong>in</strong>g Guards.<br />
In real life, basketball defenses have a far higher dimensional choice set. There are<br />
many different ways <strong>in</strong> which a defense might allocate more effort to conta<strong>in</strong> a particularly<br />
proficient scorer. The above expression views <strong>the</strong> selection and implementation of such<br />
specific strategies as a black box <strong>in</strong>side each team’s budget constra<strong>in</strong>t.<br />
Offensive strategy is more richly modeled and allows for dynamic decisions by each player<br />
over <strong>the</strong> course of <strong>the</strong> shot clock. It is variation <strong>in</strong> <strong>the</strong>se decisions as <strong>the</strong> shot clock ticks<br />
toward zero that motivates our comparison to <strong>the</strong> optimal stopp<strong>in</strong>g literature. At every one<br />
second long <strong>in</strong>terval of <strong>the</strong> shot clock, each offensive player has <strong>the</strong> opportunity ei<strong>the</strong>r to use<br />
<strong>the</strong> possession or to wait until <strong>the</strong> next period of <strong>the</strong> shot clock. Specifically, with t seconds<br />
rema<strong>in</strong><strong>in</strong>g on <strong>the</strong> shot clock, every player i ∈ Op draws an unbiased measure of <strong>the</strong> expected<br />
number of po<strong>in</strong>ts his team would get from his immediate use of possession p. Formally ∀t,<br />
denote a probability space by {Ωt, Ft, Pt}. Then Φ will denote a family of random variables<br />
on this space for <strong>the</strong> realization of scor<strong>in</strong>g opportunities.<br />
Φi,t,dp,i : Ωt → R<br />
ηp,i,t = Φi,t,dp,i (ω)<br />
The measure Pt is assumed cont<strong>in</strong>uous so that ∀i ∈ Op we may write <strong>the</strong> correspond<strong>in</strong>g<br />
distribution of ηp,i,t with measure and density below.<br />
18
Fi,t,dp,i (x) = P (ηp,i,t ≤ x)<br />
dFi,t,dp,i (x)<br />
fi,t,dp,i (x) =<br />
dx<br />
ηp,i,t represents not just expected po<strong>in</strong>ts on an immediate shot, but also whatever value<br />
his team is likely to get (<strong>in</strong>clud<strong>in</strong>g po<strong>in</strong>ts scored after offensive rebounds or from foul shots)<br />
before end<strong>in</strong>g <strong>the</strong>ir possessions. ηp,i,t is also effected by <strong>the</strong> likelihood that player i should<br />
turn <strong>the</strong> ball over <strong>in</strong> his attempt to use of possession, thus immediately nett<strong>in</strong>g his team<br />
zero po<strong>in</strong>ts.<br />
on:<br />
The various subscripts <strong>in</strong>dicate that <strong>the</strong> opportunity distribution may depend arbitrarily<br />
1. The player shoot<strong>in</strong>g <strong>the</strong> ball: i.<br />
2. The period of <strong>the</strong> shot clock we are <strong>in</strong>: t.<br />
3. The level of defensive attention focused on player i: dp,i<br />
Restrictions to <strong>the</strong> above dependencies will be discussed as necessary for identification<br />
and convenience of estimation. For now we only impose a little bit of reasonableness on<br />
<strong>the</strong> structure of defensive impact <strong>in</strong> order to <strong>in</strong>sure an equilibrium defensive allocation <strong>in</strong><br />
our model.<br />
Assumption RD: Every s<strong>in</strong>gle quantile of a players opportunity distribution is strictly<br />
decreas<strong>in</strong>g and strictly convex <strong>in</strong> <strong>the</strong> level of defensive attention received by that player.<br />
This can be <strong>in</strong>terpreted to mean that defense is effective at all marg<strong>in</strong>s, but has dim<strong>in</strong>ish<strong>in</strong>g<br />
returns to scale. F<strong>in</strong>ally, if any player is completely ignored on defense, than a small amount<br />
of defensive attention is extremely effective.<br />
∀x ∈ [0, 1] :<br />
∂F −1<br />
i,t,dp,i (x)<br />
∂dp,i<br />
∂2F −1<br />
i,t,dp,i (x)<br />
∂d2 p,i<br />
∂F<br />
lim<br />
dp,i→0<br />
−1<br />
i,t,dp,i (x)<br />
∂dp,i<br />
< 0 (9)<br />
> 0 (10)<br />
= −∞ (11)<br />
Also we make no restriction on <strong>the</strong> precision with which a player observes <strong>the</strong> quality of<br />
potential shots, but do require that he has unbiased expectations.<br />
19
Assumption UE: E(Pp,i,t|Up,i,t = 1, ηp,i,t) = ηp,i,t<br />
This restriction is necessary because we have no means of differentiat<strong>in</strong>g players who<br />
choose to shoot too much from player’s who are just deluded <strong>in</strong>to th<strong>in</strong>k<strong>in</strong>g all of <strong>the</strong>ir shots<br />
are good shots. O<strong>the</strong>rwise, this is not a mean<strong>in</strong>gful restriction.<br />
If no player chooses to exercise <strong>the</strong> possession <strong>in</strong> time period t, <strong>the</strong>n that period of <strong>the</strong><br />
shot clock is allowed to pass and each player will realize new scor<strong>in</strong>g opportunities <strong>in</strong> period<br />
t − 1. Note that we are not assum<strong>in</strong>g a viable scor<strong>in</strong>g opportunity for every player at every<br />
<strong>in</strong>terval of <strong>the</strong> shot clock, it is entirely possible that dur<strong>in</strong>g many seconds of <strong>the</strong> shot clock<br />
a player will realize scor<strong>in</strong>g opportunities of arbitrarily small expected value. S<strong>in</strong>ce <strong>the</strong>se<br />
opportunities are very rarely exercised, it is not of paramount importance how we model<br />
<strong>the</strong>m.<br />
In period 0, <strong>the</strong> shot clock is about to expire. If no player uses <strong>the</strong> possession dur<strong>in</strong>g this<br />
period, <strong>the</strong>n play is stopped and <strong>the</strong> ball is awarded to <strong>the</strong> o<strong>the</strong>r team with no po<strong>in</strong>ts for<br />
<strong>the</strong> Offense.<br />
4.2.1 General Solution: Offense<br />
Conditional on defensive strategy dp, <strong>the</strong> offense seeks to maximize <strong>the</strong> expected value of<br />
<strong>the</strong>ir possession with t second left. They do so by adopt<strong>in</strong>g monotone cut-thresholds<br />
(cp,i,t) for each player to use <strong>the</strong> possession <strong>in</strong> each period of <strong>the</strong> shot clock. If (and only<br />
if) player i realizes a scor<strong>in</strong>g opportunity of larger value than his cut threshold than he will<br />
choose to use <strong>the</strong> possession.<br />
Assumption MC: ∃cp,i,t, ηp,i,t > cp,i,t ⇐⇒ Up,i,t = 1<br />
This restricts all shoot<strong>in</strong>g decisions to be monotone <strong>in</strong> perceived shot quality. However,<br />
we place no assumptions (beyond be<strong>in</strong>g unbiased) on <strong>the</strong> abilities of player’s to accurately<br />
recognize <strong>the</strong> quality of <strong>the</strong>ir shot so this is not restrictive nor <strong>in</strong>compatible with perceived<br />
’<strong>in</strong>consistent’ shot selection <strong>in</strong> basketball games. Additionally, for a player with a given<br />
opportunity distribution, selection of a cut-threshold (cp,i,t) is equivalent to selection of a<br />
usage hazard rate (up,i,t) and a level of aggregate efficiency for each player <strong>in</strong> each time<br />
period (ep,i,t).<br />
20
∞<br />
up,i,t = P (ηp,i,t > cp,i,t) =<br />
cp,i,t<br />
ep,i,t = E(ηp,i,t|ηp,i,t > cp,i,t) =<br />
fi,t,dp,i (x)dx<br />
∞<br />
cp,i,t<br />
xfi,t,dp,i (x)dx<br />
The cut threshold carries greater economic significance, but for technical game <strong>the</strong>ory<br />
reasons we will regard <strong>the</strong> implied selection of <strong>the</strong> usage hazard rates as <strong>the</strong> ”action” of <strong>the</strong><br />
”offensive player”.<br />
The offense’s problem is now entirely reduced to <strong>the</strong> selection of cut-thresholds for each<br />
player and <strong>in</strong> each shot clock period. <strong>Optimal</strong> choices are derived <strong>in</strong> <strong>the</strong> value-function<br />
approach below 14 . In writ<strong>in</strong>g down <strong>the</strong> value function below we must formally impose our<br />
previously discussed assumption of risk neutrality.<br />
Assumption RN: The Offense’s value function on possession p with t seconds rema<strong>in</strong><strong>in</strong>g<br />
(Vp,t) is exactly equal to <strong>the</strong>ir expected po<strong>in</strong>t return on <strong>the</strong> possession. The Defense’s value<br />
function will not be seperately written down, but <strong>the</strong>ir objective is <strong>the</strong> m<strong>in</strong>imization of Vp,t.<br />
up,i,t<br />
Vp,t = E(Pp|UOp,Tt+1 = 0) = eOp,T t (12)<br />
We may now formally write <strong>the</strong> offense’s problem <strong>in</strong> (13) below.<br />
Vp,t = max Vp,t−1 +<br />
{cp,j,t}i∈Op<br />
<br />
∞<br />
(x − Vp,t−1)dFi,t,dp,i<br />
cp,j,t<br />
j∈Op<br />
Vp,−1 = 0<br />
<strong>Optimal</strong> cut thresholds are derived from a first order condition, which gives a familiar<br />
stopp<strong>in</strong>g result.<br />
Dynamic Efficiency: ∀i ∈ Op, cp,i,t = Vp,t−1<br />
A player should only choose to shoot if he realizes a scor<strong>in</strong>g opportunity more valuable<br />
than his outside option of cont<strong>in</strong>u<strong>in</strong>g <strong>the</strong> possession. Dynamic Efficiency is a direct test<br />
14 To preserve tractability we ’rule out’ cases where two players are <strong>in</strong>duced to shoot <strong>in</strong> <strong>the</strong> same shot clock<br />
period by assum<strong>in</strong>g that each player already knows none of his teammates might also use <strong>the</strong> possession <strong>in</strong><br />
this period.<br />
21<br />
(13)
of whe<strong>the</strong>r or not basketball players understand and correctly solve <strong>the</strong>ir optimal stopp<strong>in</strong>g<br />
problem. Dynamic Efficiency by all five players is a sufficient condition for optimal offensive<br />
strategy.<br />
Under knowledge of <strong>the</strong> opportunity distribution, <strong>the</strong> value function, and optimal cut-<br />
thresholds could <strong>the</strong>n be solved by recursively apply<strong>in</strong>g <strong>the</strong> criterion of Dynamic Efficiency.<br />
Thus,as seen <strong>in</strong> <strong>the</strong> follow<strong>in</strong>g lemma, <strong>the</strong> offensive player always has a unique best response<br />
to any allocation of <strong>in</strong>terference chosen by <strong>the</strong> Defense.<br />
Lemma 1. For a given set of offensive players <strong>the</strong>re exists a mapp<strong>in</strong>g BRO : R5 T ×5 → R<br />
that denotes <strong>the</strong> Offense’s best response cut-thresholds for every possible defensive allocation<br />
dp. The best response is always unique and has nice properties. Fur<strong>the</strong>r <strong>the</strong> selection of cut<br />
thresholds, cp, is equivalent to <strong>the</strong> selection of conditional usage rates, up, or unconditional<br />
usage rates, qp.<br />
Proof.<br />
cp,Op,0 = Vp,−1 = 0<br />
cp,Op,1 = Vp,0 = <br />
i∈Op<br />
...<br />
cp,Op,t+1 = Vp,t = <br />
∞<br />
i∈Op<br />
∞<br />
η · fi,0,di (η)dη<br />
0<br />
Vp,t−1<br />
(η − Vp,t−1) · fi,t,di (η)dη<br />
For <strong>the</strong> case of a uniform opportunity distribution and some T periods rema<strong>in</strong><strong>in</strong>g on<br />
<strong>the</strong> shot clock, it is straightforward to write down such a ’shooter’s sequence 15 . While our<br />
identification results will not rely on such strict assumptions, it is <strong>in</strong>terest<strong>in</strong>g to compare<br />
observed usage behavior to <strong>the</strong> implications of such a model.<br />
As shown <strong>in</strong> Figure 1, prior to t = 12, opportunities seem to differ by how <strong>the</strong> possession<br />
starts. However, from t = 12 on it looks like <strong>the</strong> orig<strong>in</strong> of <strong>the</strong> possession is no longer important<br />
and <strong>NBA</strong> offenses are do<strong>in</strong>g someth<strong>in</strong>g that closely resembles optimally solv<strong>in</strong>g a ’half-court’<br />
offense game while fac<strong>in</strong>g a uniform opportunity distribution. However, our results do not<br />
rely on dynamic programm<strong>in</strong>g to solve for <strong>the</strong> value function. Ra<strong>the</strong>r <strong>the</strong> assumption of risk<br />
neutrality allows us to compute possession values directly based on observed data, as <strong>in</strong> (12).<br />
15 See Sk<strong>in</strong>ner 2011 for an <strong>in</strong>dependent deriviation of this condition<br />
22
We also have a priori <strong>in</strong>terest <strong>in</strong> <strong>the</strong> hypo<strong>the</strong>sis that players anchor <strong>the</strong>ir shoot<strong>in</strong>g deci-<br />
sions to <strong>the</strong> shoot<strong>in</strong>g decisions of <strong>the</strong>ir teammates, ra<strong>the</strong>r than to <strong>the</strong> solution of a dynamic<br />
programm<strong>in</strong>g problem. Thus we also formulate a strictly weaker hypo<strong>the</strong>sis.<br />
Allocative efficiency: ∀i, j ∈ Op, cp,i,t = cp,j,t<br />
Allocative Efficiency requires that all five players on <strong>the</strong> court have <strong>the</strong> same reservation<br />
shot-quality <strong>in</strong> m<strong>in</strong>d. Conditional on a given team-wide shoot<strong>in</strong>g rate, Allocative Efficiency<br />
generates <strong>the</strong> best set of shot opportunities.<br />
4.2.2 General Solution: Defense<br />
For a given matrix of usage hazard rates {up,i,t} i,t∈{Op×T 12 }, we now derive <strong>the</strong> defense’s best<br />
response selection of dp <strong>in</strong> order to m<strong>in</strong>imize average po<strong>in</strong>ts given up. For convenience we<br />
iterate (13) <strong>in</strong> order to get a more convenient expression of <strong>the</strong> value function <strong>in</strong> (14).<br />
Vp,t =<br />
T<br />
( <br />
(1 − up,Op,s)) <br />
t=0<br />
s∈T 12<br />
t<br />
j∈Op<br />
up,j,t ep,j,t =<br />
T <br />
t=0 j∈Op<br />
qp,j,t ep,j,t<br />
Where qp,j,t may be considered to be <strong>the</strong> unconditional probability of player i us<strong>in</strong>g<br />
possession p with t seconds rema<strong>in</strong><strong>in</strong>g on <strong>the</strong> shot clock.<br />
Lemma 2. For a given set of offensive players, a defensive team with budget constra<strong>in</strong>t (8),<br />
<strong>the</strong>re exists a mapp<strong>in</strong>g BRD : R T ×5 → R 5 that denotes <strong>the</strong> Defense’s best response allocation<br />
of <strong>in</strong>terference to every possible level of offensive unconditional usage probabilities qp. The<br />
best response is always unique and has first derivative given <strong>in</strong> <strong>the</strong> proof.<br />
Proof.<br />
ep,k,t = 1<br />
dep,k,t<br />
ddp,k<br />
d 2 ep,k,t<br />
dd 2 p,k<br />
up,k,t<br />
= 1<br />
up,k,t<br />
1<br />
= 1<br />
up,k,t<br />
1−up,k,t<br />
1<br />
1−up,k,t<br />
1<br />
1−up,k,t<br />
F −1<br />
n (p)dp<br />
dF −1<br />
k,t,dp,k (p)<br />
dp 1<br />
< 0<br />
ddp,k<br />
d 2 F −1<br />
k,t,dp,k (p)<br />
dd 2 p,k<br />
dp 2<br />
> 0<br />
Where <strong>in</strong>equalities (1) and (2) above follow from <strong>the</strong> assumptions on defensive <strong>in</strong>terference<br />
<strong>in</strong> equations (9) and (10). Thus (14) is concave <strong>in</strong> <strong>the</strong> defensive allocation. D ≡ { d : α ′ T D d ≤<br />
1 ∩ d ≥ 0} is a compact set and thus a unique best response level of defensive allocation<br />
exists.<br />
23<br />
(14)
Under (11) we can be sure that all players face some equilibrium defense and write down<br />
<strong>the</strong> follow<strong>in</strong>g first order condition.<br />
∀j = k ∈ Op,<br />
T t=0 qp,k,t<br />
∂ep,k,t<br />
∂dp,k<br />
T ∂ep,j,t<br />
t=0<br />
qp,j,t ∂dp,j<br />
= αT Dp,k<br />
αT Dp,j<br />
The numerator and denom<strong>in</strong>ator of <strong>the</strong> left side can be viewed as <strong>the</strong> marg<strong>in</strong>al <strong>in</strong>terference<br />
created by additional defensive attention to offensive players k and j respectively. Intuitively,<br />
(15) just states that <strong>the</strong>ir ratio should be equal to <strong>the</strong> ratio of <strong>the</strong> marg<strong>in</strong>al costs of such<br />
defensive attention as specified <strong>in</strong> (8). We may <strong>the</strong>n apply <strong>the</strong> Implicit Function Theorem<br />
to establish derivatives of our best response correspondence:<br />
dBR i D<br />
dqp,i,t<br />
=<br />
i = j =⇒ dBRi D<br />
dqp,j,t<br />
<br />
dep,i,t<br />
− ddp,i + qi,t d2 <br />
ep,i,t<br />
ddp,idqp,i,t<br />
t<br />
t=0 qp,i,t d2 ep,i,t<br />
dd 2 p,i<br />
= 0<br />
= − dF −1<br />
i,t,d<br />
(1−up,i,t)<br />
p,i<br />
ddp,i<br />
t t=0 qp,i,t d2ep,i,t dd2 p,i<br />
4.2.3 Existence, Uniqueness, and Cont<strong>in</strong>uity of Equilibria<br />
Here we outl<strong>in</strong>e a series of results that demonstrate <strong>the</strong> existence of a unique equilibrium<br />
<strong>in</strong> <strong>the</strong> half-court offense game between <strong>the</strong> Offensive and Defensive player. This equilibrium<br />
will be dependent on <strong>the</strong> capabilities of <strong>the</strong> five offensive players and <strong>the</strong> <strong>in</strong>terference budget<br />
constra<strong>in</strong>t faced by <strong>the</strong> oppos<strong>in</strong>g defense. Additionally, we will demonstrate that small<br />
changes <strong>in</strong> <strong>the</strong> abilities of <strong>the</strong> offensive players will result <strong>in</strong> small changes <strong>in</strong> equilibrium<br />
play. Taken toge<strong>the</strong>r, <strong>the</strong>se results allow us to credibly average over separate possessions <strong>in</strong><br />
order to learn about <strong>the</strong> decision mak<strong>in</strong>g of <strong>in</strong>dividual players.<br />
Theorem 1. In <strong>the</strong> game of half-court offense, with preferences def<strong>in</strong>ed by <strong>the</strong> value function<br />
<strong>in</strong> (14) and defensive budget constra<strong>in</strong>t outl<strong>in</strong>ed <strong>in</strong> (8) and under <strong>the</strong> previously ma<strong>in</strong>ta<strong>in</strong>ed<br />
assumptions, <strong>the</strong>re exists an equilibrium selection of cut thresholds and allocation of defensive<br />
<strong>in</strong>terference<br />
Proof. We regard <strong>the</strong> offensive player as select<strong>in</strong>g <strong>the</strong> unconditional shoot<strong>in</strong>g probabilities<br />
q p,Op,T 12 ∈ [0, 1] 5×12 subject only to <strong>the</strong> constra<strong>in</strong>t that <strong>the</strong>y sum to exactly 1. The defensive<br />
24<br />
> 0<br />
(15)
player is still regarded as select<strong>in</strong>g {dp,j}j∈Op subject to (8). These are both closed, compact<br />
subspaces of R n . In appendix A, we demonstrate that <strong>the</strong> offensive value function (14) is<br />
concave <strong>in</strong> q and convex <strong>in</strong> <strong>the</strong> levels of defensive attention. This is a standard sufficient<br />
condition for existence of a Nash equilibrium. See Osborne and Rub<strong>in</strong>ste<strong>in</strong>, 1994, Proposition<br />
20.3 for one demonstration.<br />
Theorem 2. In <strong>the</strong> game of half-court offense, with preferences def<strong>in</strong>ed by <strong>the</strong> value function<br />
<strong>in</strong> (14) and defensive budget constra<strong>in</strong>t outl<strong>in</strong>ed <strong>in</strong> (8) and under <strong>the</strong> previously ma<strong>in</strong>ta<strong>in</strong>ed<br />
assumptions, <strong>the</strong>re may not exist multiple equilibria.<br />
Proof.<br />
5 Identify<strong>in</strong>g Violations of <strong>Optimal</strong>ity<br />
In <strong>the</strong> next section, we will consider what observable restrictions Allocative and Dynamic<br />
Efficiency impose on our data and lay <strong>the</strong> groundwork for our ultimate tests. We will<br />
beg<strong>in</strong> by only consider<strong>in</strong>g <strong>the</strong> direct implications of <strong>the</strong> model as presented. In order to<br />
<strong>in</strong>crease power we will consider an additional assumption that will allow us to take variation<br />
<strong>in</strong> shot selection across <strong>the</strong> shot clock as exogenous. This assumption enables (almost)<br />
complete identification of our structural model and will generally provide narrow regions of<br />
set identification for cut-thresholds. We will briefly consider estimators of this form, before<br />
ultimately settl<strong>in</strong>g on a parsimonious parametric specification. F<strong>in</strong>ally, we will address <strong>the</strong><br />
implications of error <strong>in</strong> our measurement of <strong>the</strong> shot clock for our identification results and<br />
show how <strong>the</strong> availability of <strong>the</strong> auxiliary data set of shot clock violations can resolve this<br />
problem.<br />
5.1 A Necessary Condition for Dynamic Efficiency<br />
In this section we develop a m<strong>in</strong>imalistic necessary condition for dynamic efficiency. We<br />
address <strong>the</strong> question “do players take shots too frequently and/or too soon <strong>in</strong> <strong>the</strong> shot<br />
clock.”<br />
On possession p, player i only shoots at t if ηp,i,t ≥ cp,i,t. Thus:<br />
ep,i,t = E(ηp,i,t|ηp,i,t ≥ cp,i,t) ≥ cp,i,t<br />
1<br />
= Vp,t−1<br />
2<br />
= ep,Op,T t (16)<br />
Equality (1) is Dynamic Efficiency and equality (2) is <strong>the</strong> appropriate valuation of a<br />
possession under risk-neutrality. Fur<strong>the</strong>r we can <strong>in</strong>tegrate this condition over any set of<br />
possessions and time periods yield<strong>in</strong>g a very general null hypo<strong>the</strong>sis.<br />
25
eA,i,T s<br />
t ≥ eA,O·,T t−1 (17)<br />
Test<strong>in</strong>g of this hypo<strong>the</strong>sis is straightforward from Central Limit Theory.<br />
êA,i,T s<br />
t<br />
∼ N(eA,T s<br />
t , σ2 A,i,T s<br />
t<br />
UA,i,T s<br />
t<br />
êA,T t−1 ∼ N(eA,T t−1, σ2<br />
A,T t−1<br />
)<br />
UA,T t−1<br />
T s<br />
t ∩ T t−1 = ∅ =⇒ êA,i,T s<br />
t ⊥ ê A,T t−1<br />
êA,i,T s<br />
t − êA,T t−1 ∼ N(eA,i,T s<br />
t − eA,T We ran this test for all players <strong>in</strong> our sample...<br />
RUN THIS TEST<br />
)<br />
t−1, σ2<br />
A,T t−1<br />
U A,T t−1<br />
+ σ2 A,i,T s<br />
t<br />
UA,i,T s<br />
t<br />
Our results show that this optimality condition is rarely violated <strong>in</strong> <strong>NBA</strong> play. However,<br />
this analysis makes no statement about <strong>the</strong> possibility that <strong>NBA</strong> players ’undershoot’. Fur-<br />
<strong>the</strong>r it cannot even be taken to fully rule out ’overshoot<strong>in</strong>g’ <strong>in</strong> <strong>the</strong> <strong>NBA</strong>. Typically <strong>the</strong>re<br />
is a wide divergence between a player’s “worst shot” ((cp,i,t)) and average shot (ep,i,t), with<br />
<strong>the</strong> average shot necessarily offer<strong>in</strong>g higher efficiency. If a player’s average is consistently<br />
below <strong>the</strong>ir team’s cont<strong>in</strong>uation value, <strong>the</strong>n <strong>the</strong> team would be better off if he did not shoot<br />
at all. We can make this statement because we know <strong>the</strong> counterfactual of a player tak<strong>in</strong>g<br />
zero shots (his team would get zero po<strong>in</strong>ts and <strong>the</strong>y would be able to cont<strong>in</strong>ue all those<br />
possessions). However, we are also <strong>in</strong>terested <strong>in</strong> situations <strong>in</strong> which players are not actively<br />
hurt<strong>in</strong>g <strong>the</strong>ir team, but are perhaps still us<strong>in</strong>g <strong>the</strong> wrong cut threshold.<br />
In order to make such statements we will need an appropriate counterfactual that gives<br />
us <strong>in</strong>formation about <strong>the</strong> value of a player’s marg<strong>in</strong>al shots. Approaches <strong>in</strong> this paper will<br />
be based off <strong>the</strong> exploitation of variation <strong>in</strong> shoot<strong>in</strong>g hazard rates observed over <strong>the</strong> course<br />
of <strong>the</strong> shot clock.<br />
As an example, we may observe a hypo<strong>the</strong>tical player i with <strong>the</strong> follow<strong>in</strong>g reduced form<br />
statistics.<br />
up,i,t = .04 ep,i,t = 1 (18)<br />
up,i,t−1 = .05 ep,i,t−1 = .95 (19)<br />
26<br />
)
We suppose <strong>the</strong>re is not important difference between periods t and t − 1 and that if our<br />
player <strong>in</strong>creased his aggression he could achieve <strong>the</strong> same levels of usage and efficiency <strong>in</strong><br />
period t. Then by choos<strong>in</strong>g his allocation <strong>in</strong> period t (and not <strong>the</strong> one he chose <strong>in</strong> period<br />
t − 1) he is pass<strong>in</strong>g up opportunities that can be approximately valued at.<br />
cp,i,t ≈ ep,i,t−1up,i,t−1 − ep,i,tup,i,t<br />
up,i,t − up,i,t−1<br />
= .75 (20)<br />
If Vp,i,t−1 < .75 <strong>the</strong>n we can conclude that player i should <strong>in</strong> fact alter his period t<br />
behavior toward what is observed to be capable of <strong>in</strong> period t − 1.<br />
In <strong>the</strong> next section we will build general identification results based on this logic. We<br />
will rely critically on <strong>the</strong> assumption that variation <strong>in</strong> <strong>the</strong> shot clock provides a valid source<br />
of exogenous variation <strong>in</strong> shot selection.<br />
5.2 Structural Identification<br />
We now consider <strong>the</strong> identification of opportunity distributions and cut-thresholds for a<br />
fixed offensive l<strong>in</strong>eup (Op) and oppos<strong>in</strong>g defense (T Dp). Given uniqueness of equilibrium <strong>in</strong><br />
<strong>the</strong> overall game, we need not consider variation <strong>in</strong> defensive <strong>in</strong>tensity from possession to<br />
possession. As such we will omit notation for defensive <strong>in</strong>tensity and <strong>the</strong> impact of teammates<br />
on <strong>the</strong> notation below. We do observe variation <strong>in</strong> player shoot<strong>in</strong>g decisions across <strong>the</strong> shot<br />
clock and we will proceed by formally tak<strong>in</strong>g this variation as exogenous.<br />
Assumption ID: A player’s opportunity distribution is <strong>in</strong>variant to <strong>the</strong> particular<br />
period of <strong>the</strong> shot clock.<br />
Fi,t,Op/i,dp,i<br />
L<br />
= Fi,s,Op/i,dp,i<br />
L<br />
≡ Fi,Op/i,dp,i<br />
Under this assumption, <strong>the</strong> reduced form quantities {up,i,t, ep,i,t} conta<strong>in</strong> structural <strong>in</strong>for-<br />
mation about our time-<strong>in</strong>variant opportunity distribution.<br />
ep,i,t =<br />
∞<br />
F −1<br />
η p,i (1−up,i,t) ηp,idFηp,i<br />
up,i,t<br />
(21)<br />
= E(ηp,i|ηp,i > F −1<br />
ηp,i (1 − up,i,t)) (22)<br />
If for some player i and ∀u ∈ [0, 1], we observed pairs {e(u), u} <strong>the</strong>n we would achieve<br />
full, non-parametric identification of <strong>the</strong> opportunity distribution and chosen cut-thresholds.<br />
27
de<br />
du<br />
−1 (Fηp,i (1 − u) − e(u))<br />
=<br />
u<br />
c(u) = F −1<br />
de<br />
ηp,i (1 − u) = e(u) + u<br />
du<br />
Given observation <strong>in</strong> a f<strong>in</strong>ite number (T ) of discrete shot clock values, we are only able<br />
to identify sub-blocks of <strong>the</strong> opportunity distribution and provide set identification of cut-<br />
thresholds.<br />
UL(t) ≡ {s ∈ T T : up,i,s < up,i,t}<br />
UH(t) ≡ {s ∈ T T : up,i,s > up,i,t}<br />
Then ∀l ∈ UL(t), h ∈ UH(t), observables place <strong>the</strong> follow<strong>in</strong>g restrictions on cut-thresholds.<br />
ep,i,tup,i,t − ep,i,lup,i,l<br />
up,i,t − up,i,l<br />
ep,i,hup,i,h − ep,i,tup,i,t<br />
up,i,h − up,i,t<br />
=<br />
=<br />
F −1<br />
η p,i (1−up,i,l)<br />
F −1<br />
η p,i (1−up,i,t) ηp,idFηp,i<br />
up,i,t − up,i,l<br />
F −1<br />
η (1−up,i,t)<br />
p,i<br />
F −1<br />
η p,i (1−up,i,h) ηp,idFηp,i<br />
up,i,h − up,i,t<br />
≤ F −1<br />
ηp,i (1 − up,i,t) = cp,i,t<br />
≥ F −1<br />
ηp,i (1 − up,i,t) = cp,i,t<br />
It is important to notice that <strong>the</strong> failure of full identification here has two different<br />
sources. The first is that no <strong>NBA</strong> basketball player is ever observed to shoot with a hazzard<br />
rate <strong>in</strong> excess of 35% or below .2%. Thus we have no <strong>in</strong>formation whatsoever about what<br />
player’s opportunity distributions look like outside <strong>the</strong>se conf<strong>in</strong>es. This is not a mean<strong>in</strong>g-<br />
ful failure. The second problem is that we do not observe cont<strong>in</strong>uous variation <strong>in</strong> player<br />
shoot<strong>in</strong>g behavior, but ra<strong>the</strong>r discrete jumps correspond<strong>in</strong>g to <strong>the</strong> lump<strong>in</strong>ess of our shot<br />
clock observation. Thus we cannot p<strong>in</strong> down exactly what cut-threshold corresponds to an<br />
observed hazard rate, because we do not have local knowledge about <strong>the</strong> behavior of <strong>the</strong><br />
opportunity distribution. In practice, 13 different periods of half-court offense is still plenty<br />
of <strong>in</strong>formation about <strong>the</strong> underly<strong>in</strong>g structure. One way to th<strong>in</strong>k about this is that as long<br />
as players adopt different hazard rates <strong>in</strong> each of <strong>the</strong> 13 periods 16 , <strong>the</strong>n we could identify a<br />
parametric model of <strong>the</strong> opportunity distribution that had up to 13 parameters.<br />
Alternatively, we will actually compute <strong>the</strong> bounds by (??) and (??) for an example<br />
16 As we will address <strong>in</strong> our Results section, <strong>the</strong>y almost exclusively do.<br />
28<br />
(23)<br />
(24)<br />
(25)<br />
(26)
player and show that identified sets are generally quite narrow. But that variability <strong>in</strong><br />
<strong>the</strong> estimation of <strong>the</strong> upper and lower bounds of <strong>the</strong>se sets tends to make <strong>the</strong> fully non-<br />
parametric approach un<strong>in</strong>formative. Below, we consider <strong>the</strong> aggregation of all <strong>NBA</strong> centers<br />
as one composite player <strong>in</strong> an attempt to m<strong>in</strong>imize <strong>the</strong> impact of sample variance on our<br />
estimator. We will comb<strong>in</strong>e <strong>the</strong> <strong>in</strong>formation from various bounds <strong>in</strong> (26) and (25) <strong>in</strong>to a<br />
s<strong>in</strong>gle MSE m<strong>in</strong>imiz<strong>in</strong>g estimator of a composite lower and upper bound 17 .<br />
Figure 2: Diagram of <strong>the</strong> Aggregate <strong>NBA</strong> Center’s preformance <strong>in</strong> our data set. Blue dots<br />
<strong>in</strong>dicate observed comb<strong>in</strong>ations of efficiency and usage <strong>in</strong> each period of <strong>the</strong> shot clock. The<br />
o<strong>the</strong>r l<strong>in</strong>es <strong>in</strong>dicates estimates of <strong>the</strong> area of set identification via (26) and (25). Even for this<br />
aggregate player, nonparamtric estimators produce very weak po<strong>in</strong>twise confidence <strong>in</strong>tervals<br />
on cut-thresholds.<br />
5.3 A Parsimonious Parametric Model<br />
The conditions of <strong>the</strong> previous section are quite permissive of potentially suboptimal play<br />
for two reasons: 1) <strong>the</strong>y provide no tests for undershoot<strong>in</strong>g 2) <strong>the</strong>y give no <strong>in</strong>sight toward<br />
allocative efficiency. Fur<strong>the</strong>rmore, <strong>the</strong>y offer no convenient way to characterize <strong>the</strong> overall<br />
distribution from which a player realizes scor<strong>in</strong>g opportunities. In this section we address<br />
<strong>the</strong>se difficulties. In order to do so w must make a somewhat restrictive assumption. Namely<br />
that ∀t ∈ T player i draws his opportunities from a common distribution. This will allow<br />
us to compare player i’s performance at <strong>the</strong> beg<strong>in</strong>n<strong>in</strong>g of <strong>the</strong> shot clock (when he will rarely<br />
shoot) to <strong>the</strong> end of <strong>the</strong> shot clock (when he will have to shoot a much higher fraction of<br />
17 In <strong>the</strong> selection of an optimal estimator, <strong>the</strong> <strong>in</strong>creased bias of looser bounds is approximated by imag<strong>in</strong><strong>in</strong>g<br />
<strong>the</strong> shot opportunity distribution is uniform with density .25.<br />
29
<strong>the</strong> time). Intuitively, we are us<strong>in</strong>g <strong>the</strong> shot clock as an <strong>in</strong>strument to identify player i’s<br />
opportunity distribution and implied usage curve.<br />
We model player i’s scor<strong>in</strong>g opportunities as drawn from a uniform distribution along<br />
<strong>the</strong> <strong>in</strong>terval [Bi, Ai]. Additionally, <strong>the</strong> player selects a shot <strong>in</strong> period t if his draw lies <strong>in</strong> <strong>the</strong><br />
<strong>in</strong>terval of [ci,t, Ai]. For <strong>the</strong> vast majority of players, Bi will take a negative value. This does<br />
not mean that players have opportunities to lose po<strong>in</strong>ts, but merely reflects <strong>the</strong> fact that<br />
<strong>the</strong> vast majority of players do not realize a good scor<strong>in</strong>g opportunity <strong>in</strong> most periods of <strong>the</strong><br />
shot clock. As long as ci,t > 0 <strong>the</strong> part of our <strong>the</strong>oretical uniform opportunity distribution<br />
that lies below zero is irrelevant.<br />
Conditional on <strong>the</strong> parameters θi = {Ai, Bi, {ci,t}t∈T }, it is straightforward that:<br />
ei,t = Ai + ci,t<br />
2<br />
P rob({Ei,t, Ui,t}t∈T |θi, Ni,t) = <br />
φ( <br />
t∈T<br />
, ui,t = Ai − ci,t<br />
,<br />
Ai − Bi<br />
dei,t<br />
dui,t<br />
Ui,t − Ai−ci,t<br />
Ai−Bi<br />
Ni,t Ai−ci,t<br />
Ai−Bi<br />
(1 − Ai−ci,t<br />
Ai−Bi )<br />
= Ai − Bi<br />
)φ( Ei,t − Ai+ci,t<br />
2<br />
σi √<br />
Ni,t<br />
We achieve parametric identification because <strong>the</strong> likelihood matrix is non-s<strong>in</strong>gular for all<br />
players who choose at least some variation <strong>in</strong> cutoff levels across periods of <strong>the</strong> shot clock.<br />
This should not be a surprise, if a player did <strong>the</strong> same th<strong>in</strong>g <strong>in</strong> every period of <strong>the</strong> shot clock,<br />
our shot clock <strong>in</strong>strument is irrelevant and we have on hope of learn<strong>in</strong>g about his tradeoff<br />
between usage and efficiency. It is doubtful that any <strong>NBA</strong> player’s shoot<strong>in</strong>g behavior is<br />
truly <strong>in</strong>variant to pressure from <strong>the</strong> shot clock. However, lack of identification at this po<strong>in</strong>t<br />
has <strong>the</strong> potential to create distortions <strong>in</strong> standard Wald statistics if <strong>the</strong> truth is <strong>in</strong> a small<br />
enough neighborhood of <strong>the</strong> unidentified region. We do not believe this difficulty is relevant<br />
to <strong>the</strong> vast majority of players <strong>in</strong> our data and <strong>in</strong> this version we will not explicitly address<br />
it. But we do note that recent advances <strong>in</strong> <strong>the</strong> econometric literature (Dufour et al. 2010;<br />
Andrews et al. 2010) may enable us to better understand and be robust to any nonstandard<br />
properties of our estimators.<br />
5.4 Measurement Error <strong>in</strong> <strong>the</strong> <strong>Shot</strong> <strong>Clock</strong><br />
Our identification argument <strong>in</strong> <strong>the</strong> previous section relies crucially on our ability to <strong>in</strong>fer <strong>the</strong><br />
correct value of <strong>the</strong> shot clock that a player observed when mak<strong>in</strong>g his shoot<strong>in</strong>g decisions.<br />
There are precise guidel<strong>in</strong>es for when <strong>the</strong> shot clock should be reset <strong>in</strong> <strong>the</strong> <strong>NBA</strong>(?) and we<br />
were able to estimate <strong>the</strong> true shot clock time by track<strong>in</strong>g <strong>the</strong> value of <strong>the</strong> game clock at<br />
which all shot clock relevant events occur. However, it is a typical occurence that <strong>the</strong> shot<br />
30<br />
)
clock operator and <strong>the</strong> person who tracks <strong>the</strong> play by play data might slightly disagree on<br />
exactly when a possessions starts. Thus our shot clock measurement may be off by as much<br />
as 5 or 6 seconds on some possessions. We were orig<strong>in</strong>ally unaware of <strong>the</strong>se complications<br />
and would like to thank participants at <strong>the</strong> Sloan Sport’s Conference 18 for mak<strong>in</strong>g us aware<br />
of this obstacle.<br />
An easy way to demonstrate <strong>the</strong> imprecision of our shot clock estimator is to consider<br />
<strong>the</strong> distribution of our estimates of <strong>the</strong> value of <strong>the</strong> shot clock when <strong>the</strong> offensive team<br />
experiences a shot clock violation. A shot clock violation only occurs when <strong>the</strong> true value<br />
of <strong>the</strong> shot clock is exactly zero, as you can see our estimates are generally centered around<br />
zero, but somewhat imprecise.<br />
Figure 3: The distribution of error for shot clock violations varies greatly across different<br />
types of possessions.<br />
As seen above, <strong>the</strong> distribution of shot clock error varies greatly depend<strong>in</strong>g on how <strong>the</strong><br />
possession starts. When <strong>the</strong> possession starts off a dead ball (<strong>the</strong> o<strong>the</strong>r team throws it out<br />
of bounds or makes <strong>the</strong> second of two free throws) <strong>the</strong>ir is very little room for error by<br />
<strong>the</strong> shot clock operator. Similarly when <strong>the</strong> possession starts off a steal, <strong>the</strong>ir is typically<br />
only a small lag before <strong>the</strong> shot clock operator notices this event and resets <strong>the</strong> shot clock,<br />
lead<strong>in</strong>g to only slightly more measurement error. A third way a possession can start is with<br />
a defensive rebound follow<strong>in</strong>g a missed field goal by <strong>the</strong> o<strong>the</strong>r team. Battles for rebounds<br />
are often harshly contested and it is not always clear when one team has secured <strong>the</strong> ball. It<br />
is entirely plausible that <strong>the</strong> shot clock operator and <strong>the</strong> person keep<strong>in</strong>g <strong>the</strong> game log could<br />
come to very different conclusions here and thus we experience significantly more shot clock<br />
measurement error on <strong>the</strong>se possessions. F<strong>in</strong>ally, it is also possible (and very common) for<br />
18 Allocative and Dynamic Efficiency <strong>in</strong> <strong>NBA</strong> Decision Mak<strong>in</strong>g was <strong>the</strong> version of our paper presented <strong>the</strong>re,<br />
account<strong>in</strong>g for measurement error <strong>in</strong> <strong>the</strong> shot clock has slightly changed <strong>the</strong> overall flavor of our results <strong>in</strong><br />
that we generally f<strong>in</strong>d more overshoot<strong>in</strong>g and less undershoot<strong>in</strong>g now.<br />
31
<strong>the</strong> possession to start after a made basket by <strong>the</strong> o<strong>the</strong>r team. On <strong>the</strong>se possessions, <strong>the</strong> <strong>NBA</strong><br />
game clock cont<strong>in</strong>ues to run after a basket is scored and before <strong>the</strong> next possession starts.<br />
However, <strong>the</strong> offense’s shot clock does not beg<strong>in</strong> to run until <strong>the</strong>y <strong>in</strong>bound <strong>the</strong> basketball.<br />
Players often lolly-gag or exchange brief communication before choos<strong>in</strong>g to <strong>in</strong>bound <strong>the</strong> ball<br />
lead<strong>in</strong>g to our perception that <strong>the</strong> shot clock should start significantly before it actually<br />
does. As seen <strong>in</strong> above graph, average measurement error on <strong>the</strong>se possessions is strongly<br />
negative (we th<strong>in</strong>k player’s are out of time when <strong>the</strong>y are not) and has significant variability.<br />
Now that we understand <strong>the</strong> source of our measurement error, we need to consider how<br />
we can overcome its bedevill<strong>in</strong>g effect on our identification. For purposes of notation, let mp<br />
be a discrete <strong>in</strong>dicator for <strong>the</strong> mechanism by which <strong>the</strong> possessions starts and let zp,t denote<br />
<strong>the</strong> observed value of <strong>the</strong> shot clock on possession p when <strong>the</strong> ’true’ value of <strong>the</strong> shot clock<br />
is t.<br />
zp,t = t + ɛp,t<br />
In order to restore identification of our model, we will need to know <strong>the</strong> distribution of<br />
’true’ shot clock values conditional on all observables. That is we want, P (t|z, m) which by<br />
application of Bayes’ rule we may write this down as:<br />
P (t|m)<br />
P (t|z, m) = P (z|t, m)<br />
P (z|m)<br />
We will do this <strong>in</strong> pieces. The denom<strong>in</strong>ator (P (z|m)) is <strong>the</strong> unconditional distribution<br />
of observed shot clock values when possessions are used, it is directly observable from <strong>the</strong><br />
data. P (t|m) is <strong>the</strong> unconditional distribution of possession use times and P (z|t, m) is <strong>the</strong><br />
conditional distriubtion of shot clock error.<br />
Learn<strong>in</strong>g about <strong>the</strong>se two terms requires <strong>the</strong> follow<strong>in</strong>g assumption on (27), which is<br />
justified by <strong>the</strong> above discussion.<br />
Assumption CME: <strong>Shot</strong> clock measurement error is classical. It is entirely caused by<br />
mismatches between <strong>the</strong> game and shot clock that occur at <strong>the</strong> beg<strong>in</strong>n<strong>in</strong>g of <strong>the</strong> possession.<br />
The distribution of this error may vary across <strong>the</strong> four different ways possessions orig<strong>in</strong>ate,<br />
but is <strong>in</strong>variant to <strong>the</strong> true value of <strong>the</strong> shot clock. This is succ<strong>in</strong>ctly expressed <strong>in</strong> <strong>the</strong><br />
equations below.<br />
∀t, p, ɛp,t = ɛp<br />
ɛp ∼ Γmp<br />
32<br />
(27)<br />
(28)<br />
(29)<br />
(30)
This assumption yields constructive identification of our desired terms.<br />
P(z|t, m) = Γm(z − t) = P (z − t|t = 0, m) (31)<br />
P(t|m) × P (z − t|m) = P (z|m) (32)<br />
t|m ⊥ (z − t)|m (33)<br />
Where <strong>the</strong> × <strong>in</strong> (32) <strong>in</strong>dicates <strong>the</strong> convolution operator. Then given (33), P (t|m) is <strong>the</strong><br />
solution to a straightforward deconvolution equation.<br />
Comb<strong>in</strong><strong>in</strong>g all <strong>the</strong>se pieces of <strong>in</strong>formation we identify our target density: P (t|z, m). Com-<br />
b<strong>in</strong><strong>in</strong>g this <strong>in</strong>formation, with <strong>the</strong> model <strong>in</strong> Section 5.3 we achieve a likelihood model that is<br />
robust to measurement error.<br />
6 Estimation<br />
L ME (Up, Pp|z, m, θ) = <br />
P (t|z, m)L(Up, Pp|t, m, θ) (34)<br />
t<br />
For purposes of tractability we impose a little bit of structure on Φ allow<strong>in</strong>g for a seperable<br />
penalty function of defensive <strong>in</strong>terference.<br />
Φi,Op/i,dp,i,t = Φi,Op/i,t − hilog(dp,i) (35)<br />
6.1 Parametric Uniform <strong>Shot</strong> Distributions<br />
The conditions of <strong>the</strong> previous section are quite permissive of potentially suboptimal play<br />
for two reasons: 1) <strong>the</strong>y provide no tests for undershoot<strong>in</strong>g 2) <strong>the</strong>y give no <strong>in</strong>sight toward<br />
allocative efficiency. Fur<strong>the</strong>rmore, <strong>the</strong>y offer no convenient way to characterize <strong>the</strong> overall<br />
distribution from which a player realizes scor<strong>in</strong>g opportunities. In this section we address<br />
<strong>the</strong>se difficulties. In order to do so w must make a somewhat restrictive assumption. Namely<br />
that ∀t ∈ T player i draws his opportunities from a common distribution. This will allow<br />
us to compare player i’s performance at <strong>the</strong> beg<strong>in</strong>n<strong>in</strong>g of <strong>the</strong> shot clock (when he will rarely<br />
shoot) to <strong>the</strong> end of <strong>the</strong> shot clock (when he will have to shoot a much higher fraction of<br />
<strong>the</strong> time). Intuitively, we are us<strong>in</strong>g <strong>the</strong> shot clock as an <strong>in</strong>strument to identify player i’s<br />
opportunity distribution and implied usage curve.<br />
We model player i’s scor<strong>in</strong>g opportunities as drawn from a uniform distribution along<br />
<strong>the</strong> <strong>in</strong>terval [Bi, Ai]. Additionally, <strong>the</strong> player selects a shot <strong>in</strong> period t if his draw lies <strong>in</strong> <strong>the</strong><br />
33
<strong>in</strong>terval of [ci,t, Ai]. For <strong>the</strong> vast majority of players, Bi will take a negative value. This does<br />
not mean that players have opportunities to lose po<strong>in</strong>ts, but merely reflects <strong>the</strong> fact that<br />
<strong>the</strong> vast majority of players do not realize a good scor<strong>in</strong>g opportunity <strong>in</strong> most periods of <strong>the</strong><br />
shot clock. As long as ci,t > 0 <strong>the</strong> part of our <strong>the</strong>oretical uniform opportunity distribution<br />
that lies below zero is irrelevant.<br />
Conditional on <strong>the</strong> parameters θi = {Ai, Bi, {ci,t}t∈T }, it is straightforward that:<br />
ei,t = Ai + ci,t<br />
2<br />
P rob({Ei,t, Ui,t}t∈T |θi, Ni,t) = <br />
φ( <br />
t∈T<br />
, ui,t = Ai − ci,t<br />
,<br />
Ai − Bi<br />
dei,t<br />
dui,t<br />
Ui,t − Ai−ci,t<br />
Ai−Bi<br />
Ni,t Ai−ci,t<br />
Ai−Bi<br />
(1 − Ai−ci,t<br />
Ai−Bi )<br />
= Ai − Bi<br />
)φ( Ei,t − Ai+ci,t<br />
2<br />
σi √<br />
Ni,t<br />
We achieve parametric identification because <strong>the</strong> likelihood matrix is non-s<strong>in</strong>gular for all<br />
players who choose at least some variation <strong>in</strong> cutoff levels across periods of <strong>the</strong> shot clock.<br />
This should not be a surprise, if a player did <strong>the</strong> same th<strong>in</strong>g <strong>in</strong> every period of <strong>the</strong> shot clock,<br />
our shot clock <strong>in</strong>strument is irrelevant and we have on hope of learn<strong>in</strong>g about his tradeoff<br />
between usage and efficiency. It is doubtful that any <strong>NBA</strong> player’s shoot<strong>in</strong>g behavior is<br />
truly <strong>in</strong>variant to pressure from <strong>the</strong> shot clock. However, lack of identification at this po<strong>in</strong>t<br />
has <strong>the</strong> potential to create distortions <strong>in</strong> standard Wald statistics if <strong>the</strong> truth is <strong>in</strong> a small<br />
enough neighborhood of <strong>the</strong> unidentified region. We do not believe this difficulty is relevant<br />
to <strong>the</strong> vast majority of players <strong>in</strong> our data and <strong>in</strong> this version we will not explicitly address<br />
it. But we do note that recent advances <strong>in</strong> <strong>the</strong> econometric literature (Dufour et al. 2010;<br />
Andrews et al. 2010) may enable us to better understand and be robust to any nonstandard<br />
properties of our estimators.<br />
Additionally, our likelihood equation is dependent upon <strong>the</strong> normality of <strong>the</strong> distribution<br />
of po<strong>in</strong>ts from an <strong>in</strong>dividual shot. In small samples, this is trivially <strong>in</strong>correct as <strong>the</strong> actual<br />
distribution of po<strong>in</strong>ts is discrete and slightly skew. Additionally, we have no reason to<br />
believe that our uniform specification is exactly correct. As such we shall not assume that<br />
<strong>the</strong> true data generat<strong>in</strong>g process falls with<strong>in</strong> our model and apply consistently estimated<br />
covariance matrices from White (1982). Fur<strong>the</strong>r we shall only perform Maximum Likelihood<br />
Estimation (MLE) for each player with at least 15 used possessions <strong>in</strong> every relevant period<br />
of <strong>the</strong> shot clock. Despite <strong>the</strong> fact that we are fitt<strong>in</strong>g a uniform distribution, <strong>the</strong> support<br />
of observed efficiency and usage rates over any f<strong>in</strong>ite sample does not change so we do not<br />
have a regularity problem and are able to calculate standard asymptotic variances from <strong>the</strong><br />
likelihood matrix.<br />
34<br />
)
Figure 1 provides <strong>the</strong> aggregate results of this estimation procedure. We see that <strong>in</strong>deed<br />
<strong>NBA</strong> players use a monotonically decl<strong>in</strong><strong>in</strong>g cut threshold consistent with <strong>the</strong> predictions of<br />
an optimal stopp<strong>in</strong>g problem with f<strong>in</strong>ite periods. More impressive is <strong>the</strong> fact that <strong>the</strong> cut-<br />
thresholds are nearly identical to <strong>the</strong> cont<strong>in</strong>uation values of <strong>the</strong> possession and <strong>the</strong> functions<br />
have <strong>the</strong> same shape. We present <strong>the</strong> results <strong>in</strong> two panels to enhance <strong>the</strong> contrast of <strong>the</strong><br />
slope, while still show<strong>in</strong>g it for all periods of <strong>the</strong> shot clock. Overall, Figure 1 is strong<br />
evidence <strong>in</strong> favor of near optimal play. <strong>NBA</strong> players appear to be well-tuned to <strong>the</strong> cont<strong>in</strong>-<br />
uation value of <strong>the</strong> possession and adjust <strong>the</strong>ir shot choice to reflect it. This is precisely <strong>the</strong><br />
mechanics of optimal stopp<strong>in</strong>g. Not only do <strong>the</strong>y get <strong>the</strong> mechanics right, but <strong>the</strong> rate at<br />
which <strong>the</strong> players lower <strong>the</strong>ir cut threshold matches <strong>the</strong> cont<strong>in</strong>uation value nearly exactly!<br />
We do note, however, that <strong>in</strong> Figure 1 <strong>the</strong> cut-threshold does lie slightly above <strong>the</strong><br />
cont<strong>in</strong>uation value, which is evidence that undershoot<strong>in</strong>g is more common that overshoot<strong>in</strong>g.<br />
To extend <strong>the</strong> analysis, we now exam<strong>in</strong>e which player’s tend to overshoot or undershoot on<br />
average. We take as a null hypo<strong>the</strong>sis that each player is dynamically efficient on average.<br />
Namely that for each player i, <br />
t∈{2,..,17} ci,t = <br />
t∈{2,..,17} eO,t∗. To maximize power to detect<br />
deviation, we test this hypo<strong>the</strong>sis with a weights <strong>in</strong>verse to <strong>the</strong> variance of our estimated<br />
cutoffs. Figure 2 provides a histogram of <strong>the</strong> result<strong>in</strong>g t-statistics. A negative t-statistic<br />
<strong>in</strong>dicates overshoot<strong>in</strong>g, a positive t-statistic <strong>in</strong>dicates undershoot<strong>in</strong>g.<br />
The results aga<strong>in</strong> <strong>in</strong>dicate that undershoot<strong>in</strong>g is much more common than overshoot<strong>in</strong>g.<br />
In fact, only 5 players are found to be significant overshooters — less than we would expect<br />
to f<strong>in</strong>d by chance alone (although if we take <strong>the</strong> mean to be 1, not 0, <strong>the</strong> evidence <strong>the</strong>se<br />
players overshoot streng<strong>the</strong>ns considerably. 19 In l<strong>in</strong>e with this reason<strong>in</strong>g, <strong>the</strong> distribution<br />
appears standard normal but shifted over about 1 (<strong>the</strong> mean 0.98) Most players appear to be<br />
optimiz<strong>in</strong>g and mistakes tend to come <strong>in</strong> <strong>the</strong> form of undershoot<strong>in</strong>g early <strong>in</strong> <strong>the</strong> shot clock.<br />
In this sense, some players wait too long to shoot or do not expend maximum effort on each<br />
possession. In contrast, lab subjects tend to pull <strong>the</strong> trigger too early, typically through<br />
<strong>the</strong> use of a fixed threshold (Lee et al., 2004). We also compute <strong>the</strong> loss <strong>in</strong> surplus due to<br />
sub-optimal shoot<strong>in</strong>g decisions (<strong>in</strong>tuitively <strong>in</strong>tegrat<strong>in</strong>g between <strong>the</strong> two l<strong>in</strong>es <strong>in</strong> Figure 1).<br />
The median value of DWL across players is 4%, consistent with nearly optimal play.<br />
In <strong>the</strong> discussion section we exam<strong>in</strong>e how much of a role effort conservation plays <strong>in</strong> this<br />
f<strong>in</strong>d<strong>in</strong>g. While we cannot observe effort directly, we can observe which type of players tend<br />
to undershoot. We f<strong>in</strong>d that under-shooters are frequently highly paid “star” players. Our<br />
<strong>in</strong>tuition is that some of undershoot<strong>in</strong>g is driven by effort conservation by star players who<br />
play relatively more m<strong>in</strong>utes-per-game.<br />
19 Lamar Odom, Monta Ellis, Rafer Alston, Russell Westbrook and Tyrus Thomas are <strong>the</strong> guilty parties.<br />
35
6.2 Allocative Efficiency<br />
Allocative Efficiency is <strong>the</strong> hypo<strong>the</strong>sis that <strong>the</strong> players on <strong>the</strong> court, (I) use <strong>the</strong> same cut-<br />
po<strong>in</strong>t <strong>in</strong> all periods of <strong>the</strong> shot clock. In our estimation we allow <strong>the</strong> <strong>the</strong> set I to be an<br />
entire five man l<strong>in</strong>eup or a three or four man ”core” that more frequently shares <strong>the</strong> court.<br />
The concept of cores is convenient to improve power <strong>in</strong> estimation Because our estimation<br />
is more precise for <strong>the</strong>se periods, we will focus on cut-po<strong>in</strong>ts <strong>in</strong> <strong>the</strong> first ten seconds of <strong>the</strong><br />
shot clock (T = {9...18}). Let c be <strong>the</strong> T I × 1 vector of relevant cut po<strong>in</strong>ts, sorted first<br />
by period of <strong>the</strong> shot clock. Based on our parametric procedure we have:<br />
ĉ ∼ N(c, Vc<br />
N )<br />
We def<strong>in</strong>e <strong>the</strong> ”true” deviation from Allocative Efficiency as spread.<br />
SI,T = <br />
t∈T i∈I (ci,t − ¯ct) 2 = (c − ¯c) ′ I(c − ¯c) = z ′ Iz<br />
Where ci,t is <strong>the</strong> cutoff chosen by player i <strong>in</strong> period t and ¯ct is <strong>the</strong> average over <strong>the</strong> five<br />
teammates of <strong>the</strong> cutoff chosen <strong>in</strong> period t and ¯c is <strong>the</strong> appropriate correspond<strong>in</strong>g T I×1<br />
vector of averages.<br />
Note that, if we def<strong>in</strong>e M = IT ⊗ (I5 − J5 ), we may also simply write z = Mc and<br />
5<br />
S = c ′ M ′ Mc = c ′ Mc.<br />
The ”seem<strong>in</strong>gly” natural empirical analog to S is:<br />
ˆS = 5 t∈T i=1 ( ˆ ci,t − ˆ¯ct) 2 = ˆz ′ I ˆz<br />
However, a little bit of calculation reveals that this measure is biased upwards, especially<br />
for small samples.<br />
E(ˆz) = c − ¯c<br />
V (ˆz) ≡ Vz = MVcM<br />
N<br />
Thus: E( ˆ S) = E(z2 i,t) = E( zi,t) ˆ 2 + V ( zi,t) ˆ = S + diagVz = S + diag(MVcM)<br />
N<br />
36
turns out to be upward biased and especially so for small n. In order to correct for this,<br />
we essentially subtract off <strong>the</strong> bias and def<strong>in</strong>e:<br />
ˆS ∗ = ˆ S − diag(M ˆ VcM)<br />
N<br />
We use this as our measure of dispersion <strong>in</strong> <strong>the</strong> marg<strong>in</strong>al output of players shar<strong>in</strong>g <strong>the</strong><br />
court and refer to simply as spread. It turns out that <strong>the</strong> same 5-man l<strong>in</strong>e-ups share <strong>the</strong><br />
court relatively <strong>in</strong> frequently. We thus need to def<strong>in</strong>e <strong>the</strong> concept of “4-man core.” A 4-man<br />
core <strong>in</strong>cludes all l<strong>in</strong>e-ups that share <strong>the</strong> same 4 players and and <strong>the</strong> fifth player is occurs less<br />
frequently than <strong>the</strong> core-4. An example is if 4 starters typically share <strong>the</strong> court with 1 of 3<br />
back-ups (<strong>the</strong> fifth player). In this case, each back-up plays less than <strong>the</strong> starters with <strong>the</strong><br />
l<strong>in</strong>e-up, so is not <strong>in</strong> <strong>the</strong> 4-core. We <strong>the</strong>n compute spread as def<strong>in</strong>ed above.<br />
Figure 3 presents <strong>the</strong> distribution of spread for <strong>the</strong> 246 most common 4-man cores <strong>in</strong><br />
our data. The median is exceed<strong>in</strong>gly small; it comes <strong>in</strong> at 0.036 (<strong>the</strong> mean is slightly larger<br />
at 0.06). We also see a large spike at 0 <strong>in</strong> <strong>the</strong> distribution. In most l<strong>in</strong>e-ups, <strong>the</strong> players<br />
show near equal marg<strong>in</strong>al efficiencies at each period of <strong>the</strong> shot clock. Aga<strong>in</strong>, this is strong<br />
evidence of near optimal play, but aga<strong>in</strong> play is not perfect. Players seems to be shar<strong>in</strong>g <strong>the</strong><br />
ball <strong>in</strong>credibly well, but still about 10% of l<strong>in</strong>e-ups are estimate to have spread <strong>in</strong> excess of<br />
0.25. These l<strong>in</strong>e-ups show clear room for improvement <strong>in</strong> terms of ball allocation.<br />
In <strong>the</strong> Appendix we present <strong>the</strong> distribution of spread for 3-man cores. The results<br />
confirm <strong>the</strong> results presented <strong>in</strong> Figure 3 and <strong>in</strong> fact <strong>the</strong> distribution is even more closely<br />
centered around 0. This makes sense, <strong>in</strong> a 3-man core we compute spread over 3 players,<br />
<strong>in</strong>stead of 4, that are more familiar with each o<strong>the</strong>r and have more similar characteristics.<br />
In <strong>the</strong> discussion section we delve <strong>in</strong>to how spread correlates with features of <strong>the</strong> l<strong>in</strong>e-up.<br />
7 Deeper analysis and discussion<br />
The analysis <strong>in</strong> <strong>the</strong> preced<strong>in</strong>g section gave sufficient statistics dynamic and allocative effi-<br />
ciency. We now l<strong>in</strong>k <strong>the</strong>se measures to features of <strong>the</strong> l<strong>in</strong>e-up such as aggregate efficiency,<br />
experience and features of <strong>the</strong> players.<br />
7.1 L<strong>in</strong>e-up performance, l<strong>in</strong>e-up attributes and allocative effi-<br />
ciency<br />
In this section we exam<strong>in</strong>e how l<strong>in</strong>e-up performance and o<strong>the</strong>r attributes of <strong>the</strong> l<strong>in</strong>e-up cor-<br />
relate with adherence to allocative efficiency. We view <strong>the</strong>se regressions and figures as at <strong>the</strong><br />
37
very least <strong>in</strong>formative correlations. Clearly we, as academic researchers, cannot exogenously<br />
improve adherence to optimality <strong>in</strong> order to gage <strong>the</strong> impact on performance. Similarly we<br />
cannot exogenously impose an a more experienced l<strong>in</strong>e-up and measure <strong>the</strong> impact on shot<br />
allocation. Indeed our estimates <strong>the</strong>mselves provide answers to <strong>the</strong> question “how much meat<br />
is left on <strong>the</strong> table” (our estimates of DWL are small). One potential endogeneity problem<br />
is that a coach might cont<strong>in</strong>ue to play a high output l<strong>in</strong>e-up (driven by talent), despite<br />
relatively poor decision mak<strong>in</strong>g, because <strong>the</strong> results are still better than experiment<strong>in</strong>g with<br />
someth<strong>in</strong>g new. This would tend to dampen <strong>the</strong> impact of offensive output on adherence to<br />
optimality. 20<br />
Figure 4 presents a scatter plot of 4-man core spread on average output per possession.<br />
There is a detectable, but not overwhelm<strong>in</strong>g, pattern of more productive l<strong>in</strong>e-ups show<strong>in</strong>g<br />
lower spread. Table 2 demonstrates that this relationship is significant at <strong>the</strong> 0.10 level<br />
us<strong>in</strong>g a two-tailed test. Given that our hypo<strong>the</strong>sis is <strong>in</strong> fact one-tailed, this significance<br />
level can be adjusted to 0.05. Table 2 also gives estimates of o<strong>the</strong>r correlates of spread. N<br />
gives <strong>the</strong> number of possessions <strong>the</strong> core shared <strong>in</strong> <strong>the</strong> sample. The impact of N is weak<br />
and positive. This is not necessarily evidence of a negative impact of repetition because of<br />
numerous endogeneity issues based on coaches’ decisions. Column (1) <strong>in</strong>cludes <strong>the</strong> average<br />
salary of <strong>the</strong> l<strong>in</strong>e-up <strong>in</strong> logs. Column (2) <strong>in</strong>cludes <strong>the</strong> simple arithmetic mean. For both<br />
cases, higher paid l<strong>in</strong>e-ups perform significantly better <strong>in</strong> terms of spread. Higher salary<br />
<strong>in</strong>equality, as measured through <strong>the</strong> standard deviation, is associated with larger values<br />
of spread. This result is potentially of <strong>in</strong>terest to labor economists concerned with salary<br />
equality and production. “Experience” gives <strong>the</strong> number of years played <strong>in</strong> <strong>the</strong> <strong>NBA</strong> at <strong>the</strong><br />
start of 2006. The mean of play<strong>in</strong>g experience across <strong>the</strong> l<strong>in</strong>e-up does not appear to impact<br />
adherence to allocative efficiency. The standard deviation of experience seems to have a<br />
slight negative impact on spread.<br />
7.2 Aggregate player characteristics and efficiency<br />
In Table 3, we regress <strong>the</strong> t-stat from adherence to dynamic optimality on <strong>in</strong>dividual player<br />
characteristics. Consistent with <strong>the</strong> reason<strong>in</strong>g offered <strong>in</strong> Section 3.3, salary is positively<br />
related to <strong>the</strong> t-stat. Recall higher t-stats <strong>in</strong>dicate under-shoot<strong>in</strong>g. Indeed <strong>the</strong> league’s star<br />
players such as Chris Paul, Lebron James and Kobe Bryant have high t’s. Table 3 shows<br />
that this <strong>the</strong> case generally, salary is highly significant <strong>in</strong> <strong>the</strong> both <strong>the</strong> l<strong>in</strong>ear specification<br />
(1) and l<strong>in</strong>ear-log specification (2).<br />
20 One might expect salary to be a nice control for l<strong>in</strong>e-up output, but unfortunately <strong>the</strong> correlation is<br />
not strong (regressions available from <strong>the</strong> authors). Essentially <strong>the</strong>re are just too many highly paid lousy<br />
38
Table 4: Impact of l<strong>in</strong>e-up characteristics on allocative efficiency (spread).<br />
(1) (2)<br />
Output -0.214* -0.205<br />
(0.124) (0.124)<br />
N 4.41e-05*** 4.88e-05***<br />
(1.61e-05) (1.69e-05)<br />
ln(Mean Salary) -0.0862**<br />
(0.0425)<br />
Mean Salary -1.16e-08*<br />
(6.23e-09)<br />
s.d. Salary 1.38e-08* 1.30e-08*<br />
(7.56e-09) (7.38e-09)<br />
Mean Experience 0.00714 0.00658<br />
(0.00598) (0.00625)<br />
s.d. Experience -0.0171 -0.0182*<br />
(0.0103) (0.0104)<br />
L<strong>in</strong>e-ups 238 238<br />
Robust standard errors <strong>in</strong> paren<strong>the</strong>ses<br />
*** p
Table 3 also shows that over-shooters tend to offensive rebound specialists; defensive<br />
rebound<strong>in</strong>g does not appear to have a reliable impact, although <strong>the</strong> estimate is noisy. The<br />
player position dummies are not significant and are thus suppressed for brevity. The fact that<br />
higher paid players are more likely to under shoot is perhaps surpris<strong>in</strong>g at first. For <strong>in</strong>stance,<br />
some readers might have <strong>the</strong> <strong>in</strong>tuition that <strong>NBA</strong> players <strong>in</strong>terests diverge from team <strong>in</strong>terests<br />
<strong>in</strong> that <strong>the</strong>y have <strong>the</strong> <strong>in</strong>centive to raise <strong>the</strong>ir po<strong>in</strong>t average through suboptimal play. Under<br />
this view, <strong>the</strong> labor market rewards <strong>the</strong> wrong attributes (po<strong>in</strong>ts per game as opposed to<br />
efficiency, for example). Our results are <strong>in</strong>consistent with <strong>the</strong> view of labor market. Boost<strong>in</strong>g<br />
<strong>in</strong>dividual production at a cost to <strong>the</strong> team is not a strategy employed by <strong>NBA</strong> players. We<br />
th<strong>in</strong>k this is <strong>in</strong>terest<strong>in</strong>g <strong>in</strong> its own right. Teams still have a pr<strong>in</strong>ciple-agent problem <strong>in</strong> that<br />
long-term contracts create a moral hazard for effort, but it is <strong>in</strong>terest<strong>in</strong>g that very few players<br />
exhibit “selfish play.” Our belief is that <strong>the</strong> prevalence of under-shoot<strong>in</strong>g among <strong>the</strong> higher<br />
paid players is evidence that <strong>the</strong> better players conserve energy at times due to <strong>the</strong>ir high<br />
play<strong>in</strong>g time and long season (over-shoot<strong>in</strong>g would be impossible to rationalize this way and<br />
we say far less over-shoot<strong>in</strong>g).<br />
7.3 MSNE, purified games and our contribution<br />
Past work has studied relatively simple games. We study a much more difficult game and<br />
<strong>in</strong> this way really put <strong>the</strong> <strong>the</strong>ory to <strong>the</strong> test. Importantly, while this game is difficult, it<br />
is far easier than equilibria implied by many models that have complicated MSNE, such as<br />
bidd<strong>in</strong>g <strong>in</strong> common value auctions. If we take our <strong>the</strong>ory seriously, we should apply <strong>the</strong>se<br />
“hard” tests.<br />
The results are supportive that experts can reach equilibrium <strong>in</strong> complex games. <strong>NBA</strong><br />
players appear to be superb optimizers. Fur<strong>the</strong>rmore this work <strong>in</strong>corporate <strong>the</strong> concept of<br />
purified games by model<strong>in</strong>g MSNE <strong>in</strong> <strong>the</strong> context of dynamic thresholds. Players adherence<br />
to <strong>the</strong>se dynamic thresholds is strik<strong>in</strong>g. Exceed<strong>in</strong>gly few are found to over-shoot and with<strong>in</strong><br />
l<strong>in</strong>e-ups spread is quite low. We do f<strong>in</strong>d that some (especially star) players under-shoot,<br />
potentially to conserve energy. Overall all <strong>the</strong> hallmarks of optimal stopp<strong>in</strong>g and MSNE are<br />
present.<br />
7.4 Dynamic stopp<strong>in</strong>g problems<br />
Past work has also studied stopp<strong>in</strong>g problems similar to this one <strong>in</strong> <strong>the</strong> lab. It is hard to know<br />
how to generalize <strong>the</strong>se studies though, because <strong>the</strong>se are difficult problems that my require<br />
l<strong>in</strong>e-ups that it weakens <strong>the</strong> relationship considerably.<br />
40
experience and tra<strong>in</strong><strong>in</strong>g. Perhaps, however, people just do typically have <strong>the</strong> capacity to<br />
solve <strong>the</strong>se complex problems, even with experience. This would be a very important results<br />
for models of labor search. There have been papers that have looked at optimal stopp<strong>in</strong>g<br />
problems us<strong>in</strong>g highly tra<strong>in</strong>ed professionals. These studies use deliberative decisions made<br />
by firms harvest<strong>in</strong>g trees (Provencher, 1997), renew<strong>in</strong>g patents (Pakes, 1986) and replac<strong>in</strong>g<br />
bus eng<strong>in</strong>es (Rust, 1987). In <strong>the</strong> field of sports, Romer (2006) is most similar to this paper<br />
<strong>in</strong> <strong>the</strong> study of coaches decisions to “go for it” on 4th down on <strong>the</strong> NFL. Aga<strong>in</strong> this is a<br />
deliberative decision made perhaps 10 times per season by <strong>the</strong> coach (<strong>in</strong> this sense it is one<br />
that he may not have a lot of experience <strong>in</strong> mak<strong>in</strong>g).<br />
This study uses a huge volume of quick decision made players observ<strong>in</strong>g a random arrival<br />
of shoot<strong>in</strong>g opportunities. The f<strong>in</strong>d<strong>in</strong>gs <strong>in</strong>dicate <strong>the</strong> players do this quite well, far better<br />
than lab subjects. Tra<strong>in</strong>ed <strong>in</strong>dividuals appear quite capable of solv<strong>in</strong>g problems with <strong>the</strong><br />
level of complexity of classical optimal stopp<strong>in</strong>g problems.<br />
7.5 Players are not perfect<br />
While <strong>NBA</strong> players do show strik<strong>in</strong>g adherence to optimality overall, <strong>the</strong>y are not perfect<br />
optimizers. We f<strong>in</strong>d that a m<strong>in</strong>ority of players over-shoot and some l<strong>in</strong>e-ups show significant<br />
deviations from allocative efficiency. Fur<strong>the</strong>rmore this can persist even for l<strong>in</strong>e-ups that play<br />
toge<strong>the</strong>r quite frequently. O<strong>the</strong>r work has established that m<strong>in</strong>or mistakes are made <strong>in</strong> <strong>NBA</strong><br />
player. For <strong>in</strong>stance Rao (2010) f<strong>in</strong>ds that a m<strong>in</strong>ority of players behaviorally respond to past<br />
shot success by tak<strong>in</strong>g more difficult shots, despite <strong>the</strong> fact <strong>the</strong>y perform no better on <strong>the</strong>se<br />
shots (hot hand fallacy). Our f<strong>in</strong>d<strong>in</strong>gs here are consistent with <strong>the</strong> idea that play is near<br />
optimal, but that some players and coaches have room for improvement. The better teams<br />
adhere more tightly, which is natural.<br />
8 Conclusion<br />
Past work has studied experts play<strong>in</strong>g relatively simple games. We study a much more<br />
difficult game and <strong>in</strong> this way really put <strong>the</strong> <strong>the</strong>ory to <strong>the</strong> test. Importantly, while this game<br />
is difficult, it is far easier than equilibria implied by many models that have complicated<br />
MSNE, such as bidd<strong>in</strong>g <strong>in</strong> common value auctions. If we take our <strong>the</strong>ory seriously, we<br />
should apply <strong>the</strong>se “hard” tests. The unique decision environment we study allows us to<br />
extend a stylized stopp<strong>in</strong>g problem from <strong>the</strong> lab to a field sett<strong>in</strong>g. Fur<strong>the</strong>rmore, we extend<br />
<strong>the</strong> analysis of allocative efficiency across pure strategies to a markedly more difficult game.<br />
The trade off is that is more challeng<strong>in</strong>g to get hard-and-fast optimality conditions, but our<br />
41
model<strong>in</strong>g approach telescopes <strong>in</strong> <strong>the</strong> strength of assumptions and still provides mean<strong>in</strong>gful<br />
conclusions about <strong>the</strong> adherence to <strong>the</strong> <strong>the</strong>oretical standards employed.<br />
The paper uses a huge volume of quick decisions made by players observ<strong>in</strong>g a random<br />
arrival of shoot<strong>in</strong>g opportunities. The decision to shoot is modeled as a dynamic allocative<br />
stopp<strong>in</strong>g problem. By us<strong>in</strong>g a realistic model<strong>in</strong>g approach we are able to derive strict tests<br />
of optimality. We f<strong>in</strong>d that players overall adhere quite closely to <strong>the</strong> <strong>the</strong>oretical predic-<br />
tions; overall <strong>the</strong>y are suburb optimizers, although mistakes are made. In <strong>the</strong> context of<br />
dynamic efficiency, <strong>the</strong> shot threshold has precisely <strong>the</strong> correct slope and nearly overlaps <strong>the</strong><br />
cont<strong>in</strong>uation value of <strong>the</strong> possession. The mistakes that are made tend to be undershoot<strong>in</strong>g,<br />
<strong>in</strong> that <strong>the</strong> cont<strong>in</strong>uation value is lower than <strong>the</strong> marg<strong>in</strong>al shot; <strong>the</strong>se “mistakes” could be<br />
rationalized by <strong>the</strong> conservation of energy across possessions. In allocative decision mak<strong>in</strong>g,<br />
most teams show a very low variance of marg<strong>in</strong>al efficiencies across players on <strong>the</strong> court for<br />
each <strong>in</strong>terval of <strong>the</strong> shot clock, consistent with near optimal shar<strong>in</strong>g of <strong>the</strong> ball. Overall our<br />
results extend <strong>the</strong> realm and difficulty level of games that humans can play accord<strong>in</strong>g to<br />
game <strong>the</strong>oretic equilbrium.<br />
References<br />
Andrews, D. and Cheng, X. (2010). Estimation and <strong>in</strong>ference with weak identification.<br />
http://www.pr<strong>in</strong>ceton.edu/erp/.<br />
Bearden, J., Murphy, R., and Rapoport, A. (2005). A multi-attribute extension of <strong>the</strong> secre-<br />
tary problem: Theory and experiments. Journal of Ma<strong>the</strong>matical Psychology, 49(5):410–<br />
422.<br />
Bearden, J., Rapoport, A., and Murphy, R. (2006). Sequential observation and selection<br />
with rank-dependent payoffs: An experimental study. Management Science, 52(9):1437.<br />
Berri, D. (1999). Who is most valuable? Measur<strong>in</strong>g <strong>the</strong> player’s production of w<strong>in</strong>s <strong>in</strong> <strong>the</strong><br />
National Basketball Association. Managerial and Decision Economics, 20(8):411–427.<br />
Chiappori, P.-A., Levitt, S., and Groseclose, T. (2002). Test<strong>in</strong>g mixed-strategy equilibria<br />
when players are heterogeneous: The case of penalty kicks <strong>in</strong> soccer. The American<br />
Economic Review, 92(4):1138–1151.<br />
Dufour, J.-M. and Valery, P. (2010). Wald-type tests when rank conditions fail: A<br />
smooth regularization approach. http://www.fields.utoronto.ca/programs/scientific/10-<br />
11/actuarialmath/Dufour Valery Regularized Wald Test 2010 09 15.pdf.<br />
42
Erev, I. and Roth, A. E. (1998). Predict<strong>in</strong>g how people play games: Re<strong>in</strong>forcement learn<strong>in</strong>g<br />
<strong>in</strong> experimental games with unique, mixed strategy equilibria. The American Economic<br />
Review, 88(4):848–881.<br />
Harsanyi, J. (1973). Games with randomly disturbed payoffs: A new rationale for mixed-<br />
strategy equilibrium po<strong>in</strong>ts. International Journal of Game Theory, 2(1):1–23.<br />
Hsu, S., Huang, C., and Tang, C. (2007). M<strong>in</strong>imax play at wimbledon: Comment. The<br />
American Economic Review, 97(1):517–523.<br />
Kahan, J., Rapoport, A., and Jones, L. (1967). Decision mak<strong>in</strong>g <strong>in</strong> a sequential search task.<br />
Attention, Perception, & Psychophysics, 2(8):374–376.<br />
Kovash, K. and Levitt, S. (2009). Professionals do not play m<strong>in</strong>imax: evidence from major<br />
league baseball and <strong>the</strong> national football league. Technical report, National Bureau of<br />
Economic Research.<br />
Lee, M., O’Connor, T., and Welsh, M. (2004). Decision-Mak<strong>in</strong>g on <strong>the</strong> Full Information<br />
Secretary Problem. In Proceed<strong>in</strong>gs of <strong>the</strong> 26th annual conference of <strong>the</strong> cognitive science<br />
society, pages 819–824. Citeseer.<br />
Lee, M. D. (2006). A hierarchal bayesian model of human decision mak<strong>in</strong>g on an optimal<br />
stopp<strong>in</strong>g problem. Cognitive Science, 30.<br />
Levitt, S., List, J., and Reiley, D. (2010). What Happens <strong>in</strong> <strong>the</strong> Field Stays <strong>in</strong> <strong>the</strong> Field:<br />
Explor<strong>in</strong>g Whe<strong>the</strong>r Professionals Play M<strong>in</strong>imax <strong>in</strong> Laboratory Experiments. Economet-<br />
rica, 78(4):1413–1434.<br />
McCabe, K., Mukherji, A., and Runkle, D. (2000). An Experimental Study of Informa-<br />
tion and Mixed-strategy Play <strong>in</strong> <strong>the</strong> Three-person Match<strong>in</strong>g-pennies Game. Economic<br />
Theory, 15(2):421–462.<br />
Metrick, A. (1995). A Natural Experiment <strong>in</strong> “Jeopardy!”. The American Economic Review,<br />
85(1):240–253.<br />
Ochs, J. (1995). Games with Unique, Mixed Strategy Equilibria: An Experimental Study.<br />
Games and Economic Behavior, 10(1):202–217.<br />
Oliver, D. (2004). Basketball on paper: rules and tools for performance analysis. Brassey’s.<br />
O’Neill, B. (1987). Non-parametric Test of <strong>the</strong> M<strong>in</strong>imax Theory of Two-person Zerosum<br />
Games. Proceed<strong>in</strong>gs of <strong>the</strong> National Academy of Sciences, 84(7):2106–2109.<br />
43
Pakes, A. (1986). Patents as options: Some estimates of <strong>the</strong> value of hold<strong>in</strong>g european patent<br />
stocks. Econometrica, 54(4):755–784.<br />
Palacios-Huerta, I. (2003). Professionals play m<strong>in</strong>imax. Review of Economic Studies,<br />
70(2):395–415.<br />
Palacios-Huerta, I. and Volij, O. (2008). Experientia Docet: Professionals Play M<strong>in</strong>imax <strong>in</strong><br />
Laboratory Experiments. Econometrica, 76(1):71.<br />
Palacios-Huerta, I. and Volij, O. (2009). Field centipedes. The American Economic Review,<br />
99(4):1619–1635.<br />
Provencher, B. (1997). Structural versus reduced-form estimation of optimal stopp<strong>in</strong>g prob-<br />
lems. American Journal of Agricultural Economics, 79(2):357–368.<br />
Rao, J. M. (2010). Experts’ Perceptions of Autocorrelation: The Hot Hand Fallacy Among<br />
Professional Basketball Players. Work<strong>in</strong>g Paper.<br />
Rapoport, A. and Tversky, A. (1970). Choice behavior <strong>in</strong> an optional stopp<strong>in</strong>g tasks. Orga-<br />
nizational Behavior and Human Performance, 5(2):105–120.<br />
Romer, D. (2006). Do firms maximize? Evidence from professional football. Journal of<br />
Political Economy, 114(2).<br />
Rust, J. (1987). <strong>Optimal</strong> replacement of gmc bus eng<strong>in</strong>es: An empirical model of harold<br />
zurcher. Econometrica, 55(5):999–1033.<br />
Seale, D. and Rapoport, A. (2000). <strong>Optimal</strong> stopp<strong>in</strong>g behavior with relative ranks: <strong>the</strong> sec-<br />
retary problem with unknown population size. Journal of Behavioral Decision Mak<strong>in</strong>g,<br />
13(4):391–411.<br />
Shachat, J. (2002). Mixed Strategy Play and <strong>the</strong> M<strong>in</strong>imax Hypo<strong>the</strong>sis. Journal of Economic<br />
Theory, 104(1):189–226.<br />
Sk<strong>in</strong>ner, B. (2010). The Price of Anarchy <strong>in</strong> Basketball. Journal of Quantitative Analysis<br />
<strong>in</strong> Sports, 6(1):3.<br />
Staw, B. and Hoang, H. (1995). Sunk costs <strong>in</strong> <strong>the</strong> nba: Why draft order affects play<strong>in</strong>g<br />
time and survival <strong>in</strong> professional basketball. Adm<strong>in</strong>istrative Science Quarterly, pages<br />
474–494.<br />
44
Walker, M. and Wooders, J. (2001). M<strong>in</strong>imax play at wimbledon. The American Economic<br />
Review, 91(5):1521–1538.<br />
White, H. (1982). Maximum likelihood estimation of misspecified models. Econometrica:<br />
Journal of <strong>the</strong> Econometric Society, 50(1):1–25.<br />
Zwick, R., Rapoport, A., Lo, A., and Muthukrishnan, A. (2003). Consumer sequential search:<br />
Not enough or too much? Market<strong>in</strong>g Science, pages 503–519.<br />
9 Appendix<br />
9.1 Proof of Theorem 1<br />
Consider two arbitrary, feasible comb<strong>in</strong>ations of ”actions” by <strong>the</strong> offensive and defensive<br />
player: { <br />
d N p , q N p }N=0,1. Additionally, pick an arbitrary λ ∈ [0, 1] and def<strong>in</strong>e <strong>the</strong> appropriate<br />
convex comb<strong>in</strong>ation of <strong>the</strong> previous two actions<br />
{ d λ p, q λ p } = λ{ d 1 p, q 1 p} + (1 − λ){ d 0 p, q 0 p}<br />
We first fix <strong>the</strong> defenders action at d 1 p and demonstrate that better than sets are convex<br />
<strong>in</strong> hazard rates.<br />
Vp,t( d 1 p, q λ p ) =<br />
=<br />
T <br />
T <br />
q λ p,j,t ep,j,t(d 1 p,j , qλ p,j,t )<br />
t=0 j∈Op<br />
∞<br />
(λ + (1 − λ))<br />
F −1<br />
η (1−qλ p,j,t )<br />
ηp,j,tdFη<br />
t=0 j∈Op<br />
T<br />
<br />
<br />
∞<br />
=<br />
λ<br />
t=0 j∈Op<br />
F −1<br />
η (1−q1 p,j,t )<br />
∞<br />
ηp,j,tdFη + (1 − λ)<br />
F −1<br />
η (1−q0 p,j,t )<br />
−1<br />
Fη ηp,j,tdFη + λ<br />
(1−qλ p,j,t )<br />
F −1<br />
η (1−q1 p,j,t )<br />
−1<br />
F<br />
ηp,j,tdFη + (1 − λ)<br />
= λVp,t( d1 p , q1 p ) + (1 − λ)Vp,t( d1 p , q0 T<br />
<br />
<br />
−1<br />
Fη p ) +<br />
λ<br />
t=0 j∈Op<br />
(1−qλ p,j,t )<br />
F −1<br />
η (1−q1 p,j,t )<br />
−1<br />
Fη ηp,j,tdFη + (1 − λ)<br />
(1−qλ p,j,t )<br />
F −1<br />
η (1−q0 p,j,t )<br />
<br />
ηp,j,tdFη<br />
≥ λVp,t( d 1 p, q 1 p) + (1 − λ)Vp,t( d 1 p, q 0 p)<br />
η (1−qλ p,j,t )<br />
F −1<br />
η (1−q0 p,j,t )<br />
ηp,j,tdFη<br />
Where <strong>the</strong> f<strong>in</strong>al <strong>in</strong>equality follows because <strong>the</strong> term <strong>in</strong> paren<strong>the</strong>sis is non-negative for ev-<br />
ery player and time period. We demonstrate below, by assum<strong>in</strong>g (without loss of generality)<br />
that q 1 p,j,t ≥ q λ p,j,t ≥ q 0 p,j,t.<br />
45
λ<br />
= λ<br />
F −1<br />
η (1−q λ p,j,t )<br />
F −1<br />
η (1−q 1 p,j,t )<br />
F −1<br />
η (1−q λ p,j,t )<br />
F −1<br />
η (1−q 1 p,j,t )<br />
−1<br />
Fη (1−q<br />
ηp,j,tdFη + (1 − λ)<br />
λ p,j,t )<br />
F −1<br />
η (1−q 0 p,j,t )<br />
−1<br />
Fη (1−q<br />
ηp,j,tdFη − (1 − λ)<br />
0 p,j,t )<br />
F −1<br />
η (1−q λ p,j,t )<br />
ηp,j,tdFη<br />
ηp,j,tdFη<br />
≥ λ(q 1 p,j,t − q λ p,j,t)F −1<br />
η (1 − q λ p,j,t) − (1 − λ)(q λ p,j,t − q 0 p,j,t)F −1<br />
η (1 − q λ p,j,t)<br />
= 0.<br />
Concavity of <strong>the</strong> value function <strong>in</strong> <strong>the</strong> defenders action is more straightforward and follows<br />
directly from <strong>the</strong> previously established fact that d2 ep,j,t<br />
dd 2 p,j<br />
Vp,t( d λ p, q 1 p) =<br />
≤<br />
T <br />
t=0 j∈Op<br />
T <br />
t=0 j∈Op<br />
9.2 Proof of Theorem 2<br />
q 1 p,j,t ep,j,t(d λ p,j, q 1 p,j,t)<br />
< 0.<br />
q 1 p,j,t (λep,j,t(d 1 p,j, q 1 p,j,t) + (1 − λ)ep,j,t(d 0 p,j, q 1 p,j,t))<br />
≤ λVp,t( d 1 p, q 1 p) + (1 − λ)Vp,t( d 0 p, q 1 p)<br />
Thus we have demonstrated <strong>the</strong> existence of a Nash Equilibrium <strong>in</strong> our game. That is we<br />
have a fixed po<strong>in</strong>t solution to <strong>the</strong> problem.<br />
Or equivalently:<br />
d ∗ p = BRd(u ∗ p)<br />
u ∗ p = BRu( d ∗ p)<br />
d ∗ p = BRd(BRu( d ∗ p)) = BRd,u( d ∗ p) (36)<br />
<br />
u ∗ p = BRu( d ∗ p) (37)<br />
Suppose, for sake of contradiction, that half-court offense had a multiplicity of equilibria:<br />
46
∃ d ∗ p, d ∗∗<br />
p , such that:<br />
BRd,u( d ∗ p) − d ∗ p = BRd,u( d ∗∗<br />
p ) − d ∗∗<br />
p = 0<br />
( d ∗∗<br />
p − d ∗ p) = ∆p = 0<br />
In particular we know ∃j ∈ Op such that ∆p,j = d ∗∗<br />
p,j − d ∗ p,j > 0.<br />
α ′ T Dp ∆d = 0<br />
Now that we have demonstrated <strong>the</strong> existence of a unique equilibrium <strong>in</strong> our game, we<br />
can proceed to estimate our model without concern for <strong>the</strong> possibility of sudden regime<br />
changes that would change our <strong>in</strong>terpretation of player shoot<strong>in</strong>g decisions.<br />
9.3 Parametric Model Specification Tests<br />
Our model is identified by assum<strong>in</strong>g <strong>in</strong>variance of a player’s ability to realize scor<strong>in</strong>g oppor-<br />
tunities across different values of <strong>the</strong> shot clock. However, because we make this assumption<br />
for 16 different periods of <strong>the</strong> shot clock, our model is overidentified and our assumptions<br />
can be tested. Suppose for example that defenses became progressively more tenacious over<br />
<strong>the</strong> course of <strong>the</strong> shot clock and that player’s opportunity distributions generally decl<strong>in</strong>ed<br />
across <strong>the</strong> shot clock. Then, we would f<strong>in</strong>d that a s<strong>in</strong>gle distribution could not accurately<br />
reflect a players ability to score <strong>in</strong> both <strong>the</strong> beg<strong>in</strong>n<strong>in</strong>g and end of <strong>the</strong> shot clock. Players<br />
would end up shoot<strong>in</strong>g less and less efficiently towards <strong>the</strong> end of <strong>the</strong> shot clock than our<br />
model would predict. In future versions of this paper, we hope to present formal tests of<br />
our specification. For now, we present plots show<strong>in</strong>g how our model preforms across <strong>the</strong><br />
shot clock as an average across all players <strong>in</strong> our sample. We take solace <strong>in</strong> <strong>the</strong> lack of any<br />
obvious trends <strong>in</strong> our residuals.<br />
9.4 Additional support for Risk Neutrality<br />
The follow<strong>in</strong>g plot shows <strong>the</strong> simulated w<strong>in</strong>n<strong>in</strong>g percentage for an underdog with basel<strong>in</strong>e<br />
mean expected po<strong>in</strong>t value of 1.38 play<strong>in</strong>g a team who averages 1.4 po<strong>in</strong>ts per possession<br />
with standard deviation 0.45. Each “game” was simulated 10,000 times. As evidenced by<br />
<strong>the</strong> figure, although <strong>the</strong> underdog wants to <strong>in</strong>crease <strong>the</strong> standard deviation of <strong>the</strong> expected<br />
value of shot attempts it does not want to trade off any mean<strong>in</strong>gful amount of mean to do<br />
so.<br />
47
Appendix Figure 1: Underdog w<strong>in</strong>n<strong>in</strong>g percentage as a function of standard deviation and mean<br />
9.5 When Does Half-Court Offense Beg<strong>in</strong>?<br />
We decided that half-court offense began with 17 seconds on <strong>the</strong> shot clock. Our reason for<br />
do<strong>in</strong>g so, is that prior to 17 seconds <strong>the</strong> average value of exercis<strong>in</strong>g a possession is found to<br />
be strongly correlated with <strong>the</strong> mechanism by which <strong>the</strong> possession orig<strong>in</strong>ated (steal, dead<br />
ball, or defensive rebound). However, from 17 seconds on team’s are <strong>in</strong> a half court set and<br />
<strong>the</strong> average value of possession use is now <strong>in</strong>dependent of how <strong>the</strong> possession orig<strong>in</strong>ated.<br />
48
9.6 3-man Cores<br />
Distribution of spread for 3-man cores. Higher values <strong>in</strong>dicate larger deviations from<br />
optimality.<br />
49
Figure 5: Player-by-player t-statistic for deviations from dynamic optimality. Positive values<br />
<strong>in</strong>dicate “under-shoot<strong>in</strong>g.”<br />
51
Figure 6: Distribution of spread for 4-man cores. Higher values <strong>in</strong>dicate larger deviations<br />
from optimality.<br />
52