2014 Research Paper Competition

Presented by:


1
The Hot Hand: A New Approach to an
Old Fallacy

Andrew Bocskocsky, John Ezekowitz, and Carolyn Stein
Harvard University,
Cambridge, MA, 02138
Email: [email protected],
[email protected], [email protected]


Abstract
The vast literature on the Hot Hand Fallacy in basketball rests on the assumption that shot
selection is independent of player-perceived hot or coldness. In this paper, we challenge this
assumption using a novel dataset of over 83,000 shots from the 2012-2013 National Basketball
Association (NBA) season, combined with optical tracking data of both the players and the ball.
We create a comprehensive model of shot difficulty using relevant initial shot conditions, and use
it to show that players who have exceeded their expectation over recent shots shoot from
significantly further away, face tighter defense, are more likely to take their teams next shot, and
take more difficult shots. We then turn to the Hot Hand itself and show that players who are
outperforming will continue to do so, conditional on the difficulty of their present shot. Our
estimates of the Hot Hand effect range from 1.2 to 2.4 percentage points in increased likelihood
of making a shot.

1 Introduction

Humans are inherently bad at handling uncertainty and chance. We crave order and symmetry, and this tendency
often leads us astray in the face of randomness. This is evidenced by the belief in local representativeness, or the
law of small numbers, first formally defined by Kahneman and Tversky (1971). Humans, the authors write,
expect the essential characteristics of a chance process to be represented not only globally in the entire sequence,
but also locally, in each of its parts (1985).

The Hot Hand or streak shooting has been used as an example of local representativeness. An average NBA
viewer observes Player X, a sixty percent shooter, make four shots in a row. He thinks to himself, this isnt
representative of his normal sixty percent shooting Player X must be hot! The statistician, however, knows that
streaks of four makes in a row are inevitable over the course of a season. Short sequences of shots need not be
consistent with the players overall shooting percentage. This debate, seemingly long-ago settled in favor of the
statistician, has given rise to perhaps the most well-known theories in behavioral economics: the Hot Hand Fallacy.

The seminal piece of research that attempted to answer this question was done by Thomas Gilovich, Robert
Vallone, and Amos Tversky in 1985. In this paper, the authors analyzed series of shots by players on the
Philadelphia 76ers, looking for a positive correlation between successive shots, and found none. They also analyzed
series of free throws by the Boston Celtics and the Cornell mens and womens basketball teams, and again found no
evidence of serial correlation. Subsequent studies, including Adams, 1992; Koehler and Conley, 2003; Bar-Eli,
Avugos and Raab, 2006; and Rao 2009 have confirmed this finding.

Today, among the academic crowd, the Hot Hand is almost universally considered a fallacy. Larry Summers even
chastised the Harvard mens basketball team for their belief in the phenomenon (Davidson, 2013). Yet, among
basketball fans and players, the Hot Hand is a myth that refuses to die. Professional players themselves reported
feeling that they almost cant miss after making several shots in a row (Kahneman and Tversky, 1971). Players
actions confirm this sentiment, as shot difficulty tends to increase following several made shots (Rao, 2009).
Therefore, we believe that the validity of the Hot Hand remains an open question. For, buried in the introduction of
the famous Gilovich, Vallone, and Tversky paper lies a key assumption that:

Each player has an ensemble of shots that vary in difficulty (depending, for example, on the
distance from the basket and on defensive pressure), and each shot is randomly selected from this ensemble
(Gilovich, Vallone and Tversky, 1985, emphasis added).



2014 Research Paper Competition
Presented by:


2
The strength of the authors conclusion hinges heavily on this assumption, and yet it is not difficult to envision a
scenario in which it is violated. There is strong evidence that players themselves believe in the Hot Hand and may
not select shots at random. Players taking more difficult shots after past makes may cancel out or mask the Hot
Hand effect.

In this paper, we use a novel dataset, provided by the SportVU tracking cameras of STATS, Inc. This dataset
consists of over 83,000 shot attempts from the 2012-2013 NBA season. Synthesizing this dataset, we are able to
know almost all relevant characteristics of the shot at the moment it is taken. This allows us to investigate the
following two questions: First, do players (both offensive players and defenders) believe in the Hot Hand, as
evidenced by their playing decisions? And second, if we control for the difficulty of the shot, does the Hot Hand
effect indeed emerge?

We show that players who perceive themselves to be hot based on previous shot outcomes shoot from significantly
further away, face tighter defense, are more likely to take their teams subsequent shot, and take more difficult shots.
These results invalidate the shot selection independence assumption. To account for this bias, we create a
comprehensive model of shot difficulty that depends on conditions of the shot the moment it is taken. These
conditions include variables relevant to game situation, shot location, and defender locations. Next, we create a
measure of heat which reflects the extent to which a player outperformed over his past few shots, bearing in mind
how difficult those shots were. By having both a measure of shot difficulty and a measure of heat, we have a way to
test for the Hot Hand, holding shot difficulty constant. The results of this test suggest that once we control for the
dependence of shot selection, there may be a small yet significant Hot Hand effect.

2 Data and Empirical Preliminaries

2.1 Raw Data

STATS, Inc. introduced the SportVU optical tracking system into the NBA in 2010. The system uses six cameras,
three on each side of the court, to provide precise three-dimensional image tracking of the players, referees, and ball
every 1/25th of a second. Our dataset consists of every game played at the 15 SportVU equipped arenas in the
2012-2013 regular season.

We create a shot log which synthesizes the optical tracking data from SportVU and the play-by-play data from the
NBA to create robust characterization of each shot. For each shot, we have information on the player who took the
shot, the type of shot taken, and the time and score at the time of the shot. From the optical tracking data, we have
the precise location of the ball and all ten players, both offensive and defensive, on the court.

2.2 Predicted Shot Difficulty

Using the shot log data described in Section 2.1, we estimate a model that predicts the difficulty of each shot for
player i taking shot s, based on four broad categories of determinants of shot difficulty:

P

is
= + * (Game Condition Controls
is
)+ * (Shot Controls
is
)
+ * (Defensive Controls
is
)+ * (Player Fixed Effects
i
) ( 1 )

We use Game Condition Controls, such as time remaining and the score differential between the teams to proxy for
differences in pressure, player fatigue, and effort across shots. We use Shot Controls, such as the precise distance
from the basket of each shot and its play-by-play categorization (i.e., fadeaway jump shot, reverse layup, etc), to
estimate the difficulty of the shot. Because shot efficiency varies non-linearly with shot distance (see Figure 1 for a
visual representation), we break distance into two-by-two foot increments and include this set of mutually exclusive
and exhaustive dummy variables in the shot difficulty model. This allows for a non-parametric specification, where
each two-by-two foot increments coefficient is allowed to vary.

We use Defensive Controls to measure defensive intensity. Using the SportVU data, we are able to determine and
use both the absolute distance between the player shooting and the closest defender and the angle of that defender
relative to a straight line between the shooter and the basket. We also include the height differential between the
closest defender and the shooter. This is interacted with the distance between the two players to give a measure of
size mismatch.



2014 Research Paper Competition
Presented by:


3
Finally, we use Player Fixed Effects to control for differences between players. If Kevin Durant and Tyson Chandler
both take identical jump shots, the two shots likely have different likelihoods of being made. Player Fixed Effects
allow us to capture this difference.

Shot Difficulty Regression Results

Table 1 shows the fit of this regression, as well as the specific controls used. For ease of notation, we will refer to
the estimated probability of a make as simply P

throughout the paper. We used OLS rather than a Logit or Probit

regression because of issues with model convergence. Therefore, when the model predicted values of greater than
one or less than zero, we replaced them with 0.99 and 0.01 respectively.2 Because we used such a large number of
variables and interactions, along with a very non-parametric specification (i.e. breaking distance into two-by-two
zones rather than using a linear measure), most of the hundreds of coefficients do not have an easy interpretation.
Thus, no coefficients are reported.

To test the accuracy of the model, we ran it on a randomized training set consisting of half of the data. We then
predicted P

s for the validation set, creating a pseudo-out of sample test. We grouped the P

s into bins of one
percent increments, and calculated the actual make percentage for each bin. If the model is accurate, the P

bins
should correspond closely with actual made percentages. Figure 2 presents the scatter plot of the data and illustrates
that model generally fits the data very well.

Defining Heat

Before we can consider the effect of heat, we must first define what heat actually is. Note that we only consider
consecutive shots that occur within the same game. The conventional definition of heat is:

Simple Heat
n
= Actual % over past n shots ( 2 )

However, we believe that in order to test for the Hot Hand, we must define what we call Complex Heat. Note that
the P

values allow us to calculate the expected shooting percentage over the past n shots. Therefore, we define:

Complex Heat
n
= Actual % over past n shots - Expected % over past n shots ( 3 )

Simple Heat is the commonly understood measure of heat - it simply reflects how successful a player has been over
his past few shots. However, a drawback of this is that it does not account for shot difficulty. A player who has gone
five-for-five on five consecutive layups will be labeled hotter than a player who has made four out of five three-
point attempts. Complex Heat corrects for this, and looks for players who are shooting better than expected, given
the difficulty of the shots they are taking.

Another drawback of Simple Heat is illustrated by the following example. Suppose a player is playing against a very
short defender. It is likely that he is making more of his shots than usual, and his Simple Heat will be high.
However, as long as he keeps shooting against this low-quality defender, he is still likely to make his shots - in other
words P

will also be high. Therefore, we see that Simple Heat and P

have a mechanical correlation. This is not the
case with Complex Heat, because it controls for the difficulty of the past few shots (i.e. the P

of the previous few
shots). Continuing with the short defender example, the defenders short stature will be controlled for when we use
Complex Heat, because the difficulty of the past few shots will be lower.

Finally, we note that there is some discretion in selecting how many shots we look back over when defining heat
(i.e. what value n takes when defining Simple and Complex Heat). We ran our results for all values of n from two
through seven, and found similar results. For simplicity, we report all results here for n = 4, which we believe is a
reasonable number of shots. Moreover, unless otherwise noted, all results that use heat refer to Complex Heat,
which we believe is the more correct measure.



3 Do Players Believe in the Hot Hand?



2014 Research Paper Competition
Presented by:


4
The first question we are interested in is: Do players believe in the Hot Hand? Gilovich, Vallone and Tversky (1985)
showed that players on the 76ers claimed to believe the Hot Hand was true. We are interested in whether over a
quarter century later, players actions reflect this belief.

3.1 Empirical Strategy

Shot Distance

If players truly believe that the Hot Hand exists, they may attempt more difficult shots as they heat up. One way to
test this is to see whether players take shots that are further away from the basket as they become hot. To see if this
is the case, we run the following specification:

Shot Distance
is
= + * (Heat
is
)+ * (Controls
is
)+ * (Player Fixed Effects
i
) ( 4 )

The controls include quarter, score differential, quarter/score differential interaction, closest defender distance, and
angle, and a dummy for fast breaks.

We would hypothesize that if players believe in the Hot Hand, the coefficient on heat would be positive. As players
think they are becoming hot (i.e. better shooters), they optimize by taking shots that are more difficult than shots
they would ordinarily attempt. An alternative explanation is that the defense buys into the Hot Hand belief and
covers hot players more tightly, causing them to attempt shots from further away.

Defender Distance

As mentioned previously, if defenders believe the Hot Hand exists, they may cover hot players more tightly. To test
this hypothesis, we run the following regression:

Defender Distance
is
= + * (Heat
is
)+ * (Controls
is
)+ * (Player Fixed Effects
i
) ( 5 )

The controls in this regression include quarter, score differential, quarter/score differential interaction, fast break,
shot type, and shot distance.

We would hypothesize that if defenders believe in the Hot Hand, then the coefficient on heat would be negative.
Intuitively, as a player becomes hotter, the defenders give him less space to shoot.

Likelihood of Taking Next Shot

Further, we can investigate the effect of heat on overall shot selection by players by evaluating how heat impacts the
probability that a given player takes his teams next shot. More specifically, we look at the probability that after
Player X takes a shot, Player X also takes his teams next shot as a function of heat. To do this, we estimate the
following Probit model:

P (Same
is
) = [ + * (Heat
is
)+ * (Controls
is
)+ * (Player Fixed Effects
i
)] ( 6 )

In this regression, we control for game environment factors - quarter, score differential, and the quarter/score
differential interaction. Here, if players believe in the Hot Hand, we would expect the coefficient on heat to be
positive. If the shooter thinks he is heating up, he is more likely to shoot. Moreover, if his teammates also believe he
is hot, they are more likely to give him the ball in a shooting position.

Overall Shot Difficulty

Finally, we can use our P

model to investigate how heat affects overall shot difficulty. If players attempt shots that
are further away and more closely guarded when they perceive themselves to be hot, we would expect that these
shots overall have a lower value of P

. To see if this is the case, we run:

P

is
= + * ( Heat
is
) ( 7 )




2014 Research Paper Competition
Presented by:


5
Note that we dont include controls or player fixed effects, since P

includes these already. If players do indeed

attempt all-around more difficult shots when hot, we would expect to be negative (recall that a low P

corresponds
to a more difficult shot).

3.2 Results

Shot Distance

Looking at the column (1) of Table 2, we see that shot distance increases with heat. The size of the effect - 2.144
feet - is significant. If a player made rather than missed one more of his past four shots (i.e. increased Complex Heat
by 0.25), his estimated shot distance increases by just over 6 inches. Compared to the average shot distance of 12.5
feet, this is about a 5% increase. This supports the hypothesis that as players become hot, they attempt more
difficult shots.

Defender Distance

Considering column (2) of Table 2, we also see that defender distance shrinks with heat. Again, to better understand
the size of these coefficients, consider what happens if a player makes an additional one out of his past four shots.
In this case, defender distance shrinks by 0.046 feet, or about half an inch. This is a small effect, but it is worth
noting that the average defender distance is only about four feet. Therefore, the effect size is about equal to a 1%
decrease in defender distance. This supports the hypothesis that as a player becomes hot, the defense plays tighter.

Likelihood of Taking Next Shot

Next, we consider how likely it is that a player takes his teams next shot. Column (3) of Table 2 shows how the
probability that the same player takes a teams next shot as a function of heat. If a shooter makes an additional
average-difficulty shot (i.e. all other covariates are at their mean values), then the probability of the same shooter
taking the shot increases by 1.5 percentage points (0.061* 0.25). This sounds small, but remember that there are five
players on a team. If the average player has about a 20% chance of taking his teams next shot, this corresponds to a
7.6% increase in the overall probability.

Overall Shot Difficulty

Finally, we consider overall shot difficulty in column (4) of Table 2. Consistent with what we observed in regards to
shot distance and defender distance, being hot is associated with taking overall more difficult shots. A player who
makes an additional shot will on average see his P

value drop by 1.0 percentage points. Given that the average shot
has about a 40% chance of going in, this can be thought of as attempting a 2.5% more difficult shot. This result, and
the three above it, illustrate that players behave as though they believe in the Hot Hand.

4 Testing the Hot Hand: Shot Difficulty

4.1 Empirical Strategy

Now armed with evidence that shot difficulty increases with heat, we can turn our attention to whether the Hot
Hand truly exists once we control for shot difficulty. In the following section, we present results using our P

shot
difficulty model, comment on those results, and add some discussion of potential drawbacks of this approach.

First, as a baseline, we look for the Hot Hand without any control for shot difficulty. We are interested in how a
players probability of hitting a shot varies with heat and nothing else. Therefore, we can run the following simple
specification:

P(Make
is
)= + * (Heat
is
) + * (Player Fixed Effects
i
) ( 8 )

We can run this using an OLS specification to understand how the probability of making a shot varies with heat,
unconditional on shot type or difficulty. This is analogous to the analysis done by Gilovich, Vallone, and Tversky
(1985) and others, and will provide us with a nice baseline.



2014 Research Paper Competition
Presented by:


6
Next, we can control for difficulty by using P

, which encapsulates all of the relevant, quantifiable controls. We can

use the following relatively simple specification to test for the difficulty-controlled Hot Hand:

P(Make
is
)= + * (Heat
is
) + * P

is
( 9 )

Note that we no longer need player fixed effects, because P

encapsulates those.

If the Hot Hand does not exist, and each shot is truly independent, we would expect = 0, = 0, and = 1. In
other words, the only thing that predicts P (Make) is the difficulty of the shot. However, if the Hot Hand does exist,
we would expect to find > 0.

4.2 Results and Discussion

Column (1) of Table 3 loosely replicates the work done by Gilovich, Vallone, and Tversky by regressing a simple
measure of heat and player fixed effects against the probability of hitting a shot, with no attempt to control for shot
difficulty. Our results mirror the original authors, with the coefficient on heat being negative but insignificant.
The second column introduces shot difficulty controls via P

, but still uses Simple Heat rather than Complex Heat.

As discussed previously, Simple Heat alone may be a misleading measure of heat because it does not account for
factors that affect shot difficulty and remain constant across game stretches. In column (2), we actually see a
supposed significant cold hand effect. This dramatic shift is evidence of the bias of Simple Heat: it does not
control for within-player correlation such as the short defender example. Therefore, we believe that Complex Heat
is a superior measure for testing for the Hot Hand.

As shown in the previous section, shot difficulty increases as a player heats up. Therefore, it seems possible that
once we control for difficulty, the Hot Hand effect would indeed emerge. Our final specification uses Complex
Heat and also controls for difficulty. Looking at column (3), we finally see a positive and significant coefficient on
heat. Though significant, we acknowledge that the effect size is modest. A player who makes one more of his past
four shots sees his shooting percentage increase by 0.54 percentage points. Given that the average NBA player has a
field goal percentage of about 45%, this represents about a 1.2% improvement. In the same vein, if a player makes
two more of his past four shots (perhaps more indicative of what it truly means to be hot), we see a 2.4%
improvement.

Based on these results, concluding that the Hot Hand exists is contingent on defining hot correctly. A player who
makes three out of his past four layups is not hot, but a player who makes three out of his past four three-point
attempts is. In other words, being hot is not about the absolute number of shots a player has previously made, but
rather is about how much he has outperformed, conditional on the types of shots he has taken.

There are several drawbacks to the two stage empirical strategy we employ here of first estimating a Shot Difficulty
regression and then using those results to test for the Hot Hand. The principle worry is errors-in-variables: if our P

s
are measured too imprecisely, they may produce a biased estimator. There is reason to suspect that, despite our best
efforts and the extraordinary dataset at our disposal, our Shot Difficulty model does not control for or does not
correctly specify individual shot difficulties. Specific concerns include: not being able to track appendages, not
specifying relationships between variables (i.e., distance from the basket, defender distance, etc.) correctly, and that
player fixed effects are not precise enough to accurately estimate P

for individual players. Player fixed effects change

the intercept of the probability that a shot goes in for each player, but they do not account for the fact that certain
players specialize in certain shots - in other words, they do not adjust the slopes.

5 Conclusion

For thirty years, the empirical consensus that the Hot Hand in basketball is a fallacy of the human mind has been
confirmed time and again. In the same way that evolutionary biologists might regard creationists as completely
misguided, economists, psychologists and statisticians have viewed the persistent belief in the Hot Hand as entirely
fallacious. Amos Tversky, co-author of the canonical paper on the subject, typifies this view when he says, Ive
been in a thousand arguments over this topic, won them all, but convinced no one (Bar-Eli, Avugos and Raab,
2006).

In this paper, thanks to SportVUs optical tracking data, we are able to show that the key shot selection
independence assumption is not a good one. Players who have either simply made more of their last few shots, or


2014 Research Paper Competition
Presented by:


7
done better than expected on those shots, tend to take more difficult shots. Additionally, hot players are much more
likely to take the teams next shot and thus are not choosing shots independently.
We then extend our analysis to ask if the Hot Hand exists once we control for this dependent shot selection using a
regression framework that controls for past expectation and finds a small, positive, and significant Hot Hand effect.
This conception of the Hot Hand as exceeding expectations is different from the popular conception of absolute
outperformance. Given the endogeneity of shot selection, however, assessing the Hot Hand only through the
absolute outperformance seems flawed.

We caution that our use of OLS regression as the functional form of choice for most of our analysis may not be
optimal. Moreover, our effect sizes are small, and it is more likely that our estimates of standard errors are imprecise
than our estimates of mean effects. It is plausible that better specifications might render these small effects
insignificant. Nevertheless, our effects remain significant across different specifications and with both clustered and
heteroskedasticity robust standard errors.

We hope this paper will spark debate and further work on the mechanisms of the Hot Hand effect in basketball.
Under- standing how both players and defenses react to perceived hotness can provide valuable insight into teams
and players optimal strategies and could lead to a re-evaluation of current strategies. At the very least, our findings
cast doubt on the overwhelming consensus that the Hot Hand is a fallacy. Perhaps the next time a professor
addresses the Harvard Mens basketball team, the Hot Hand will not be so quickly dismissed.

Acknowledgements

We would like to thank Matt Davis, Roland Fryer, Lawrence Katz, Carl Morris and Joe Blitzstein for their help and
insight. We also thank Brian Kopp, Ryan Warkins, and STATS, Inc. for providing us with the raw data. All
remaining errors are our own.

References

[1] Kahneman, Daniel, and Amos Tversky. Subjective Probability: A Judgment of Representativeness. Cognitive
Psychology 3, no. 3 (1972): 430-454.

[2] Gilovich, Thomas, et al. The Hot Hand in Basketball: On the Misperception of Random Sequences. Cognitive
Psychology 17, no. 3 (1985): 295-314.

[3] Adams, Robert M. The Hot Hand Revisited: Successful Basketball Shooting as a Function of Intershot
Interval. Perceptual and Motor Skills 74, no. 3 (1992): 934-934.

[4] Koehler, Jonathan, and Caryn Conley. The Hot Hand Myth in Professional Basketball. Journal of Sport &
Exercise Psychology 25 (2003): 253.

[5] Bar-Eli, Michael, et al. Twenty Years of Hot hand Research: Review and Critique. Psychology of Sport and
Exercise 7, no. 6 (2006): 525-553.

[6] Rao, Justin M. Experts Perceptions of Autocorrelation: The Hot Hand Fallacy Among Professional Basketball
Players. 2009.

[7] Davidson, Adam. Boom, Bust or What? Larry Summers and Glenn Hubbard Square Off on Our Economic
Future. New York Times, May 2, 2013.













2014 Research Paper Competition
Presented by:


8

Appendix: Figures and Regression Tables


Figure 1: Shot Efficiency By Zone



Figure 2: Predicted versus Actual Make Percentages





2014 Research Paper Competition
Presented by:


9


Table 1: Predicting Makes - List of Variables Used in Specification


VARIABLES

Quarter Dummies
Score Differential
Shot Location Dummies
Shot Category Dummies
Fastbreak Dummy
Distance of Closest Defender
Angle of Closest Defender
Shooter-Defender Height Difference
Double Covered Dummy
Quarter x Score Differential
Shot Location Dummies x Shot Category Dummies
Distance of Closest Defender x Angle of Closest Defender
Distance of Closest Defender x Shooter-Defender Height Difference
Player Fixed Effects

Observations 70,862
R
2
0.149





Table 2: Do Players Believe in the Hot Hand?

(1) (2) (3) (4)
VARIABLES Shot Distance Defender Distance P(Same Shooter)
P


Heat 2.144*** -0.183*** 0.0610*** -0.0419***

(0.180) (0.0393) (0.00927) (0.00380)
Constant 8.067*** 4.126*** 0.457***

(0.262) (0.138) (0.000904)
Observations 43,475 43,475 43,467 43,475

R
2
0.296 0.167 0.003
Notes: Shot Distance, Defender Distance, and Shot Difficulty regressions are OLS, while P(Same Shooter) is
probit, and coefficients are marginal effects. Regressions also include controls as described in the body of the
paper. Complex Heat has n=4.
Robust standard errors in parentheses *** p<0.01, **p<0.05, * p<0.1







2014 Research Paper Competition
Presented by:


10




Table 3: Does the Hot Hand Exist?

(1) (2) (3)
VARIABLES P(Make) P(Make) P(Make)

P

1.020*** 1.020***
(0.00984) (0.00982)

Simple Heat -0.0106 -0.0237***
(0.0108) (0.00875)

Complex Heat 0.0215**
(0.00947)

Constant 0.459*** -0.000662 -0.0116**
(0.00506) (0.00629) (0.00505)

Observations 43,475 43,475 43,475

R
2
0.014 0.149 0.149


Notes: Regression (1) also includes player FEs, while regressions (2) and (3)
do not. Simple and Complex Heat both have n = 4.


Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1