Current Biology 24, 19181922, August 18, 2014 2014 Elsevier Ltd All rights reserved http://dx.doi.org/10.1016/j.cub.2014.07.

013
Report
Asymmetric Predictability
and Cognitive Competition
in Football Penalty Shootouts
Erman Misirlisoy
1,
* and Patrick Haggard
1,
*
1
Institute of Cognitive Neuroscience, University College
London, 17 Queen Square, London WC1N 3AR, UK
Summary
Sports provide powerful demonstrations of cognitive stra-
tegies underlying competitive behavior [1]. Penalty shoot-
outs in football (soccer) involve direct competition between
elite players and absorb the attention of millions. The pen-
alty shootout between Germany and England in the 1990
World Cup seminal was viewed by an estimated 46.49%
of the UK population [2]. In a penalty shootout, a goalkeeper
must defend their goal without teammate assistance while
an opposing series of kickers aim to kick the ball past
them into the net. As in many sports [3], the ball during a
penalty kick often approaches too quickly for the goal-
keeper to react to its direction of motion; instead, the goal-
keeper must guess the likely direction of the kick, and dive
in anticipation, if they are to have a chance of saving the
shot [46]. We examined all 361 kicks from the 37 penalty
shootouts that occurred in World Cup and Euro Cup
matches over a 36-year period from 1976 to 2012 and
show that goalkeepers displayed a clear sequential bias.
Following repeated kicks in the same direction, goal-
keepers became increasingly likely to dive in the opposite
direction on the next kick. Surprisingly, kickers failed to
exploit these goalkeeper biases. Our ndings highlight
the importance of monitoring and predicting sequential
behavior in real-world competition. Penalty shootouts pit
one goalkeeper against several kickers in rapid succession.
Asymmetries in the cognitive capacities of an individual
versus a group could produce signicant advantages over
opponents.
Results and Discussion
Previous game-theoretic models have viewed football penalty
kicks as a zero-sum, simultaneous-move game. Analyses of
penalty kicks during match play have suggested that goal-
keepers and kickers both use a mixed strategy, randomly
choosing between left and right dives or kicks [79]. In this
case, the penalty kick situation resembles a mixed-strategy
Nash equilibrium. Crucially, this equilibrium can hold only if
both goalkeepers and kickers choices are serially indepen-
dent and uncorrelated, because any predictable pattern of
sequential behavior could be exploited by the other party.
Remarkably, choices in match-play penalty kicks were indeed
found to be random and serially independent [7, 8], even
though human attempts to act randomly are generally poor
[10, 11]. However, penalty kicks in match play are infrequent
and are generally separated by long periods of time, so serial
independence may be unsurprising.
In contrast, the penalty shootout is based on a series of mul-
tiple kicks taken in rapid succession, after normal match play
has failed to produce a clear winner. Each party could easily
monitor for sequential regularities in the others behavior and
could exploit any departure fromrandomness. The goalkeeper
might anticipate the direction of the kick, for example by
second-guessing the kickers behavioral choices [12, 13]. In
particular, sequences of human choices exhibit a gamblers
fallacy [10, 14]. This involves the belief that in a series of
independent random binary events, such as a coin toss, the
alternative event becomes increasingly likely to occur after
progressively longer runs of one outcome. Following repeated
kicks of the ball in the same direction, a goalkeeper subject to
the gamblers fallacy will anticipate the ball going in the oppo-
site direction on the next kick and will therefore dive accord-
ingly. Importantly, the gamblers fallacy is not based on the
value of any single previous outcome, but on the run length
over which an outcome has already occurred. If goalkeepers
exhibit this regularity, kickers could potentially exploit it to
gain advantage. Previous game-theoretic studies mostly
considered match-play penalty kicks [7]. Although some
examined penalty shootouts [15], they focused primarily on
motivational factors predicting performance, rather than on
runs of decisions by individual players.
We examined online videos and statistics of all penalty
shootouts from FIFA World Cup and UEFA Euro Cup nals
tournaments from 1976 to 2012, comprising 361 penalty kicks
in 37 shootouts. We recorded the direction each ball was
kicked (left/right/center) and also the direction in which the
goalkeeper moved. These correspond to action selection deci-
sions by kickers and goalkeepers, respectively. Goalkeepers
remained in the center only very rarely (2.49%), and kicks to
the center were also rare (9.14%). Moreover, we had no strong
prior assumptions about sequential effects on center choices,
so we removed them, leaving a data set of 321 penalty kicks
involving clear left/right choices.
Goalkeepers were approximately equally likely to dive left or
right, showing no signicant difference from 50%(46.73%left;
p = 0.27, binomial test). Similarly, kickers were equally likely to
shoot left or right (46.42% left; p = 0.22, binomial test). A chi-
square test showed no signicant difference between these
proportions (p = 0.94). Importantly, goalkeepers were no
more likely to dive in the direction of the kick than would be ex-
pected by chance (53.58%; p = 0.22, binomial distribution).
This suggests that goalkeepers are indeed unable to react to
kicker behavior and endogenously choose each dive direction.
We next investigated possible exploitable regularities, by
analyzing effects of repeated ball direction across consecutive
kicks on goalkeeper behavior and determining whether goal-
keepers showed a gamblers fallacy. We also examined
whether sequences of kick direction did indeed depend on
previous kick history. We thus measured the direction of the
kick, and of the goalkeepers dive, after runs of one (n = 159),
two (n = 66), or three (n = 16) consecutive kicks in the same di-
rection during a single penalty shootout. We also considered
the possibility that goalkeepers previous dive directions,
rather than previous kick directions, could predict future
behavior (see Figures S1 and S2 available online). However,
*Correspondence: e.misirlisoy.11@ucl.ac.uk (E.M.), p.haggard@ucl.ac.uk
(P.H.)
we found no evidence that dive history was relevant for kickers
or goalkeepers: kick direction was more important than dive
direction. Therefore, we present here the results based on
runs of kick directions. We used bootstrap resampling [16,
17] to compare the actual data to the same data randomly
shufed 10,000 times. We tested whether goalkeepers and
kickers were more likely than chance to switch direction in
their behavior following long repetitions of kicks in the same
direction.
When kickers repeatedly kicked in the same direction, goal-
keepers became progressively more likely to dive in the oppo-
site direction on the next kick, conrming a gamblers fallacy.
Figure 1 shows an example in which the goalkeeper dived to
the right after three successive kicks to the left, thus failing
to save the ball that the fourth kicker again directed to the left.
Figure 2 shows overall goalkeeper data (red line) for percent-
age of dives in the opposite direction to the last kick. This per-
centage increased monotonically as run length increased and
then crossed the bootstrapped 95% condence interval after
three kicks in the same direction. Goalkeepers action choices
thereforeshowa patternsimilar tothe gamblers fallacy [10, 14].
Figure 3 indicates that kickers showed a rather different
sequential pattern. Kickers showed a trend to switch to a
new direction after a run of two kicks in the same direction
(p = 0.07). Importantly, this trend did not escalate further as
run lengths became longer: kickers were no more likely to
switch directions after a run of three than would be expected
by chance. Overall, kickers seem to show less predictable
behavioral sequences than goalkeepers, with a less obvious
gamblers fallacy pattern. One important reason for this differ-
ence couldbe that the kicks in penalty shootouts are generated
by multiple agents, who act relatively independently. In
contrast, the behavioral sequence of dives is generated by
the goalkeeper alone, so that cognitive limitations on an indi-
viduals random generation gure prominently in his action
decisions.
The goalkeepers choices could arise for two reasons. They
may believe that the sequence of kick directions is genuinely
random and simply be subject to the pervasive gamblers fal-
lacy [10, 14]. Alternatively, the goalkeepers behavior could
reect a cognitive hierarchy: the goalkeeper may believe
that the kickers will display a gamblers fallacy [18], though in
fact kickers do not. In any case, goalkeepers appear to behave
nonoptimally, sincetheyshowasequential biasnot alignedwith
kicker behavior. The goalkeepers sequential bias was not func-
tional, since the dive direction that they choseafter runs of three
repeatedkickdirectionswas nomorelikely thanchancetocoin-
cide with the actual direction of the kick on that attempt (same
direction 56.25%; p = 0.79, binomial distribution).
Analyses of gamblers fallacy assume that incrementing
runs of repeated events affect behavioral choices [19, 20].
However, other forms of serial dependence could also occur.
Lagged regression offers a general framework for considering
independent contributions of multiple previous events in pre-
dicting the current state. We therefore also applied discrete-
choice regression for unbalanced panel data to our data set.
This analysis approach has been popular in econometrics
(e.g., [21]) and also in behavioral economics, including game-
theoretic analyses of tournaments [8, 15]. We investigated
whether current dive direction depended on kicking direction,
and on dive direction one to three kicks previously (see
Supplemental Experimental Procedures for data, analysis,
and summary output). First, we used a random-effects model,
as recommended for short panels [8, 15]. The overall model
did not provide signicant evidence for serial dependence
(F
6,95
= 1.63, p = 0.15), consistent with a previous report [7].
Interestingly, the partial coefcients nevertheless showed a
signicant negative relation between current dive direction
and the kick direction three kicks previously (p = 0.04), after
the contribution of other events was taken into account. We
also tted an additional model using rst-difference estima-
tors; these have been preferred because they can be shown
to provide unbiased and consistent modeling that can account
for unobserved heterogeneity [22]. The overall rst-difference
model was signicant (p < 0.001). There was again a signicant
negative effect of the kick direction three kicks before the cur-
rent dive (p = 0.03). However the goalkeepers behavior now
showed a negative relation to their own previous dive direction
at lags one to three (all p < 0.01). Recent studies have pointed
to additional factors that may affect penalty shootout scores
as a result of their motivational signicance, notably kicking
order (rst versus second kicking team in the shootout) and
difference between the two teams scores immediately before
each kick [15]. Adding these motivational predictors did not
change the pattern of signicance for other terms in either
regression model.
Such lagged regressions do not consider effects of repeti-
tions within sequences and therefore provide a different
perspective from our runs analysis. However, the results are
in broad agreement with the runs analysis, in showing a serial
dependence on kick direction three kicks before the current
dive. They also suggest that goalkeepers may have a general
tendency to avoid repetitions in their sequential diving direc-
tion behavior, consistent with human behavior during random
sequence generation [23].
Kickers tend to use the inside of their dominant foot to make
contact with the ball during a penalty kick. This means that
right-footed and left-footed players nd it easier to kick to their
left and right, respectively [7, 8]. A goalkeeper may use infor-
mation about the kickers natural side in deciding which
way to dive. In our data set, goalkeepers dived to the kickers
natural side 178 times, against 143 dives to the nonnatural side
(p = 0.06, binomial test), suggesting that this information may
be important. We therefore repeated our analyses after recod-
ing all kick directions as natural or nonnatural for that
kicker. Interestingly, this analysis showed that goalkeepers
were signicantly more likely than chance to dive in the oppo-
site direction to the last kick after a run of one repeated kick di-
rection (p = 0.03) (i.e., they more likely to dive to the nonnatural
side of kicker 2 if kicker 1 had just kicked to their natural side,
and vice versa). However, goalkeepers did not use information
about natural kicking direction for longer runs. In fact, the
probability of switching dive direction relative to current kicker
preference (i.e., natural side) did not increase with run length
(see Figure S3). Goalkeepers therefore may use information
about the kickers preferences at the start of a run, but they
Figure 1. Four Consecutive Shots by Portuguese Kickers against the
English Goalkeeper during the Euro 2004 Seminal Penalty Shootout
The Portuguese kickers are shown in red and the English goalkeeper in blue;
the green circle is the ball. All four kicks were to the left of the goal. The goal-
keeper made a rightward dive following three consecutive kicks to the left.
Predictability in Football Penalty Shootouts
1919
do not appear to keep track of kicker-specic preferences
over extended sequences. To do so, the goalkeeper would
need reliable information about each preceding kickers
preferred foot and the corresponding actual ball direction.
This information would rapidly approach memory capacity.
Instead, goalkeepers initially take kicker preference into ac-
count in choosing dive direction but thereafter simply track se-
quences of kicks to left and right. The gamblers fallacy effect
identied in our main analysis is therefore based on accumu-
lating information about left and right ball directions without
comprehensive modeling of kicker preference.
The gamblers fallacy is well known in both game theory and
cognitive psychology. Perruchet and colleagues link the gam-
blers fallacy to conscious expectancy about the likely pattern
of events [19], while others suggest that the gamblers fallacy
arises because of a cognitive default based on sampling
without replacement [20]. Paradoxically, in simple laboratory
reaction tasks, the conscious expectancy that a long run will
shortly end seems to coexist with increasingly fast reaction
times [19]. Extending this idea to penalty shootouts, the goal-
keeper shown in Figure 1 might have dived very rapidly, and
thus had a particularly high chance of saving the fourth kick,
if the kick had in fact gone in the direction of his dive.
Previous game-theoretic studies of elite sports have not
specically considered the gamblers fallacy but have investi-
gated serial dependence of choices in general. In-match pen-
alty kicks did not show serial dependence [7, 8]. Kovash and
Levitt [24] found clear negative serial correlations in baseball
pitches and NFL football passes, though their data sets were
much larger than that used here. The gamblers fallacy is
conventionally measured as increasing perceptual expec-
tancy of a novel event with increasing run length [19]. Kovash
and Levitt instead reported lag-one correlations in production
of events. Nevertheless, their data are consistent with the
broad viewthat the gamblers fallacy may be an important fac-
tor limiting minimax play (i.e. an optimum strategy that mini-
mizes worst possible losses) in elite sports [7].
Since goalkeepers are carefully selected and highly trained,
one may ask why their vulnerability of regular behavioral se-
quences persists. Kickers may simply fail to detect, or other-
wise fail to exploit, the fundamental limitation in goalkeepers
random generation, which means that the vulnerability can
continue without being penalized. Goalkeepers may not be
aware of the vulnerability, perhaps because it has not been
prioritized in selection and training. In fact, penalty shootouts
are relatively rareour sample contained 168 matches that
could potentially have produced a penalty shootout, but only
37 were eventually settled in this way. Moreover, in-match
penalty kicks do not show regular patterns of kicking direction
[7, 8]. Therefore, the occurrence and implications of the gam-
blers fallacy in penalty shootouts may have gone unnoticed.
The penalty shootout pits the will of one goalkeeper against
the will of many kickers. The cognitive functions of monitoring
past performance and generating novel actions are highly rele-
vant to both parties. These cognitive functions have been
extensively studied in individuals, but little is known about
how they are coordinated across individuals within a group
[25]. We suggest that the goalkeepers and kickers perfor-
mance can be understood by considering how asymmetric
competition between a single individual (the goalkeeper) and
a group of individuals (the kickers) depends on behavior-moni-
toring and behavior-generating processes. For example, goal-
keepers may have complete autobiographical memory of the
sequence of kick directions, and of their own diving behavior,
because they face every kick. In contrast, kickers have auto-
biographical memory for their own kick only and may rely on
less-direct experience for their teammates kicks. Kickers
may therefore fail to monitor the goalkeepers diving behavior.
They may then fail to detect and exploit regular patterns of
goalkeeper dives.
The goalkeeper should have a monitoring advantage over
kickers, since self-related actions and material are better re-
tained than other classes of material [26, 27]. However, goal-
keepers have the disadvantage of producing predictable
sequential behavior, due to limitations in generating random
sequences. In contrast, the group of kickers may collectively
produce a more random sequence than any single individual,
because each kicker chooses their kick direction only once.
Distributed cognition across the group of kickers may increase
randomness. Similarly, other animals can produce stochastic
or random behavior at the group level [28, 29]. Thus, in sum-
mary, the individual goalkeeper has an advantage in the moni-
toring function but a disadvantage in random generation. The
groupof kickers, in contrast, has a disadvantage in monitoring,
but an advantage in random generation.
Game-theoretic analyses have shown the value of unpre-
dictable, mixed strategies in many games [30], and specically
in football penalty kicks during match play [7]. Experimental
studies conrm that departures from randomness can be
readily exploited [31, 32]. Penalty shootouts involve a unique
combination of a simple, competitive decision-making game
and a cognitive asymmetry between the individual goalkeeper
and the group of kickers. There has been relatively little
research on games in which the same individual competes
iteratively against a set of several individuals, each of whom
takes a turn, since most studies focus on two-player games.
However, cognitive hierarchy models [18] suggest that players
try to anticipate opponents decisions but fail to take into ac-
count the possibility that their opponents may also be doing
this as well as they are, or even better. Importantly, this stra-
tegic anticipation is known to be constrained by working mem-
ory capacity: the typical number of strategic thinking steps
Figure 2. Percentage of Goalkeeper Dives in the
Opposite Direction to the Last Ball Direction
following Runs of One, Two, or Three Repeated
Kicks in the Same Direction
Blue bars indicate 10,000 iterations of random
data across 20 bins. Dashed black lines are
one-tailed 95% condence intervals of the shuf-
ed distribution. Red lines indicate actual goal-
keeper data.
Current Biology Vol 24 No 16
1920
about opponents likely decisions is around 1.5 [18]. Our data
show that the kickers collectively lack the degree of strategic
thinking required to exploit goalkeeper behavior. Groups that
communicate freely generally reach better perceptual
decisions than isolated individuals [33], though it remains un-
clear whether the same group advantage applies in strategic,
competitive tasks involving random generation. Interestingly,
kickers in penalty shootouts typically act as a series of individ-
uals, with little intercommunication among the group during
the shootout. Better coordination of the kickers collective
cognitive capacities might in principle give them a strategic
advantage over the goalkeeper. Paradoxically, our results
suggest that the kickers should not monitor the goalkeepers
dive direction sequences. Rather, they should monitor previ-
ous kick directions and use these to simulate the goalkeepers
predictions, along the lines of the cognitive hierarchy model.
This is because the goalkeepers gamblers fallacy is driven
by previous kick directions. Thus, groups of kickers should
deploy a triad of advanced cognitive functions to exploit goal-
keeper bias: working memory for monitoring kick directions,
theory of mindfor estimating the goalkeepers likely dive direc-
tion, and communication to pool the groups cognitive
capacities.
In the penalty shootout, both fairness and success may
depend on the balance between the individual goalkeepers
ability in random generation and kickers collective ability to
predict the goalkeepers intention by monitoring sequential
behavioral patterns. Our research might motivate cognitive
training in random generation for goalkeepers, and in inter-
communicative monitoring and prediction for kickers. These
could help teams better prepare for nail-biting penalty shoot-
outs in future football tournaments.
Supplemental Information
Supplemental Information includes four gures, Supplemental Experi-
mental Procedures, and a data set used for lagged regression analyses
and can be found with this article online at http://dx.doi.org/10.1016/j.
cub.2014.07.013.
Acknowledgments
This work was supported by the Economic and Social Research Council
(ESRC, grant RES-062-23-2183). P.H. was additionally supported by an
ESRC Professorial Fellowship and by European Research Council
Advanced Grant HUMVOL (Human Volition, Agency and Responsibility).
We would additionally like to thank Professor Nick Chater, Dr. Nobuhiro
Hagura, Dr. Max-Philipp Stenner, and three anonymous reviewers for crit-
ical comments and revisions to the manuscript.
Received: March 18, 2014
Revised: June 5, 2014
Accepted: July 4, 2014
Published: July 31, 2014
References
1. Yarrow, K., Brown, P., and Krakauer, J.W. (2009). Inside the brain of an
elite athlete: the neural processes that support high achievement in
sports. Nat. Rev. Neurosci. 10, 585596.
2. BBC News (2004). Record audience for England match. June 25, 2004.
http://news.bbc.co.uk/2/hi/entertainment/3839067.stm.
3. Land, M.F., and McLeod, P. (2000). From eye movements to actions:
how batsmen hit the ball. Nat. Neurosci. 3, 13401345.
4. Savelsbergh, G.J.P., Williams, A.M., Van der Kamp, J., and Ward, P.
(2002). Visual search, anticipation and expertise in soccer goalkeepers.
J. Sports Sci. 20, 279287.
5. Savelsbergh, G.J.P., Van der Kamp, J., Williams, A.M., and Ward, P.
(2005). Anticipation and visual search behaviour in expert soccer goal-
keepers. Ergonomics 48, 16861697.
6. McMorris, T., and Hauxwell, B. (1997). Improving anticipation of goal-
keepers using video observation. In Science and Football III, T. Reilly, J.
Bangsbo, and M. Hughes, eds. (London: Taylor & Francis), pp. 290294.
7. Palacios-Huerta, I. (2003). Professionals play minimax. Rev. Econ. Stud.
70, 395415.
8. Chiappori, P.-A., Levitt, S., and Groseclose, T. (2002). Testing mixed-
strategy equilibria when players are heterogeneous: The case of penalty
kicks in soccer. Am. Econ. Rev. 92, 11381151.
9. Azar, O.H., and Bar-Eli, M. (2011). Do soccer players play the mixed-
strategy Nash equilibrium? Appl. Econ. 43, 35913601.
10. Tune, G.S. (1964). Response preferences: a review of some relevant
literature. Psychol. Bull. 61, 286302.
11. Wagenaar, W.A. (1972). Generation of random sequences by human
subjects: A critical survey of literature. Psychol. Bull. 77, 6572.
12. Gallagher, H.L., and Frith, C.D. (2003). Functional imaging of theory of
mind. Trends Cogn. Sci. 7, 7783.
13. Siegal, M., and Varley, R. (2002). Neural systems involved in theory of
mind. Nat. Rev. Neurosci. 3, 463471.
14. Tversky, A., andKahneman, D. (1971). Belief in the lawof small numbers.
Psychol. Bull. 76, 105110.
15. Apesteguia, J., and Palacios-Huerta, I. (2010). Psychological pressure in
competitive environments: Evidence from a randomized natural exper-
iment. Am. Econ. Rev. 100, 25482564.
16. Efron, B., and Tibshirani, R.J. (1994). An Introduction to the Bootstrap
(Boca Raton: CRC Press).
17. Wasserman, S., and Bockenholt, U. (1989). Bootstrapping: applications
to psychophysiology. Psychophysiology 26, 208221.
18. Camerer, C.F., Ho, T.-H., and Chong, J.-K. (2004). A cognitive hierarchy
model of games. Q. J. Econ. 119, 861898.
19. Perruchet, P., Cleeremans, A., and Destrebecqz, A. (2006). Dissociating
the effects of automatic activation and explicit expectancy on reaction
times in a simple associative learning task. J. Exp. Psychol. Learn. Mem.
Cogn. 32, 955965.
20. Ayton, P., andFischer, I. (2004). Thehot handfallacyandthegamblersfal-
lacy: two faces of subjective randomness? Mem. Cognit. 32, 13691378.
21. Arellano, M., and Honore , B. (2001). Chapter 53. Panel data
models: some recent developments. In Handbook of Econometrics
Volume 5, J.J. Heckman and E. Leamer, eds. (Amsterdam: Elsevier),
pp. 32293296. http://www.sciencedirect.com/science/article/pii/
S1573441201050061.
22. Wooldridge, J.M. (2002). Econometric Analysis of Cross Section and
Panel Data (Cambridge: MIT Press).
23. Spatt, J., and Goldenberg, G. (1993). Components of randomgeneration
by normal subjects and patients with dysexecutive syndrome. Brain
Cogn. 23, 231242.
Figure 3. Percentage of Kicker Shots in the
Opposite Direction to the Last Ball Direction
following Runs of One, Two, or Three Repeated
Kicks in the Same Direction
Blue bars indicate 10,000 iterations of random
data across 20 bins. Dashed black lines are
one-tailed 95% condence intervals of the shuf-
ed distribution. Green lines indicate actual
kicker data.
Predictability in Football Penalty Shootouts
1921
24. Kovash, K., and Levitt, S. (2009). Professionals Do Not Play Minimax:
Evidence from Major League Baseball and the National Football
League, NBER Working Paper 15347 (Cambridge: National Bureau of
Economic Research). http://www.nber.org/papers/w15347.
25. Adolphs, R. (2003). Cognitive neuroscience of human social behaviour.
Nat. Rev. Neurosci. 4, 165178.
26. Symons, C.S., and Johnson, B.T. (1997). The self-reference effect in
memory: a meta-analysis. Psychol. Bull. 121, 371394.
27. Macrae, C.N., Moran, J.M., Heatherton, T.F., Baneld, J.F., and Kelley,
W.M. (2004). Medial prefrontal activity predicts memory for self.
Cereb. Cortex 14, 647654.
28. Bednekoff, P.A., and Lima, S.L. (1998). Randomness, chaos and confu-
sion in the study of antipredator vigilance. Trends Ecol. Evol. 13, 284287.
29. Perony, N., Tessone, C.J., Ko nig, B., and Schweitzer, F. (2012). How
random is social behaviour? Disentangling social complexity through
the study of a wild house mouse population. PLoS Comput. Biol. 8,
e1002786.
30. Sato, Y., Akiyama, E., and Farmer, J.D. (2002). Chaos in learning a simple
two-person game. Proc. Natl. Acad. Sci. USA 99, 47484751.
31. Lee, D., Conroy, M.L., McGreevy, B.P., and Barraclough, D.J. (2004).
Reinforcement learning and decision making in monkeys during a
competitive game. Brain Res. Cogn. Brain Res. 22, 4558.
32. Vickery, T.J., Chun, M.M., and Lee, D. (2011). Ubiquity and specicity of
reinforcement signals throughout the human brain. Neuron 72, 166177.
33. Bahrami, B., Olsen, K., Latham, P.E., Roepstorff, A., Rees, G., and Frith,
C.D. (2010). Optimally interacting minds. Science 329, 10811085.
Current Biology Vol 24 No 16
1922
Current Biology, Volume 24
Supplemental Information
Asymmetric Predictability
and Cognitive Competition
in Football Penalty Shootouts
Erman Misirlisoy and Patrick Haggard
Supplemental Information

Supplemental figures
We performed an additional analysis using repetition lengths of goalkeeper dives on
consecutive penalties instead of repetition lengths of kick directions. In other words, we
examined how often goalkeepers dived in the opposite direction following 1, 2, or 3 previous
consecutive dives in the same direction. Similarly we examined how often kickers shot in the
opposite direction following 1, 2, or 3 previous consecutive dives in the same direction. The
data can be seen below in figures S1 and S2. Neither goalkeepers nor kickers showed a
pattern of behaviour significantly linked with runs of consecutive goalkeeper behaviour, as
none of the actual data crossed 95% confidence intervals for random data.
Figure S1. Percentage of dives in the opposite direction to last dive direction following runs
of 1, 2, or 3 repeated dives in the same direction. Blue bars indicate 10,000 iterations of
random data across 20 bins. Dotted black lines are one-tailed 95% confidence intervals of
the shuffled distribution. Red lines indicate actual goalkeeper data.


Figure S2. Percentage of shots in the opposite direction to last dive direction following runs
of 1, 2, or 3 repeated dives in the same direction. Blue bars indicate 10,000 iterations of
random data across 20 bins. Dotted black lines are one-tailed 95% confidence intervals of
the shuffled distribution. Green lines indicate actual kicker data.

We also investigated whether goalkeepers use information about natural kicking
directions for each kicker. Right-footed kickers prefer to kick towards their left while left-
footed kickers prefer their right. This is due to the tendency to kick with the inside of the
dominant foot. We found that goalkeepers dived to the natural side of the kicker 178 times
overall while they dived to the non-natural side 143 times overall (trend towards a difference
based on a binomial test, p = 0.06). We recoded the binary dataset for left/right kicking
directions into a new binary dataset representing natural/non-natural kicking directions. We
then ran the same bootstrap analysis as before based on runs of repeated events.
Figure S3 shows that goalkeepers were more likely than chance to dive in the opposite
direction to the last kick, following a run of 1 repetition of natural or non-natural kicks. This
effect reduced with longer runs of repetitions, suggesting that goalkeepers did not use
information about the natural sides of each kicker to drive longer runs of several choices. In
fact, the pattern shown in figure S3 is an inverse of the gamblers fallacy pattern shown in
main figure 2. The gamblers fallacy is defined as a tendency to expect a novel event which
grows with increasing run length. This comparison suggests that the gamblers fallacy
operates in the external spatial coordinates of left and right, rather than in the social
coordinates of estimating the kickers preferred action.
Figure S4 shows that kickers in fact showed no runs-based bias for kicking towards
natural or non-natural sides (compare with main figure 3).


Figure S3. Percentage of dives in the opposite direction to last ball direction following runs
of 1, 2, or 3 repeated kicks in the same direction. Direction is coded in terms of natural and
non-natural kicking directions depending on which foot the kicker uses. Blue bars indicate
10,000 iterations of random data across 20 bins. Dotted black lines are one-tailed 95%
confidence intervals of the shuffled distribution. Red lines indicate actual goalkeeper data.

Figure S4. Percentage of kicks in the opposite direction to last kick direction, following runs
of 1, 2, or 3 repeated kicks in the same direction. Direction is coded in terms of natural and
non-natural kicking directions depending on which foot the kicker uses. Blue bars indicate
10,000 iterations of random data across 20 bins. Dotted black lines are one-tailed 95%
confidence intervals of the shuffled distribution. Green lines indicate actual kicker data.





Supplemental Experimental Procedures
Section A
We examined online videos and statistics of all penalty shoot-outs from FIFA World Cup and
UEFA Euro finals tournaments between 1976 (when penalty shoot-outs were introduced into
the Euro finals tournament) and 2012. There were 361 penalty kicks in 37 shoot-outs. We
recorded the direction each ball was kicked (left/right/centre), and also the direction in which
the goalkeeper moved. Kicks to the goal centre were rare (9.14%), and goalkeepers rarely
remained in the centre (2.49% of dives), so only left/right choices were analysed. This left
321 penalties.
We analysed the data based on repetition lengths of kick directions from consecutive
kickers on a single team. When a left or right kick was the first in a sequence (e.g. the very
first penalty taken or a new direction chosen later in the sequence), we recorded whether the
goalkeeper opponent on the subsequent penalty dived in this same direction or in the opposite
direction. The same direction was coded as a stay while the opposite direction was coded as
a switch. This was repeated for kickers if the next kicker kicked in the same direction on
the subsequent penalty, this was coded as a stay while the opposite direction was coded as a
switch. The same process was repeated for a run of two repeated kick directions, in which
the kickers on a team consecutively kicked left twice in a row or right twice in row. The same
again applied for runs of 3 in which the kickers consecutively kicked in the same direction
three times in a row. The percentage of switches relative to stays was used in analysis and in
Figures 2 and 3 in the main report. Analyses used bootstrap resampling

[S1, S2] to compare
the actual data, to results obtained from the same data shuffled 10,000 times.

Section B

Time series regression approach to modelling goalkeeper dive direction.
We used Data Set S1 in this analysis. This contains 247 rows (one for each penalty kick
excluding the first kick of every penalty shootout as there are no preceding lags for this).
Data is read from the clipboard.

R code for panel data logit regression with random effects using plm package
(plm package is available at http://cran.r-project.org/web/packages/plm/index.html)

Explanation of variables:
currdivedir = current goalkeeper dive direction (0: goalkeepers right, 1: goalkeepers left)
kick.1 = kick direction 1 kick ago (0: goalkeepers right, 1: goalkeepers left)
kick.2 = kick direction 2 kicks ago (0: goalkeepers right, 1: goalkeepers left)
kick.3 = kick direction 3 kicks ago (0: goalkeepers right, 1: goalkeepers left)
dive.1 = dive direction 1 kick ago (0: goalkeepers right, 1: goalkeepers left)
dive.2 = dive direction 2 kicks ago (0: goalkeepers right, 1: goalkeepers left)
dive.3 = dive direction 3 kicks ago (0: goalkeepers right, 1: goalkeepers left)
shootout = penalty shootout identifier
kicknum = kick number in penalty shootout sequence

Code fragment:
library(plm)
data = read.delim("clipboard")
attach(data)
Y <- cbind(currdivedir)
X <- cbind(kick.1,kick.2,kick.3,dive.1,dive.2,dive.3)
pdata <- plm.data(data, index=c("shootout","kicknum"))
logit <- plm(Y~X,data=pdata,family=binomial('logit'),model="random")
summary(logit)

Output:
Oneway (individual) effect Random Effect Model
(Swamy-Arora's transformation)

Call:
plm(formula = Y ~ X, data = pdata, model = "random", family =
binomial("logit"))
Unbalanced Panel: n=59, T=1-6, N=102

Effects:
var std.dev share
idiosyncratic 0.196363 0.443128 0.953
individual 0.009758 0.098783 0.047
theta :
Min. 1st Qu. Median Mean 3rd Qu. Max.
0.02396 0.02396 0.04627 0.05028 0.06712 0.12230

Residuals :
Min. 1st Qu. Median Mean 3rd Qu. Max.
-0.79600 -0.43800 -0.09480 0.00005 0.44500 0.73900

Coefficients :
Estimate Std. Error t-value Pr(>|t|)
(Intercept) 0.782557 0.130744 5.9854 3.811e-08 ***
Xkick.1 -0.180560 0.102040 -1.7695 0.08002 .
Xkick.2 -0.023792 0.100421 -0.2369 0.81323
Xkick.3 -0.206217 0.100279 -2.0564 0.04248 *
Xdive.1 0.041726 0.100650 0.4146 0.67939
Xdive.2 -0.169427 0.099576 -1.7015 0.09212 .
Xdive.3 -0.119162 0.098696 -1.2074 0.23029
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1
Total Sum of Squares: 24.459
Residual Sum of Squares: 22.179
R-Squared : 0.09319
Adj. R-Squared : 0.086795
F-statistic: 1.62715 on 6 and 95 DF, p-value: 0.14804

R code for panel data logit regression with first difference estimator using plm package

Code fragment:
library(plm)
data = read.delim("clipboard")
attach(data)
Y <- cbind(currdivedir)
X <- cbind(kick.1,kick.2,kick.3,dive.1,dive.2,dive.3)
pdata <- plm.data(data, index=c("shootout","kicknum"))
logit <- plm(Y~X,data=pdata,family=binomial('logit'),model="fd")
summary(logit)

Output:
Oneway (individual) effect First-Difference Model

Call:
plm(formula = Y ~ X, data = pdata, model = "fd", family =
binomial("logit"))

Unbalanced Panel: n=59, T=1-6, N=102

Residuals :
Min. 1st Qu. Median 3rd Qu. Max.
-0.930 -0.430 0.053 0.368 1.200

Coefficients :
Estimate Std. Error t-value Pr(>|t|)
(intercept) 0.070923 0.091991 0.7710 0.445751
Xkick.1 -0.305010 0.162295 -1.8793 0.068311 .
Xkick.2 -0.223139 0.166961 -1.3365 0.189779
Xkick.3 -0.364524 0.163869 -2.2245 0.032474 *
Xdive.1 -0.778950 0.160577 -4.8509 2.367e-05 ***
Xdive.2 -0.528189 0.180080 -2.9331 0.005806 **
Xdive.3 -0.429678 0.144687 -2.9697 0.005279 **
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Total Sum of Squares: 22.977
Residual Sum of Squares: 11.728
R-Squared : 0.48957
Adj. R-Squared : 0.40987
F-statistic: 5.75475 on 6 and 36 DF, p-value: 0.00028101








Supplemental references
S1. Efron, B., & Tibshirani, R. J. (1994). An Introduction to the Bootstrap. CRC Press.

S2. Wasserman, S., & Bockenholt, U. (1989). Bootstrapping: Applications to
Psychophysiology. Psychophysiology, 26(2), 208221. doi:10.1111/j.1469-
8986.1989.tb03159.x