See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/242570262

Ice Hockey

Article · January 2010

CITATIONS READS
0 1,811

1 author:

Armann Ingolfsson
University of Alberta
68 PUBLICATIONS 3,143 CITATIONS

SEE PROFILE

All content following this page was uploaded by Armann Ingolfsson on 31 August 2015.

The user has requested enhancement of the downloaded file.

 Wiley Encyclopedia of Operations Research and Management Science

Ice Hockey
Fo
Wiley Encyclopedia of Operations Research and Management
Journal:
Science

Manuscript ID: EORMS-09-0077.R1

r

Wiley - Manuscript type: Introductory

Pe

Date Submitted by the

Author:

Complete List of Authors: Ingolfsson, Armann; University of Alberta, School of Business

er

ice hockey, OR in sports, statistical analysis, Poisson process,

Keywords:
Monte Carlo simulation
Re
vi
ew

John Wiley & Sons

Page 1 of 17 Wiley Encyclopedia of Operations Research and Management Science

1
2
3
4 Notes:
5
6 • Suggested cross-references are underlined
7
8
9
10
11 Ice Hockey
12
13 Armann Ingolfsson
14 School of Business, University of Alberta
15 Edmonton, Alberta T6G 2R6, Canada
16 [email protected]
17
18 Abstract: I describe the game of ice hockey and the use of data analysis and
Fo
19 mathematical modeling to answer questions of interest to ice hockey fans, coaches, and
20
managers. Most of the discussion is in the context of the National Hockey League
21
22 (NHL), the primary professional ice hockey league. I pay particular attention to two
r
23 questions: when to pull the goalie, and how to estimate the probability that a particular
24 NHL team will qualify for the playoffs.
Pe

25
26 Keywords: ice hockey, OR in sports, statistical analysis, Poisson process, Monte Carlo
27
28 simulation, pulling the goalie, playoff qualification
er

29
30
31 Ice hockey is a popular team sport, particularly in cold climates such as in Canada, the
32 northern part of the USA, Russia, and northern Europe. The primary professional ice
Re

33 hockey league is the National Hockey League (NHL), which operates in USA and
34
35
Canada and draws players from throughout the world. As is the case with baseball,
36 football, soccer, and basketball, various operations research methods and statistical
analyses have been used to address questions of interest to ice hockey fans, coaches, and
vi

37
38 managers. This article outlines published work on such questions and elaborates on two
39 issues: (1) determining the optimal time to “pull the goalie” near the end of an ice hockey
ew

40 game and (2) estimating the probability that a team will qualify for the NHL playoffs.
41
42 Rules vary among ice hockey leagues, but all the descriptions and empirical results
43 described in this article refer to the NHL, unless otherwise noted.
44
An ice hockey team normally has six players on the ice during a game: a goaltender
45
46 (“goalie”), two defensemen, and three forwards (center, right wing, and left wing).
47 Players rotate onto the ice in shifts of about one minute. The players use sticks to propel
48 the “puck” (a black rubber disc), with the aim of scoring goals by shooting the puck into
49 the opposing team’s goal. Penalties to players for rule infractions are common (an
50 average 2007-8 NHL game had 27.3 penalty minutes), resulting the penalized team
51
52
playing “shorthanded”; the opposing team is said to be on the “power play”. A game lasts
53 for 60 minutes, comprising three 20-minute periods. During the NHL's regular season of
54 82 games for each team, if the score is tied after three periods (so-called “regulation
55 time”), the teams play one overtime period that lasts for five minutes or until one of the
56 teams scores. If the score remains tied after the five-minute overtime period, the game
57
outcome is determined by a “shootout,” during which a chosen set of players from the
58
59
60
John Wiley & Sons
 Wiley Encyclopedia of Operations Research and Management Science Page 2 of 17

1
2
3
two teams alternate in taking penalty shots (where a single skater attempts to score on an
4
5 undefended goalie). The overtime rules change during the NHL playoffs in that overtime
6 continues, in 20-minute periods, until one team scores a goal. During the regular season,
7 a win is worth two points and an overtime loss is worth one point in the standings. A loss
8 during regulation time is worth no points.
9
10 The NHL has 30 teams, divided into two 15-team Conferences, each of which is
11 subdivided into three five-team Divisions. The top eight teams from each Conference
12 advance to the playoffs, in which pairs of teams play a series of a maximum of seven
13
games, with the first team to win four games advancing to the next round of the playoffs.
14
15 The top team from each Conference plays in the final series, and the winning team is
16 awarded the Stanley Cup.
17
18
Researchers from various fields, including operations research, statistics, economics, and
management strategy, have developed mathematical models of ice hockey. Although I
Fo
19
20 focus on operations research models, I touch on contributions from these other fields.
21 The interplay between statistical analysis and decision modeling is important. For
22 example, the utility of models to choose an optimal player order for a shootout (for
r
23 example, see [1]) depends on how precisely the shootout scoring probabilities of different
24
players are known.
Pe

25
26
27 Poisson Goal Scoring Processes
28
er

29
30 Perhaps the most important and useful modeling device for understanding the dynamic of
31 hockey games is the assumption that scoring follows a Poisson process, an assumption
32 that can be justified by taking a first-principles approach. Call the teams in a hockey
Re

33 game A and B and imagine dividing the 60 minutes of regulation play into intervals of,
34
35 say, ∆t = 30 seconds. Assume that during each interval, Team A scores with probability
36 α ∆t and does not score with probability 1 – α ∆t, independent of scoring in other
vi

37 intervals and independent of whether or not Team B scores in the interval. The parameter
38 α is Team A’s goal-scoring rate. Given these assumptions, NA, the number of goals that
39
ew

40 Team A scores in the game, will follow a binomial distribution with mean αT, where T is
41 the length of the game. Letting ∆t approach zero, the binomial distribution approaches a
42 Poisson distribution with mean αT (for example, see [2, Section 5.2]). A symmetric set
43 of assumptions for Team B implies that its scoring will follow a Poisson process with rate
44
45 β, which implies that, NB, the number of goals that Team B scores in the game, has a
46 Poisson distribution with rate βT. Thus, we have reduced the dynamics of the game to
47 two independent and competing Poisson processes.
48
49 Before scrutinizing the validity of the assumptions that led to the Poisson processes, I
50 illustrate their usefulness by demonstrating how one can compute the probability that
51 Team A wins the game during “regular time,” i.e., before overtime. (Reference [3]
52 describes an alternative approach, based on the same assumptions.) Starting from the law
53
of total probability and the property that the superposition of two independent Poisson
54
55 processes is a Poisson process, we see that
56
57
58
59
60
John Wiley & Sons
Page 3 of 17 Wiley Encyclopedia of Operations Research and Management Science

1
2
3
4
∞
((α + β)T ) n e − (α +β)T
(1) Pr{ A wins} = Pr{N A > N B } = ∑ Pr{N A > N B | N A + N B = n} .
5 n!
n =1
6
7
8 Now, let us view NA as being obtained from NA + NB by thinning; that is, whenever a goal
9
10 is scored, Team A scored the goal with probability p = α/(α + β), independent of all other
11 goals. This implies that if the total number of goals in the game is n, then the number of
12 goals scored by Team A follows a binomial distribution with parameters n and p.
13 Therefore, the term Pr{NA > NB | NA + NB = n} equals the probability that a binomial
14 random variable with parameters n and p is larger than n/2, which is easily calculated.
15
16
We approximate the infinite series in (1) with the sum of the first m terms. The resulting
∞
17
18
truncation error is bounded by ∑n=m+1 ((α + β)T ) n exp(−(α + β)T ) / n! —that is, by the
complementary distribution function for a Poisson random variable with mean (α + β)T.
Fo
19
20 To estimate how many terms are needed for an accurate approximation, suppose that (α +
21 β)T, which is the average number of goals per game, equals 6 (which is roughly
22
consistent with the last few NHL seasons). With this goals-per-game average, truncating
r
23
24 the infinite series after 20 terms results in a truncation error of about 10-6.
Pe

25 Near the end of the NHL’s 2008-09 regular season, the Detroit Red Wings were the top
26
27 scoring team, with a goal scoring average of 3.63 per game, followed by the Boston
28 Bruins, with an average of 3.28 per game. The lowest scoring team, the Phoenix
er

29 Coyotes, averaged 2.43 goals per game. The calculation I have outlined results in the
30 Red Wings having a 0.48 probability of winning a game against the Bruins in regular
31 time (with a 0.37 probability of a loss and 0.15 probability of a tie) and a 0.61 probability
32
of winning a game against the Coyotes in regular time (with a 0.24 probability of a loss
Re

33
34 and 0.15 probability of a tie). Table 1 illustrates the calculation for the Red Wings vs. the
35 Bruins. To refine this analysis, instead of using average goal scoring rates against all
36 teams, one could attempt to classify teams based on ability and estimate average goal
vi

37 scoring rates for each team as a function of the opposing team’s ability.
38
39 Let us now revisit the assumptions that led us to model scoring by Teams A and B as
ew

40 independent Poisson processes. References [3] and [4] demonstrate empirically that the
41 Poisson approximation is fairly accurate. Two assumptions are particularly suspect, if
42
taken literally. First, there are various reasons to expect that the goal-scoring rate of one
43
44 team may depend on the goal-scoring rate of the other team. For example, one could
45 expect that if Team A scores the first goal, then Team B will respond by “trying harder”
46 and its goal-scoring rate will increase. Such hypotheses say nothing, however, about
47 whether or not the dependence between the two processes is large enough to cause
48 concern. This approximation appears to be defensible, as the only study that, to my
49
50
knowledge, has investigated this issue [3] found no statistical evidence against the
51 independence assumption. Second, one would expect that a certain minimum time must
52 elapse between one goal and the next, thus violating the assumption that the probability
53 of a goal in the next infinitesimal interval does not depend on when the last goal was
54 scored. Reference [5] investigated this issue in detail, and found that a generalization of
55
the Poisson process, whereby the goal-scoring rate increases gradually from zero to a
56
57 constant rate as time from the last goal increases, provides a good fit to empirical data on
58
59
60
John Wiley & Sons
 Wiley Encyclopedia of Operations Research and Management Science Page 4 of 17

1
2
3
the times at which goals were scored in the NHL. Based on the analysis in [5], the
4
5 “warm-up” time required before the goal-scoring rate reaches the constant value appears
6 to be approximately 20 seconds. Although the model in [5] is undoubtedly more realistic,
7 the Poisson model provides a reasonable and relatively tractable first approximation to
8 the dynamics of a hockey game.
9
10 Goal scoring by individual players has also been modeled as a Poisson process. For
11 example, [6] and [7] demonstrate that the distribution of the number of points (goals plus
12 assists) per game for former NHL player Wayne Gretzky—the “Great One”—is
13
remarkably well approximated by a Poisson distribution, with a mean of 2.39 points per
14
15 game.
16 The frequency of “hat tricks” (a player scoring three or more goals in a game) is another
17
18
measure that can be understood in terms of a Poisson process model. In the fall of 2008,
Teemu Selanne scored his 21st career hat trick. This was described as a “statistical
Fo
19
20 oddity” [8] by one sports writer, because Selanne has scored more hat tricks in 16 seasons
21 than Gordie Howe, the NHL’s most durable player, had scored in 26 seasons, while
22 scoring fewer total goals than Howe did. Is Selanne’s achievement unusual, from a
r
23 statistical point of view? Let’s use the Poisson model try to answer that question. At the
24
time, Selanne had played 1080 regular season NHL games and had scored 560 goals, for
Pe

25
26 a 0.52 goals per game average. If Selanne’s goal production follows a Poisson process
27 with a constant rate of 0.52 goals per game, then we can compute the probability that he
28 scores zero goals (exp(-0.52) = 0.60), one goal (0.52 exp(-0.52) = 0.31), two goals (0.522
er

29 exp(-0.52)/2 = 0.08), or three or more goals (1 – 0.60 – 0.31 – 0.08 = 0.016) in a game.
30
31
Using the 0.016 estimate of Selanne’s hat trick probability, we can estimate an expected
32 number of hat tricks for him to be 1080 × 0.016 = 17.10. Because a Poisson
Re

33 distribution’s variance equals the mean, the standard deviation equals 17.10 = 4.13 ; if
34 the Poisson model is accurate, therefore, we would expect Selanne’s actual number of hat
35
36 tricks (21) to have a 68% probability (based on a normal distribution approximation) of
falling within 17.10 ± 4.13, as indeed it does, albeit at the high end of the range. From
vi

37
38 this perspective, Selanne's achievement is close to what one would expect, given his total
39 number of games played and total number of goals scored. If one does the same analysis
ew

40 for Howe, one obtains an expected value of 19.60 total hat tricks with a standard
41
42
deviation of 4.43. Howe’s actual number of hat tricks was 19—very close to the
43 expected value. Thus, the reasoning that players with a higher total number of goals and
44 higher total number of seasons are likely to have a higher total number of hat tricks is
45 false. In order to be likely to score hat tricks, a player must have a high goals-per-game
46 average, and this matters more than the total number of goals.
47
48 If one repeats the hat trick analysis for several other players that were mentioned in [8],
49 two stand out as having had a significantly greater number of hat tricks than expected:
50 Wayne Gretzky (50 actual, 34.55 expected) and Glenn Anderson (21 actual, 11.64
51
52
expected). The players mentioned in [8] are primarily ones that are at the top of the list
53 of hat trick scorers, so these players do not constitute a random sample of NHL players.
54 For that reason alone, one would expect the number of hat tricks for this group of players
55 to be biased upwards, a point on which [9] elaborates. However, there are other factors
56 involved. Let’s take a closer look at Gretzky. His expected number of hat tricks is 34.55
57 but he actually had 50 hat tricks, which is well outside the standard deviation of 5.88.
58
59
60
John Wiley & Sons
Page 5 of 17 Wiley Encyclopedia of Operations Research and Management Science

1
2
3
One of many unusual things about Gretzky is how much his goals-per-game average
4
5 changed during his career, from 1.15 in the 1981-82 season to 0.13 in the 1999-2000
6 season. Computing the expected number of hat tricks separately for every season (which
7 corresponds to assuming that Gretzky's goal-scoring process was a Poisson process with
8 a time-varying rate) should result in a better estimate. Adding the expected number of
9 hat tricks for each season results in an expected value of 47.6 career hat tricks, with a
10
11
standard deviation of 6.91. With this approach, Gretzky's 50 actual career hat tricks are
12 well within one standard deviation of the expected value.
13
14 Empirical Evidence about Ice Hockey
15
16 A number of other questions about ice hockey have been investigated empirically,
17 including:
18
• Can a team’s win/loss percentage be predicted from the number of goals, for (GF)
Fo
19
20 and against (GA)? Baseball statistician Bill James [10] found that for major
21 league baseball, teams’ win/loss percentages were well approximated by (runs
22 scored)2/((runs scored)2 + (runs allowed)2) and he referred to this as the
r
23 Pythagorean method. In an attempt to develop a similar formula for ice hockey,
24
[11] found that NHL teams’ regulation time win/loss percentages over a season
Pe

25
26 (adjusted for ties) were well approximated by GF1.927/(GF1.927 + GA1.927).
27
28 • How likely is a player to score on a penalty shot during a shootout? According to
[12], the overall average scoring probability is 0.33. The probability is only
er

29
30 slightly higher (0.34) for the first six shooters (who were presumably chosen for
31 their scoring ability) than the later shooters (0.31). According to the same
32 reference, scoring probabilities differ only marginally, depending on the
Re

33
importance of the shot (“must make” vs. “could win” vs. “no pressure”).
34
35 • The longest shootout in NHL history involved 30 shooters, in a 2005 game
36 between the New York Rangers and the Washington Capitals. How long will we
vi

37
38 have to wait for a 40-shooter shootout? According to estimates in [12], 40-
39 shooter shootouts can be expected to occur once every 156 seasons, on average.
ew

40
41 • More than half the regular season games that reach overtime remain tied at the
42 end of overtime and are therefore determined by a shootout. Is the overtime
43 period too short? According to estimates in [13], if the overtime period were
44 extended from 5 to 10 minutes, then the fraction of overtime games that are
45 decided by a shootout would drop from 0.53 to 0.28.
46
47 • If a team is leading after two periods, how likely is it that it will win the game?
48 According to [4], the answer is 0.72 if the team is playing at home and is leading
49
by one goal, 0.92 if it is leading by two goals, and 0.98 if leading by three goals.
50
51 The probabilities are slightly lower for teams playing away from home.
52 • A related question is “when is a goal most valuable?” Or, in other words, under
53
54
what circumstances is goal scoring most likely to increase the probability of
55 winning the game? According to [5], not surprisingly, a goal is most valuable if
56 the score is close to being tied and little time is left in the game. As an example,
57 [5] estimates that a game-tying goal with 5 minutes remaining in regulation time
58
59
60
John Wiley & Sons
 Wiley Encyclopedia of Operations Research and Management Science Page 6 of 17

1
2
3
increases the probability of winning the game for the scoring team from 0.10 to
4
5 0.40.
6 • Are referees biased, or as [14] puts the opposite possibility, are referees Markov
7
8 (that is, “memoryless”)? These authors find that the home team receives fewer
9 penalties on average, a finding that is consistent with similar analyses for soccer.
10 (Instead of referee bias, a possible alternative explanation that has not been
11 investigated to my knowledge, is that the home team commits fewer penalties, on
12 average.) A second type of bias that [14] investigated is “follow-up bias”, in
13
14
which a referee that penalized Team A last is more likely to penalize Team B
15 next. According to estimates in [14], if the home team gets the first penalty, the
16 probability that the away team gets the second penalty is 0.62. If the home team
17 receives the first two penalties, the probability that the away team gets the third
18 penalty is 0.65.
Fo
19
20 • If Wayne Gretzky and Mario Lemieux had reached their peak in the same year,
21 and played on equally strong teams, who would have scored more points?
22 Reference [7] develops a “statistical time machine” to answer such questions and
r
23
24
it gives the edge to Lemieux.
Pe

25 • Is the “hot goalie” a myth, as [15] argues is the case for the “hot hand” in
26
27
basketball? Reference [16] examines the performance of the Detroit Red Wings
28 in their four-game sweep of the Stanley Cup final series in 1997, with Mike
er

29 Vernon in goal, and concludes that the “hot goalie hypothesis is alive and well.”
30
31 • When a team performs poorly, a typical reaction is to fire the coach. Does firing
32 the coach work? Management strategists have investigated team leadership in the
Re

33 context of NHL hockey, particularly for coach succession. According to [17] and
34 [18], when a coaching change occurs between seasons, the team’s performance
35 remains about the same, on average. When a coach is replaced during a season,
36
the team performance worsens, on average, during the remainder of the season. If
vi

37
38 a coach taking over during a season continues to coach in the following season,
39 the team’s performance improves, on average during the second season. The
ew

40 explanation offered by [17] is that a coaching change during the season gives the
41 new coach an opportunity to evaluate players, make roster changes, and influence
42
player training during the offseason, all of which result in improved performance
43
44 during the next season.
45 • Does awarding one point for an overtime loss cause teams to play more
46
47 defensively if they are tied near the end of regulation time? A series of articles by
48 economists on this topic [19-22] uses models of goal scoring (similar to our
49 Poisson process assumptions) and empirical data to suggest that the answer is yes,
50 and that this phenomenon has increased the frequency of overtime games.
51
52 • What determines player salaries? There is a sizable literature in sports economics
53 (see [23] for a review) that attempts to relate player salaries to such factors as
54 market structure, nationality, and player type. For example, [24] found that NHL
55
players from Sweden and Finland were higher-paid, on average, than players from
56
57 Canada, the US, Russia, and the former Czechoslovakia, perhaps because
58
59
60
John Wiley & Sons
Page 7 of 17 Wiley Encyclopedia of Operations Research and Management Science

1
2
3
Swedish and Finnish players have relatively more attractive options for playing in
4
5 their home countries rather than the NHL. As another example, after the short-
6 lived World Hockey Association (WHA; an NHL competitor) folded, [25]
7 estimated NHL players’ contribution to team revenue and compared them to
8 players’ salaries. They found that, on average, player salaries exceeded their
9 contribution to team revenue—a result that is consistent with the NHL having
10
11
increased player salaries beyond the point of profitability, in order to drive the
12 WHA out of business.
13
14 When to Pull the Goalie
15
16 Moving on to decision models, operations researchers have studied one decision
17 particularly intensely: When should a coach “pull the goalie?” If a hockey team is down
18 by one goal near the end of a game, the coach will often replace the goalie with another
Fo
19 player, so that the team has six skating players against the opponent’s five. Pulling the
20
goalie increases the team’s goal-scoring rate, thus increasing the probability that the team
21
22 will tie the game. The opponent’s goal scoring rate is likely to increase even more, thus
r
23 increasing the probability that the team loses by more than one goal, but in terms of
24 points in the standings, a loss is worth zero points, regardless of how many goals separate
Pe

25 the teams. (Although goal differential could matter at the end of the season, as it is used
26 as a tiebreaker in determining which teams enter the playoffs.) Therefore, when the
27
28
probability of a win is close to zero, a strategy that increases the probability of a tie and
thus entering overtime is worthwhile, even if the strategy also increases the probability of
er

29
30 losing by a larger margin.
31
32 How do NHL coaches decide when to pull the goalie? Former Los Angeles Kings Head
Re

33 Coach, Andy Murray, was quoted in 1994 [26] as saying “if you’re down by one goal,
34 you’re looking at the one minute mark.” The “one minute when trailing by one” rule is
35 also mentioned in [27-31], all of whom analyze the decision mathematically and
36 conclude that it would be better to pull the goalie more than one minute before the end of
vi

37
38
the game. Recall our assumption that scoring by the two teams can be modeled as
39 independent Poisson processes with rates α and β. If Team A pulls its goalie, suppose
ew

40 that its goal-scoring rate increases from α to α′ and Team B’s goal-scoring rate increases
41 from β to β′ , and further assume that β′ > α′. With T minutes remaining in the game and
42
43
Team A trailing by one goal, suppose that Team A decides to pull its goalie when t < T
44 minutes are left, if it is still trailing by one goal at that point. This strategy will result in a
45 tie if Team A scores the next goal. Following references [28-30], I computed the
46 probability that Team A scores the next goal separately for two cases: (1) the next goal is
47 scored in the interval from T to t minutes and (2) the next goal is scored in the last t
48 minutes. Under case (1), Team A scores the next goal if the following two events occur:
49
50 • The teams score at least one goal in the interval from T to t (with probability
51 1 − exp(−(α + β)(T − t ) ))
52
53 • Conditional on the preceding event, Team A scores the first goal (with probability
54
55
α /(α + β) )
56
Under case (2), Team A scores the next goal if the following three events occur:
57
58
59
60
John Wiley & Sons
 Wiley Encyclopedia of Operations Research and Management Science Page 8 of 17

1
2
3
4
• Neither team scores in the interval from T to t minutes (with probability
5 exp(−(α + β)(T − t )) )
6
7 • The teams score at least one goal in the last t minutes (with probability
8 1 − exp(−(α ′ + β′)t )
9
10 • Conditional on the preceding event, Team A scores the first goal (with probability
11 α ′ /(α ′ + β′) )
12
13 By combining terms, we obtain the probability that Team A obtains a tie sometime before
14 the game ends, as a function of T and t:
15
16
17
(
(2) f (T , t ) = 1 − e −(α +β)(T −t ) ) α α+ β + e − ( α +β )(T −t )
(1 − e − ( α′+β′)t
) α′α+′β′
18
Fo
19 One obtains the optimal time to pull the goalie by maximizing f with respect to t. One
20 can find the optimal time t* analytically [28] and t* has the desirable property that it does
21
22
not depend on T, but it is perhaps more insightful to plot how f depends on t for a given T.
As an example, suppose that in a game between the Red Wings and the Coyotes, the Red
r
23
24 Wings are trailing by one goal with T = 5 minutes remaining. Figure 1 shows how the
Pe

25 probability that the Red Wings will tie the game before it ends (that is, f(5, t)) depends on
26 the time t at which the Red Wings coach pulls the goalie, if the Red Wings have not tied
27 it already. The probability of a tie is computed using (2) and the goal-scoring rate
28
estimates discussed earlier. Pulling the goalie with one minute remaining results in a tie
er

29
30 probability of 0.279. The optimal time to pull the goalie is considerably earlier than one
31 minute before the end of the game: approximately 2.5 minutes. By pulling the goalie
32 earlier, the probability of a tie increases from 0.279 to 0.295.
Re

33
34 The models and the goal-scoring rate estimates in [27-31] differ, but all of these papers
35 reach the same conclusion: it is optimal to pull the goalie with more than one minute
36 remaining (the estimates of t* in these papers range from 1.3 to 4 minutes). As one
vi

37
would expect, strong teams (with high α) should pull their goalie sooner than weak teams
38
39 do, and teams playing against a strong opponent (with high β) should pull their goalie
ew

40 later than should teams playing a weak opponent [29].

41
42 Will my Team Qualify for the Playoffs?
43
44 As of 20 March 2009, the Edmonton Oilers had played 70 of their 82 regular season
45 games and had earned 77 points, putting them in seventh place in the NHL’s Western
46 Conference. As is usually the case at this time of year, for me and other Oilers fans there
47
was considerable uncertainty about whether or not the Oilers will be one of the eight
48
49 Western Conference teams to earn a playoff spot. Simulation and optimization can
50 provide different and complementary answers to this question. Optimization (for
51 example, see [32-33]) can sometimes provide a definite “yes” or “no” answer, whereas
52 simulation (see [34]) can provide an estimate of the probability that the answer will be
53 “yes.” Simulation can also provide additional information—about how many points will
54
55
most likely be needed to qualify for the playoffs, for example, or the probability that a
56 team will qualify, given that it earns X points in its last Y games.
57
58
59
60
John Wiley & Sons
Page 9 of 17 Wiley Encyclopedia of Operations Research and Management Science

1
2
3
I now describe the Monte Carlo simulation model in [34] in greater detail. The web site
4
5 http://www.sportsclubstats.com/NHL.html shows the results of a similar simulation
6 model. The building block of the model is the simulation of the outcome of every
7 remaining regular season game. To simulate the outcome of a game between Teams A
8 and B, one requires as input the probability that Team A will win the game. A simple
9 approach, adopted in [34], is to use the average of Team A’s winning percentage and
10
11
Team B’s losing percentage at that point in the season ([34] provides additional details,
12 including how to simulate overtime games). Alternatively, one could use the approach
13 described at the beginning of this article, an approach that involves computing win
14 probabilities based on the team’s goal-scoring rates and Poisson assumptions. Another
15 option is a statistical model that estimates win probabilities by incorporating such factors
16
as home-ice advantage, offensive capability, defensive capability, and player injuries.
17
18 In each replication of the simulation, one simulates every remaining regular season game.
Fo
19 Then, one computes the simulated standings at the end of the season, according to the
20
21
NHL’s ranking criteria, which are the following, in decreasing order of priority:
22 • Top ranking within a Division guarantees a top three standing in the Conference
r
23
24 • Total points
Pe

25
26 • Total wins
27
28 • Results of head-to-head games for teams that are tied based on the first three
criteria
er

29
30
31 • Total goals for (GF) minus total goals against (GA)
32 Interestingly, before 1970, the NHL used GF instead of GF – GA as the final criterion; as
Re

33
discussed in [35], this led to perverse incentives near the end of the 1969-1970 regular
34
35 season. In the Montreal Canadiens’ final game that season, trailing 5-2 against the
36 Chicago Blackhawks, the Canadiens could have qualified for the playoffs if they had
vi

37 scored three more goals, regardless of how many more goals the Blackhawks scored.
38 The Canadiens followed the rational strategy, given the ranking criteria, of pulling their
39 goalie at the beginning of the third period—but without success. Reference [35] uses
ew

40
41
regression analysis to argue that GF – GA is a better tie breaker, and the NHL has since
42 adopted this approach.
43 Simulation has an advantage over optimization in handling arbitrarily complex ranking
44
45 criteria. The approach used in [34] is to rank teams based on “pseudo points,” computed
46 as:
47
(3) Pseudo points = M(1) × (Criterion 1) + … + M(n) × (Criterion n),
48
49 where M(1) >> … >> M(n) > 0, in order to enforce the strict prioritization of the factors.
50 Put differently, the teams are ordered lexicographically, based on the ranking criteria. If
51
52
some of the criteria are difficult to implement in the simulation model (for example, the
53 results of head-to-head games) and are unlikely to make a difference, then one can either
54 ignore them (and eliminate replications where those criteria made a difference) or replace
55 those criteria with a random number in order to break ties that remain after applying the
56 criteria that are explicitly included in the simulation.
57
58
59
60
John Wiley & Sons
 Wiley Encyclopedia of Operations Research and Management Science Page 10 of 17

1
2
3
Table 2 shows an example of simulation results for the NHL’s Western Conference, as of
4
5 19 March 2009 and based on 5,000 replications, focusing on the Edmonton Oilers. Each
6 row in the table corresponds to one replication. Based on these replications, one can
7 estimate the probability that the Oilers qualify for the playoffs (0.624) and the average
8 number of points needed to qualify (88.9), by computing simple averages.
9
10 I show two examples to illustrate how one can dig deeper into the simulation results to
11 answer interesting questions. First, how does the probability that the Oilers qualify for
12 the playoffs depend on the number of points they earn from their remaining 12 games?
13
Table 3 shows a cross-tabulation obtained from the simulation results and it answers this
14
15 question (there were no replications in which the Oilers earned 0, 23, or 24 points in their
16 final 12 games). If Edmonton were to earn 12 of the 24 available remaining points, their
17 probability of qualifying would be 0.73; to bring the probability to 0.99, they would need
18 to earn 14 of the 24 available points.
Fo
19
20 Second, sports writers sometimes use a two-step process to estimate how many points a
21 team will need to achieve to qualify for the playoffs: (1) estimate the number of points X
22 that will be needed to qualify and (2) subtract the team’s current number of points from
r
23 X. Implicit in this approach is an assumption that the random variable X is independent
24
of whether or not the team of interest qualifies. Put in more concrete terms, the question
Pe

25
26 is whether the distribution of the minimum number of points needed to qualify for the
27 playoffs depends on whether the Edmonton Oilers qualify. By cross-tabulating the
28 simulated values of X against an indicator variable for whether or not Edmonton
er

29 qualifies, one can produce a graph like the one shown in Figure 2. As that figure shows,
30
31
more points are needed to qualify, on average, if Edmonton qualifies than if it does not.
32 To understand this outcome, keep in mind that in the current standings, Edmonton is in
Re

33 the top 8. If Edmonton ends up not qualifying, it means that the Oilers performed
34 relatively poorly during the remainder of the season, which would tend to reduce the
35 qualification threshold for teams that are below the Oilers in the current standings.
36
vi

37 In conclusion, operations research and statistical analyses can provide insight and
38 predictions about game strategies, whether a team will make the playoffs, the
39 consequences of rule changes, and a variety of other issues of interest to ice hockey
ew

40
sportswriters, analysts, and fans. I hope that this article will stimulate readers to learn
41
42 more about the many issues touched upon, and that it will motivate other researchers to
43 undertake further study of ice hockey.
44
45 References
46
47 [1] Hurley, WJ. How should you order the players in a shootout? Chance 1998; 8(1):25-
48 26.
49
50 [2] Bertsekas, DP, Tsitsiklis, JN. Introduction to Probability 2002, Athena Scientific,
51 Belmont, MA.
52
53 [3] Mullet, GM. Simeon Poisson and the National Hockey League. The American
54 Statistician 1977 Feb; 31(1):8-12.
55
56
[4] Gill, PS. Late-game reversals in professional basketball, football, and hockey. The
57 American Statistician 2000 May; 54(2):94-99.
58
59
60
John Wiley & Sons
Page 11 of 17 Wiley Encyclopedia of Operations Research and Management Science

1
2
3
[5] Thomas, AC. Inter-arrival times of goals in ice hockey. Journal of Quantitative
4
5 Analysis in Sports 2007; 3(3), http://www.bepress.com/jqas/vol3/iss3/5.
6 [6] Schmuland, B. Shark attacks and the Poisson approximation. Pi in the Sky 2001
7
8
(December); 12-14.
9 [7] Berry, SM, Reese, CS, Larkey, PD. Bridging different eras in sports. Journal of the
10 American Statistical Association 1999; 94(447):661-676.
11
12 [8] Matheson, J. Late bloomer Selanne still defies odds; Ducks veteran has 21 career hat
13 tricks despite starting play in NHL at age 22. Edmonton Journal 2008 (Nov. 4), pg.
14
C.3.
15
16 [9] Hurley, WJ. Jussi Jokinen, Regression to the Mean, and the Assessment of
17 Exceptional Performance. Chance 2009; 22(1):28-33.
18
Fo
19 [10] James, B. The Bill James Abstract. Self-published, 1980.
20
21 [11] Cochran, JJ, Blackstock, R. Pythagoras and the National Hockey League. Journal
22 of Quantitative Analysis in Sports 2009; 5(2),
r
23 http://www.bepress.com/jqas/vol5/iss2/11. DOI: 10.2202/1559-0410.1181.
24
[12] Hurley, WJ. Using operational research to answer some interesting questions about
Pe

25
26 NHL shootouts. International Journal of Operational Research 2009; 4(1):117-123.
27 DOI 10.1504/IJOR.2009.021621.
28
er

29 [13] Brimberg, J, Hurley, WJ. Is the overtime period in an NHL game long enough? An
30 example for teaching estimation and hypothesis testing in the presence of censored
31 data. The American Statistician 2008 May; 62(2):151-154.
32
Re

33 [14] Brimberg, J, Hurley, WJ. Are National Hockey League Referees Markov? Working
34 paper.
35
36 [15] Tversky, A, Gilovich, T. The cold facts about the “hot hand” in basketball. Chance
vi

37 1989; 2(1):16-21.
38
39 [16] Morrison, DG, Schmittlein, DC. It takes a hot goalie to raise the Stanley Cup.
ew

40 Chance 1998; 11(1):3-7.

41
42
[17] Rowe, WG, Cannella, Rankin, D, Gorman, D. Leader succession and organizational
43 performance: integrating the common-sense, ritual scapegoating, and vicious-circle
44 succession theories. The Leadership Quarterly 2005; 16:197-219.
45
46 [18] Audas, R, Goddard, J, Rowe, WQ. Modelling employment durations of NHL head
47 coaches: turnover and post-succession performance. Managerial and Decision
48 Economics 2006; 27:293-306. DOI: 10.1002/mde.1259.
49
50 [19] Abrevaya, J. Fit to be tied: the incentive effects of overtime rules in professional
51 hockey. Journal of Sports Economics 2004; 5(3):292-306. DOI:
52 10.1177/1527002503260560.
53
54 [20] Barnejee, AN, Swinne, JFM, Weersink, A. Skating on thin ice: rule changes and
55 team strategies in the NHL. Canadian Journal of Economics 2007; 40(2):493-514.
56
57
58
59
60
John Wiley & Sons
 Wiley Encyclopedia of Operations Research and Management Science Page 12 of 17

1
2
3
[21] Longley, N, Sankaran, S. The incentive effects of overtime rules in professional
4
5 hockey: a comment and extension. Journal of Sports Economics 2007; 8(5):546-554.
6 DOI: 10.1177/1527002506294938.
7
8
[22] Shmanske, S, Lowenthal, F. Overtime incentives in the National Hockey League
9 (NHL): more evidence. Journal of Sports Economics 2007; 8(4):435-442. DOI:
10 10.1177/1527002506292581.
11
12 [23] Leadley, JC, Zygmont, ZX. National Hockey League. Pp 49-98 in Fizel, J, ed.
13 Handbook of Sports Economics Research. 2006. M. E. Sharpe, Armonk, NY.
14
[24] Williams, B, Williams, D. Are salaries in the National Hockey League related to
15
16 nationality? Chance 1997; 10(3):20-24.
17 [25] Jones, JCH, Walsh, WD. The World Hockey Association and player exploitation in
18
the National Hockey League. Quarterly Review of Economics and Business 1987;
Fo
19
20 27(2):87-101.
21 [26] Laskaris, S. Behind the bench: pulling your goalie. Hockey Player Magazine 1994;
22
http://www.hockeyplayer.com/artman/publish/article_462.shtml
r
23
24 [27] Morrison, DG. On the optimal time to pull the goalie: a Poisson model applied to a
Pe

25
common strategy used in ice hockey. Pp. 137-144 in Machol, RE, Ladany, SP,
26
27 Morrison, DG, eds. TIMS Studies in Management Science, Vol. 4: Management
28 Science in Sports. 1976. North-Holland, Amsterdam.
er

29
30
[28] Morrison, DG, Wheat, RD. Misapplications reviews: pulling the goalie revisited.
31 Interfaces 1986 Nov-Dec; 16(6):28-34.
32 [29] Erkut, E. More on Morrison and Wheat’s “Pulling the goalie revisited.” Interfaces
Re

33
34 1987 Sep-Oct; 17(5):121-123.
35 [30] Nydick, Jr., RL, Weiss, HJ. More on Erkut’s “More on Morrison and Wheat’s
36
‘Pulling the goalie revisited.’” Interfaces 1989 Sep-Oct; 19(5):45-48.
vi

37
38 [31] Washburn, A. Still more on pulling the goalie. Interfaces 1991 Mar-Apr; 21(2):59-
39 64.
ew

40
41 [32] Adler, I, Erera, AL, Hochbaum, DS, Olinick, EV. Baseball, optimization, and the
42 world wide web. Interfaces 2002; 32(2):12-22.
43
44 [33] Russell, T, van Beek, P. 2009. Determining the number of games needed to
45 guarantee an NHL playoff spot. Proceedings of the Sixth International Conference on
46 Integration of AI and OR Techniques in Constraint Programming for Combinatorial
47
Optimization Problems (CPAIOR 2009), Pittsburgh, 233-247.
48
49 [34] Ingolfsson, A. Simulating NHL games to motivate student interest in OR/MS.
50 INFORMS Transactions on Education 2004; 5(1):37-46.
51
52 [35] Morris, C. Breaking deadlocks in hockey: prediction and correlation. Pp. 119-140 in
53 Mosteller, F, ed. Statistics by Example: Detecting Patterns. 1973. Addison-Wesley,
54 Reading, MA.
55
56
57
58
59
60
John Wiley & Sons
Page 13 of 17 Wiley Encyclopedia of Operations Research and Management Science

1
2
3
4 Table 1: Calculation of the probability that the Detroit Red Wings win in regular time, if
5 playing the Boston Bruins.
6
7
8
9
# of goals (n) Pr{n goals} # of goals needed to win Pr{Red Wings win}
10 1 0.015 1 0.525
11
2 0.045 2 0.276
12
13 3 0.089 2 0.538
14 4 0.134 3 0.351
15 5 0.161 3 0.547
16 … … … …
17 20 3.7 × 10-6 11 0.503
18
Fo
19
20 Pr{Red Wings win} = 0.48 Error bound: 1.2 × 10-5
21
22
r
23
24
Pe

25
26
27
28
er

29
30
31
32
Re

33
34
35
36
vi

37
38
39
ew

40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
John Wiley & Sons
 Wiley Encyclopedia of Operations Research and Management Science Page 14 of 17

1
2
3
4 Table 2: NHL Western Conference simulation results, as of 19 March 2009.
5
6
7 Replication Edm. rank Edm. points Edm. qualifies? Points needed
8
9
to qualify
10 1 6 94 Yes 92
11 2 11 86 No *
12 3 10 86 No 87
13 … … … … …
14
15
5,000 5 95 Yes 89
16 *: Simulated criteria insufficient to determine top 8 teams
17
18
Fo
19
20
21
22
r
23
24
Pe

25
26
27
28
er

29
30
31
32
Re

33
34
35
36
vi

37
38
39
ew

40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
John Wiley & Sons
Page 15 of 17 Wiley Encyclopedia of Operations Research and Management Science

1
2
3
4 Table 3: Cross-tabulation of Edmonton Oilers points and whether Edmonton qualifies for
5 the playoffs.
6
7
8 Pts. in last Edm. does Edm. qualifies Pr{Edm. qualifies}
9 12 games not qualify
10
1 1
11
12 2 4
13 3 6
14 4 22
15 5 37
16 6 86
17
18
7 152
8 278 2 0.01
Fo
19
20 9 315 17 0.05
21 10 404 115 0.22
22 11 327 207 0.39
r
23
12 173 469 0.73
24
13 66 537 0.89
Pe

25
26 14 8 583 0.99
27 15 2 416 1.00
28 16 316 1.00
er

29 17 206 1.00
30
31
18 143 1.00
32 19 55 1.00
Re

33 20 37 1.00
34 21 11 1.00
35 22 5 1.00
36
Totals 1881 3119
vi

37
38
39
ew

40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
John Wiley & Sons
 Wiley Encyclopedia of Operations Research and Management Science Page 16 of 17

1
2
3
4
5 0.30 0.295
6 0.293 0.294
7 0.29 0.289
0.287
8
9 0.28 0.280 0.279
10 Pr{tie}
11 0.27 0.270
12 0.263
0.26 0.258
13
14
0.25
15
16 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0
17 Time remaining when goalie is pulled (min.)
18
Fo
19
20 Figure 1: Probability of a tie as a function of the time the goalie is pulled.
21
22
r
23
24
Pe

25
26
27
28
er

29
30
31
32
Re

33
34
35
36
vi

37
38
39
ew

40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
John Wiley & Sons
Page 17 of 17 Wiley Encyclopedia of Operations Research and Management Science

1
2
3
4
5 1.00
0.90 Edmonton does not
6

Cumulative probability
0.80 qualify (mean = 88.7)
7
8 0.70
9 0.60
10 0.50
11
0.40
12
13 0.30 Edmonton does
14 0.20 qualify (mean = 89.0)
15 0.10
16 0.00
17 82 84 86 88 90 92 94 96
18
Points needed to qualify
Fo
19
20
21 Figure 2: Distribution of number of points needed to qualify, depending on whether
22 Edmonton qualifies.
r
23
24
Pe

25
26
27
28
er

29
30
31
32
Re

33
34
35
36
vi

37
38
39
ew

40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
John Wiley & Sons

View publication stats