Professional Documents
Culture Documents
Research Papers in Management Studies: Chris Hope
Research Papers in Management Studies: Chris Hope
STUDIES
Chris Hope
WP 4/2002
Chris Hope
Judge Institute of Management
April 2002
Abstract
By the end of October 2001, less than three months into the 2001-02 football season,
21 managers from the 92 English Premiership and Nationwide League clubs had lost
their jobs. Many commentators thought that the clubs were too quick to act. So what
strategy should a football club adopt when deciding whether to sack its manager?
This paper introduces a simple model assuming that a club’s objective is to maximize
the number of league points that it scores per season. The club’s strategy consists of
three choices:
•= the length of the honeymoon period during which it will not consider sacking
a new manager,
•= the level of the trapdoor, the average number of points scored per game; if the
manager’s performance falls below this, he will get the sack,
•= the weight that it will give to more recent games compared to earlier ones.
Data from the last six seasons of the Premiership are used to calibrate the model. The
best strategy is to have only a short honeymoon period of 8 games, to set the trapdoor
at 0.74 points per game, and to put 47% of the weight on the last five games. A club
adopting this strategy would obtain on average 56.8 points per season, compared to a
Premiership average of 51.8 points. It would employ an average of 5.7 managers
every ten seasons, against the Premiership average of 4.5 managers. It would have
sacked John Gregory at Aston Villa and Walter Smith at Everton, but not Ruud Gullit
or Gianluca Vialli at Chelsea, nor Joe Royle at Everton, just.
Priorities for further work include developing the model to include the distinction
between home and away games, the quality of the opposition, the importance of
avoiding relegation, and the different aspirations of different clubs, and applying it to
other leagues than the Premiership.
Introduction
The 2001 – 2002 season was an uncomfortable one for the managers of the 92 clubs
in English Premiership and Nationwide League football. By the 26th of October 2001,
18 managers had lost their jobs since the start of the season in August, and Walter
Smith, then Manager of Everton, was lamenting
“I feel managers are being sacked too quickly. I think circumstances placed upon directors in the
financial sense means some are getting panicked into changing their managers too quickly. It means
clubs never get the stability which I believe is so vital to them.”
http://news.bbc.co.uk/sport/hi/english/football/newsid_1621000/ 1621182.stm
Smith’s lament went unheeded, and at the end of October, just five days later, it was
reported that
“Oldham Athletic have sacked manager Andy Ritchie - the 21st manager to leave his club since the
start of August.”
http://news.bbc.co.uk/sport/hi/english/football/teams/o/oldham_athletic/newsid_1629000/1629700.stm
So were the clubs too quick to act? What can we say about the strategy that a football
club should follow in deciding whether to sack its manager?
This paper introduces a simple model of this decision problem. The club is assumed
to want to maximize the number of league points that it scores per season. To do this
it would like to employ the best manager it can. There are five types of manager: poor,
fair, good, excellent and world class. But the club cannot observe the quality of its
manager directly. Instead it looks at results on the pitch.
•= the length of the honeymoon period during which it will not consider sacking a
new manager,
•= the level of the trapdoor, the average number of points scored per game; if the
manager’s record falls below this, he will get the sack,
•= the weight that it will give to more recent games compared to earlier ones. The
club increases this weight by increasing the amount of smoothing it applies
when keeping track of the manager’s results.
The choice is complicated by the likelihood that a change of manager will initially
inspire the team and get a boost in performance, and then require some time to
rebuild, during which the team’s performance will drop. And even the best manager
will eventually see his performance drop somewhat with age, and even more as his
skills and relationship with the club decay.
2
So the essence of the club’s dilemma is this: every time it sacks a manager it may get
a short-lived boost in performance, but it incurs a substantial cost and a subsequent
period of rebuilding, both of which cost points. But if it doesn’t sack a mediocre
manager, it will continue to perform badly.
If the club sets the honeymoon period too short, it risks wasting money by sacking a
lot of managers, some of whom might have turned out to be excellent or world class,
but unlucky in their first few games. If it sets the honeymoon period too long, it will
keep even poor managers for longer than their performance would merit.
If the club sets the trapdoor too high, it will sack a lot of managers, some of whom
could be superb managers going through a sticky spell. Too low, and even mediocre
managers will never get the sack.
If the club relies too much on the most recent results, it will sack a lot of managers, as
even a short bad patch will lead to dismissal. If the club uses too little smoothing, it
will take a long while to sack even those managers whose performance has aged and
decayed.
The next section describes a model built using the terms in italics in the above
description. Then, results from the English Premiership are used to calibrate the
model, and the records of some well-known managers are examined. The model is
used to explore a range of strategies that a club could adopt, and some general
guidelines are found. Then an optimal strategy is calculated, its implications
described, and it is applied to some real managers. Finally the simplifications of the
model are discussed and some proposals for further research are made.
3
The model
In the English Premiership and Nationwide football (soccer) leagues, games are
scored at 3 points for a win, 1 point for a draw, and 0 points for a loss. Each team
plays every other team twice during a season, once at home and once away.
Assume that, for each game, a club has a probability of winning of p(win), of drawing
p(draw), and of losing 1 - p(win) - p(draw). The mean number of points scored per
game, g, will be given by
g = 3p(win) + p(draw)
The league
Averaged over a whole season and all the teams in the league, the number of wins
must equal the number of losses, so, using capital letters to represent the average over
the season for the league as a whole,
P(draw) = 1 – 2P(win)
So
G = 3P(win) + P(draw)
= 3P(win) + 1 – 2P(win)
= 1 + P(win)
At the two extremes, if all games in the league are drawn, P(win) = 0, and G = 1; if all
games are either won or lost P(win) = 0.5 and G = 1.5. P(win) is an input to the
model.
The managers
Table 1 shows the five types of manager that are assumed to be available, their quality
measured by g(normal), the mean points per game they will obtain, and the
probability of obtaining them each time a new manager is hired.
4
Table 1 The five types of manager
Table 1 shows that a good manager will equal the mean performance in the league as
a whole, while the excellent and world class managers will do better than this, with
fair and poor managers their mirror image in both quality and probability. If the
population of managers in the league is drawn randomly from this table, then the
mean number of points per game over the season will be G, as required. The model
inputs are d, P(2d) and P(d). G is derived from P(win).
0
1 39 77 115 153 191 229
games
5
Figure 1 shows the profile for an excellent manager, with P(win) = 0.35 (so G = 1.35
points per game), d = 0.2 points per game, inspire = 4 games, boost = 0.2 points per
game, rebuild = 30 games, drop = 0.45 points per game, age = 152 games (4
Premiership seasons), and decay = 38 games (1 Premiership season).
Game g
From to
If inspire + rebuild > age - 1, there is no period when the manager is performing
‘normally’. However, this is not a problem with any reasonable set of input values.
It is assumed that all types of managers, at all stages of their career, have the same
proportion of draws. Data from the English Premiership, shown later, support this
assumption.
Since
g = 3p(win) + p(draw)
we have
p(win) = (g – P(draw))/3
= (3 – 2P(draw) – g)/3
Clearly this assumption will cease to make sense if p(win) < 0 or p(lose) < 0, so the
model checks to ensure that these conditions are not broken, for any type of manager,
in any period of his career.
For each game, a uniform random number between 0 and 1 is generated, and if the
number is less than p(lose), 0 points are scored, if the number is between p(lose) and
6
p(lose) + p(draw), 1 point is scored, and if the number is above p(lose) + p(draw), 3
points are scored. Formally, where m represents the number of the game,
perf(1) = result(1)
where smooth is between 0 and 1, and is a choice variable in the model. The higher
the value of smooth, the more weight is given to the most recent results. At the
extreme, if smooth = 1, only the result of the most recent game is considered. This
form of smoothing is commonly used in quality control in industry, and has the great
advantage that only the most recent result and the most recent previous value of
performance need to be considered (Eppen, Gould and Schmidt, 1990).
If (m >= honeymoon) and perf(m) < trapdoor, then the manager is sacked.
Once the manager has been sacked, a new manager is drawn at random from the types
in table 1, and m is reset to 1.
The aim of the club is to score as many points per season as possible. A running tally
of the points scored is kept for 380 games (10 Premiership seasons). This is then
adjusted in two ways.
1. Every time a manager is sacked before his contract expires, points are deducted
from the club’s total. The number of points deducted is found by multiplying the
7
amount of time left on the contract by the manager’s salary, and by the number of
points that each £million pound’s investment can buy.
Formally, if the manager is sacked after msack games, then points are deducted if
(msack < contract). The number of points deducted is
where
contract is the length of the contract in games,
salary is the manager’s salary in £000 per game,
and
buy_success is the number of points that can be bought for £1 miilion.
Given the progression of the Premiership towards a pure business proposition, there
was a temptation to cast the whole model in terms of discounted cash flow, but, on
reflection, I opted to stay with the traditional league points as the measure of success.
2. At the end of the 380 games, the club’s points are adjusted upwards by an amount
equal to the points that would be lost through rebuilding and paying off the managers’
contracts by a club that had employed an average number of managers over the
period. This adjustment has no effect on the operation of the model, but is necessary
to allow the number of points scored by the club to be compared fairly with the
premiership average.
(manave – 1) * aveloss.
The total points scored by the club over the 380 games including the two adjustments
is divided by 10, to give an average number of points per season. It is this value of
points per season that the club is trying to maximize.
8
The English Premiership
Since the 1995-96 season, the English Premiership has been made up of 20 clubs (24
clubs before that date), so a season has 38 games. The six seasons 1995-96 to 2000-01
make up the database of complete seasons in the present format. Mables-tables.com
provides results of each game played in the Premiership in these seasons, in a
spreadsheet format. Over the six seasons, the proportion of wins is 0.363, and so the
average number of points per game is 1.363.
At the end of each season, the bottom 3 clubs are relegated and replaced by clubs
from the Nationwide League division 1. The top 3 clubs (4 from the 2001-2002
season) qualify for the European Champions League.
Table 3 shows the mean number of points scored per season and per game by clubs in
the top 3 places, the next 4 places, the middle 6 places, the next 4 places, and the
bottom 3 places.
Finishing Points
position
per season per game
1– 3 74.78 1.968
4– 7 60.96 1.604
8 – 13 50.31 1.324
14 – 17 41.38 1.089
18 – 20 33.56 0.883
Source: mables-tables.com
The proportion of drawn games in the five categories was 0.25, 0.26, 0.29, 0.29, and
0.26 respectively. This near constancy inspired the model assumption that all types of
manager have the same proportion of drawn games. A P(win) value of 0.37 implies
P(draw) = 0.26.
9
Time profile of manager performance
It is not so simple to find reasonable values of inspire, boost, rebuild, drop, age, and
decay to describe the time profile of a manager’s performance.
Most commentators seem confident that there is an initial period during which the
team’s performance is boosted:
“If every incoming manager has a honeymoon [inspired in our terminology] period where they inspire
good form and fortune, then Yugoslavia will be hoping that their recently appointed management trio
can have triple the effect of a normal new arrival.” Beale, 2001
“Affable and enthusiastic as he is, does anyone really believe Keegan is the type to rebuild a club like
City?” Kelly, 2001.
But the statistical evidence, particularly on the initial inspired period, is not so strong.
It could be that commentators are comparing the performance of the new manager
with the last few games of the outgoing manager, which are likely to have been rather
poor to earn him the sack. However what is needed for the model is a comparison of
the new manager’s first few games with his subsequent normal performance.
Figures 2 below shows the performance of five high profile managers from their first
game until their sacking (or until the end of the 2000 – 01 season in the case of
Wenger).
10
Figure 2 The performance of five Premiership managers
0
1 11 21 31 41 51 61 71 81 91 101 111 121 131 141 151 161 171 181
Games
0
1 11 21 31 41 51 61 71 81 91 101 111 121 131 141
Games
11
Figure 2c Ruud Gullit at Chelsea
Points per game Recent games
3
Trapdoor
Trend 1
Trend 2
2
0
1 11 21 31 41 51 61
Games
0
1 11 21 31 41 51 61 71 81 91
Games
0
1 11 21 31 41 51 61 71 81 91 101 111 121 131 141
Games
12
The five charts show the exponentially weighted average points per game using a
smoothing value of 0.121 (the jagged line labelled ‘recent games’), and two
polynomial approximations to the recent games line. Trend 1 is a 3rd order polynomial
approximation (the lowest order that can detect an initial boost, a drop, a normal
period, then another drop), and trend 2 a 4th order polynomial. Ignore the trapdoor line
for now; it shows an output from the optimal strategy found in a later section.
Visual inspection of the five charts shows that only Wenger and Gregory show any
sign of the initial boost in performance. However all five show some evidence of a
drop in performance during a rebuilding period, and at least a suspicion of a late
decline, which of course does not have to come from physical aging, but could just be
due to personality clashes or the desire for a fresh start.
From visual inspection of these graphs, plus commonsense, the input values shown in
table 4 are used in the model (see figure 1 and the accompanying explanation for a
definition of the variables).
As these values do not come from a statistical analysis of the underlying data, they are
obviously open to criticism, which is one reason for presenting the raw data of the
charts in figure 2. Anyone wishing to suggest alternative values can use these charts
as a starting place for their own investigations.
The mean values of inspire, boost, rebuild and drop imply that the sacking of a
manager leads on average to the loss of 10.2 points (excluding any loss due to paying
up the manager’s contract) while the new manager rebuilds the team.
The normal length for a contract in the Premiership seems to be about three years.
This is the reported length of Alex Ferguson’s contract extension at Manchester
United (McClaren, 2002).
13
“McClaren would prove very difficult to dislodge from Middlesbrough as he signed a lucrative contract
five-year contract last summer”
http://www.myboro.co.uk/news.asp?page=1&sort=2&storyid=650
As the details of contracts are not published, it is necessary to rely upon intelligent
speculation to deduce salaries. At the top of the Premiership, we have
“the salary on offer more than doubles Arsene Wenger's existing salary, and a figure of three million
pounds a year plus bonuses is being talked about” Parry, 2000.
and
“Ferguson, who is getting a reported salary of two million pounds this season, would expect to receive
a sizable pay increase that would vault him past Arsenal manager Arsene Wenger's reported 2.5 million
salary.”
www.canoe.ca/Slam020205/soc_fer-ap.html
“despite the insurance of a £1.2 million salary, McClaren is finding the Premiership a cruel
environment” Fitton, 2001.
and
When it comes to buying success, there is one piece of solid statistical evidence. The
Deloitte & Touche Annual Review Of Football Finance for 1997/98 reported that
league position was correlated with spending so that a rise of 16 league positions was
related to an extra spend of £22M (Deloitte & Touche, 1998).
Putting these together, we have the values shown in table 5 for a manager’s contract
14
Table 5 Input value distributions for manager’s contract and buying success
The mean contract of 114 games equates to 3 Premiership seasons, while the mean
salary of 53 000 per game equates to £2 million per year. The value for buy_success
is very uncertain, with a mean of 1.5 points per £million, slightly below the figure
implied in the Deloitte and Touche report.
The final piece of information needed is the average number of managers employed
by a club over the ten seasons that the model looks at. The numbers of managers
employed in the ten seasons to May 2001, excluding caretaker managers, and
counting joint managers as one, for the 20 clubs making up the Premiership in 2000-
01 are shown in table 6.
Man Utd 1
Arsenal 4
Liverpool 3
Leeds 3
Ipswich 2
Chelsea 7
Sunderland 5
Aston Villa 4
Charlton 2
Southampton 8
Newcastle 5
Tottenham 8
Leicester 5
Middlesbrough 4
West Ham 2
Everton 6
Derby 3
Man City 5
Coventry 6
Bradford 7
Mean 4.5
Source: www.soccerbase.com
15
The average value in table 6, 4.5, is used as the value of manave in the model. With
this value of manave, the average tenure of a manager is 84.4 games, and the mean
number of points lost from paying up a sacked manager’s contract is
Adding this to the 10.2 mean points lost through disruption, and multiplying by the
average number of sackings, 3.5, shows that an average club will lose 44 points, or
4.4 per season through the turnover of managers.
16
Exploratory analysis
The three choice variables in the model are honeymoon, trapdoor and smooth.
Many commentators imply that new managers have a honeymoon period during
which they will not be sacked:
“Alex Miller was the ideal appointment to stabilise a club which lacked leadership and direction. Now,
almost twelve months on, the honeymoon is over and Miller too will soon be under pressure unless the
situation alters radically.” Gordon, 1998.
The shortest tenure (excluding caretakers) in the Premiership in the last season is 17
games (98 days) for Colin Todd at Derby in October 2001 to January 2002, but only
15 of the games were Premiership games. In 2000-01, Bradford were in the
Premiership and sacked their manager Chris Hutchings in November 2000 after just
12 league games in charge (Austin, 2002).
In 1974, in the precursor to the Premiership, Brian Clough was in charge of Leeds
United for only 44 days and 9 division 1 (old) games
(http://www.soccerbase.com/footballlive).
So one aspect of the investigation with the model concerns the appropriate length of
any honeymoon period.
The trapdoor is the level of performance below which a manager will get the sack.
Over the last six seasons, if a club averaged fewer than 39.67 points in a season, or
1.04 points per game, it could expect to be relegated. So a first guess at an appropriate
level for trapdoor might be just over 1.
But the fluctuating nature of results means that even an excellent or world class
manager might average below 1 point a game for short periods, particularly during
their rebuilding period. So the interaction of trapdoor with smooth and honeymoon
needs to be considered.
Any value of smooth above about 0.3 implies that the most recent games will have far
more weight than earlier ones. Table 7 illustrates this.
percent
smooth weight on last 5 games
0.1 41
0.2 67
0.3 83
0.4 92
0.5 97
0.6 99
17
With a value of smooth above about 0.3, any manager with a run of as few as five
games without a win would see their performance drop close to or below 1, and so be
sacked if the value of trapdoor was as high as this.
To quantify this effect, it is necessary to run the model. Clearly it is not enough to just
run the model once for ten seasons. Many of the model inputs are uncertain, and even
if they were not, the results of individual games certainly are. So the model is run
5000 times for ten seasons with each combination of choice variables, sampling
different values from the uncertain inputs, recording the sackings and results of each
game, and calculating the mean number of points per season for the club. The risk
analysis software @RISK from Palisade Corporation is used to perform the
calculations.
Table 8 shows 27 results of this exploratory analysis, for three values of each of the
three choice variables. The three values for each choice variable are chosen to cover
the range of plausible values. Each entry in the table is the mean points per season for
one combination of the choice variables. So the first entry shows that a combination
of honeymoon = 1, trapdoor = 0.1 and smooth = 0.1 gives a mean result of 46.5 points
per season.
honeymoon = 1
smooth:
0.1 0.3 0.5
trapdoor:
0.1 46.5 51.8 55.8
0.5 52.9 53.3 27.6
1.0 52.5 13.4 -13.0
honeymoon = 8
smooth:
0.1 0.3 0.5
trapdoor:
0.1 45.4 51.0 55.7
0.5 52.7 53.6 36.7
1.0 55.4 33.8 27.9
honeymoon = 40
smooth:
0.1 0.3 0.5
trapdoor:
0.1 45.3 49.6 53.9
0.5 51.5 53.5 47.7
1.0 53.9 46.3 43.7
18
The values in table 8 are only shown to 1 decimal place, as the mean number of points
per season over the ten seasons has a standard deviation of about 7. So even with
5000 runs, the standard error of the mean is about 7/sqrt(5000) = 0.1 points per
season, giving a 95% confidence interval for each of the results in table 8 of about
plus or minus 0.2 points per season.
With this in mind, there are still several insights that can be gained from table 8.
The first is that a combination of a short honeymoon, a high trapdoor and a lot of
weight on recent games is not a good idea. At the extreme of a honeymoon of 1, a
trapdoor of 1.0 and smooth of 0.5, it even leads to an obviously unreal negative mean
number of points per season, as an average of 90 managers are tried, fall below the
trapdoor of 1.0 points per game, and are sacked and paid off during the ten seasons.
But even the less extreme combinations in this bottom right hand area of the three
blocks produce poor results.
The second insight is that the combination of a low trapdoor of 0.1 and a low smooth
of 0.1 produces only about 46 points per season, a poor result, whatever the
honeymoon period. The problem here is the opposite one, with the first manager
employed nearly always being retained for practically the whole of the ten seasons,
whatever his quality.
The third insight is that a long honeymoon period of 40 games generally does not
produce the best results. It leads to poor managers being retained for longer than their
performance would suggest.
Finally, the two best results in the table, with mean points per season of 55.8 and 55.7,
both have a low trapdoor of 0.1 and a high smooth of 0.5, and honeymoons of 1 and 8
games respectively. However, there is another result with a honeymoon of 8 games, a
high trapdoor of 1.0 and a low smooth of 0.1 that gives nearly the same number of
points per season, 55.4.
Looking in more detail at the two best solutions with honeymoons of 8 games shows
that they are of a different character, despite giving a similar result. The solution with
a low trapdoor and high smoothing employs on average fewer than 6 managers in the
ten seasons, and their average quality is almost mid-way between good and excellent
(strictly speaking, the type of manager is an ordinal measure, and so does not have a
mean, but by specifying the symmetric distribution of manager types shown in table
1, it becomes an interval measure whose mean can be taken). The solution with a high
trapdoor and low smoothing employs more managers, just over 9 in the ten seasons on
average, but their average quality is higher, 4 times closer to excellent than good. The
only reason it doesn’t end up with the higher result is that the consequence of sacking
so many managers is a big loss of points through rebuilding and paying off contracts.
This is about as far as the exploratory analysis can take us. There is nothing to
guarantee, or even suggest, that one of the 27 combinations of choice variables in
table 8 is the true optimal combination, the best that is available. Indeed, having two
such different strategies coming so close to giving the best result of the 27 suggests
that finding the optimal strategy will be a challenge.
19
The optimal strategy
Running RISKOptimiser with 5000 iterations per simulation, and all other settings at
their standard values allows about 700 simulations to be performed overnight on a 1.5
GHz Pentium 4 desktop computer. Three overnight runs were performed and gave the
three optimal strategies shown in table 9. All of them give a mean result of about 57
points per season.
Since each of these results has a 95% confidence interval of about 0.2 points per
season, we can say that they are all significantly better than the best of the 27
strategies used in the exploratory analysis in table 8, which gave a mean result of
under 56 points per season.
However, they are not distinguishable from each other with any degree of confidence
(two mean results from 5000 iterations need to differ by 0.28 points per season to be
95% sure that they are actually different). To attempt to distinguish them, an @RISK
run of 38000 iterations was performed with each of the three candidate optimal
strategies from table 9 (38000 iterations gives a 95% confidence interval of 0.07
points per season for each result, or 0.1 for the difference between results). The three
candidates from table 9 scored 56.66, 56.78 and 56.81 points per season respectively.
We can say with 95% confidence that the candidate strategy from run 1 is worse than
the two very similar candidate strategies from runs 2 and 3.
So the optimal strategy for a club would seem to be to allow a manager a honeymoon
period of 8 games, and then sack him if his weighted average performance with a
smoothing value of 0.121 (putting 47% of the weight on the last five games) falls
below 0.74 points per game.
If a club follows this strategy it can expect to obtain 56.81 points per season. It can
expect to employ 5.7 managers on average over the ten seasons, and the average
quality of those managers will be 3/5ths of the way between good and excellent. To
put this into perspective, 56.81 points would have secured a finishing position of 10th,
8th, 8th, 6th, 8th and 8th in the six seasons so far of a 20-club Premiership.
20
Applying the optimal strategy to real managers
Looking again at figure 2 for the five managers, the red line shows the trapdoor and
honeymoon implied by the optimal strategy. The managers’ performance in the charts
is already weighted using the optimal value of smooth (that is why the rather unusual
smoothing value of 0.121 was used to draw the charts), so the charts can be used to
see directly whether a club using the optimal strategy would have sacked any of the
five managers. If at any point, the blue, weighted performance, line falls below the
red, trapdoor, line, the club would have sacked the manager at that point.
Arsene Wenger’s weighted performance has never dropped remotely close to the
trapdoor. This is reassuring, but hardly a stern test of the model, as there would be
severe doubts about the utility of any model that recommended the sacking of
Wenger!
Ruud Gullit’s weighted performance at Chelsea never dropped close to the trapdoor.
In fact it only once dropped as low as 1 point per game, after his 34th game in charge,
the 3-1 defeat away at Newcastle on 16th April 1997. At the time of his sacking in
February 1998, his weighted performance was at a very creditable 1.6 points per
game.
His successor, Gianluca Vialli, flirted with the sack immediately by losing five of his
first seven Premiership games. But he was saved by the honeymoon period of 8
games, and by his eighth game had pulled his weighted performance up to 0.86 points
per game, just keeping clear of the trapdoor. From that point on, his weighted
performance was comfortably clear of the trapdoor, remaining above 1 point per
game. At the time of his sacking in September 2000, it was standing at 1.4 points per
game, and his sacking was greeted with a lot of criticism.
“The email response has been massive, and the majority believe Blues chairman Ken Bates has made a
mistake.” BBC Sport Online, 2000.
Walter Smith’s weighted performance at Everton dropped to 0.75 points per game
after his 24th Premiership game, the 2-1 away defeat at Derby County on 7th February
1999, and to 0.69, below the trapdoor of 0.74, after his 32nd game, the 1-2 home
defeat by Sheffield Wednesday on 5th April 1999. A club using the optimal strategy in
the model would have sacked him at that point. Although his performance dropped
below 1 point per game several times after that, the next time it fell below the
21
trapdoor was after his 143rd Premiership game, the last one before he was actually
sacked (although in reality there was also a cup defeat after this last league game).
One final comparison with a real manager is instructive. Figure 3 shows the end of the
Everton career of Joe Royle, one of Walter Smith’s predecessors. This was not shown
in figure 2 as the early part of his Everton career was during the 24-club Premiership.
His weighted performance dropped to 0.76 points per game after the 4-1 away defeat
at Newcastle on 29th January 1997. This was just above the trapdoor of 0.74 points per
game. So the model would just not have sacked him, although in reality he was
dismissed in March 1997.
Recent games
Trapdoor
1
0
1 11 21 31 41
Games
Source: mables-tables.com
22
Discussion and suggestions for further research
The results presented in the previous section sounded reassuringly precise, and the
comparison with the record of six real managers gives further confidence that the
model is reasonable, but figure 4 shows how much the results can vary even with this
optimal strategy.
0.04
0.03
0.02
0.01
0
0 10 20 30 40 50 60 70 80
mean points per season
Although the mean number of points per season over ten seasons is 56.81, for 5% of
the ten season periods the mean number of points per season falls below 44.42, and
for 5% of them it rises above 66.68.
The former would see the club on the fringes of relegation, as the top club relegated in
the six seasons of a 20-club Premiership so far has scored 38, 40, 40, 36, 33 and 34
points. Football is an uncertain game, and following the best strategy gives no
guarantee of success, even over ten seasons. Perhaps this is a good thing for the long-
term health of the sport.
The latter would see the club regularly challenging for a European place, as 66.68
points would have secured a finishing position of 4th, 5th, 3rd, 5th, 5th and 5th in the six
seasons so far of a 20-club Premiership.
Figure 5 shows the uncertain inputs to the model and their correlation with the
number of points per season. All of the correlations are in the expected direction (the
longer the time to age, the more points per season, the more points that can be bought
for £1 million, the lower the points per season etc), but none of the correlations are
particularly strong. The main lesson from figure 5 is that it would probably be
worthwhile making some effort to find a better estimate of the time it takes for a
manager’s performance to begin to decline (age); the value used in the model is little
more than an educated guess.
23
Figure 5 Correlation of inputs with mean points per
season using the optimal strategy
age 0.204
drop -0.108
salary -0.071
contract -0.058
inspire 0.057
rebuild -0.055
decay 0.027
boost 0.024
Correlation Coefficient
The previous section used 38000 @RISK iterations to show that the very similar
optimal strategies from runs 2 and 3 were better than the candidate from run 1.
However, in each overnight run, 25 or so of the top strategies gave results within
about 0.28 points per season of the optimal strategy. It needs a difference of 0.28
points per season to be able to dismiss a strategy as sub-optimal. Table 10 shows these
strategies for run 3. Checking each of these with 38000 @RISK runs would be too
time consuming, so it remains an open question whether any of these are actually the
optimal strategy.
•= The honeymoon periods are short, ranging from a minimum of three games to
a maximum of eight.
•= The trapdoors lie in a fairly small range, between 0.69 and 0.92 points per
game.
24
•= The smoothing coefficients are fairly gentle, between 0.09 and 0.14. If the
trapdoor values are towards the top of the range, the smoothing coefficients
are towards the bottom, and vice versa.
What table 10 also illustrates is that there are many strategies that give almost the
same results as the optimal strategy. All the strategies in table 10 appear to give a
mean result within 0.3 points per season, or 3 points over 10 seasons, of the optimal
strategy. 3 points is the difference between losing and winning one game in ten
seasons – perhaps a single penalty miss by your captain and a flash of brilliance from
an opposition striker.
25
What’s missing from the model
The model described here is clearly only a first attempt at applying management
science techniques to the hiring and firing of football managers. The more glaring
simplifications include:
•= No consideration of whether games are played at home or away. In the last six
seasons of the Premiership, clubs averaged just under 9 home wins, but only
just under 5 away wins per season. In reality a club might be less willing to
forgive a poor run of results at home, than a poor run away from home.
26
and transfer dealings, and probably take this into account when deciding
whether to sack him or not.
Despite all these shortcomings, and no doubt others that are obvious to the reader if
not the author, the model does seem to give reasonable insights, and the calculations
required to apply the optimal strategy (as opposed to those required to discover it) are
simple. Any club considering the future of its manager might find it worthwhile at
least to take a few minutes to check how close his performance is to the trapdoor
before pulling the lever that consigns him to the sack.
Further research
The first priority is to update the data in the model once the complete information for
the 2001-02 season becomes available. This will allow the optimal strategy to be
recalculated using the most recent data. It will also enable the performance of all the
current Premiership managers to be monitored and compared to the trapdoor.
The second task is to adapt the model so that it can be applied to the Nationwide
League clubs in England, where the seasons consist of 46 games, and the financial
circumstances are very different to those in the Premiership. Looking at other soccer
leagues, in Spain and Italy particularly, or other sports, would also seem to be
possible.
It would also be worth trying to improve the accuracy of the inputs to simulate the
actual Premiership more closely. In particular, the restriction to five types of manager,
symmetrically arranged around the ‘good’ manager who has an average performance,
is not a perfect representation of the Premiership. Inspection of table 3 shows that the
‘good’ managers actually do slightly worse than this on average, while both the ‘poor’
and ‘world class’ managers do rather better than the model assumes. Figure 5 shows
that more systematic investigation to estimate the time profile of a manager’s
performance, particularly any drop in performance after several seasons in the job,
would also be worthwhile.
27
The foremost football research centres in the UK at Leicester University
(www.le.ac.uk/snccfr ), Birkbeck College (www.football-research.org/ ) and
Liverpool University (www.liv.ac.uk/footballindustry/) concentrate on the social and
governance aspects of the game. Deloitte & Touche Sport produces excellent annual
publications on football finance. But quantitative modelling so far seems to have been
absent. For a £1 billion per year industry (Deloitte & Touche, 2000), this seems an
odd omission, and surely an opportunity to be grasped.
28
References
Austin S, 2002, Quick-fire managers,
news.bbc.co.uk/sport/hi/english/football/newsid_1761000/ 1761595.stm
BBC Sport online, 2000, Fans stand behind Vialli,
http://news.bbc.co.uk/sport/hi/english/football/teams/c/chelsea/newsid_924000/92461
7.stm
Deloitte & Touche, 1998, Annual Review Of Football Finance 1997/98, p41.
Eppen G , Gould F and Schmidt C, 1990, Quantitative Concepts for Management, 3rd
Ed, Prentice Hall.
Kelly D, 2001,
http://www.megasoccer.com/page/genpage.php?nt=world&nv=World&ns=home&pa
ge=columns%2Fkelly10.html
McClaren S, 2002, Fergie quit? He's the last of a dying breed, Observer, 3 March
2002, http://www.observer.co.uk/sport/story/0,6903,660885,00.html
29
Appendix: Cautionary lessons from an alternative formulation
An earlier version of the model assumed that all managers would keep the same
proportion of wins to draws as in the league as a whole. So
p(win)/p(draw) = P(win)/P(draw)
= P(win)/(1-2P(win))
g = 3p(win) + p(draw)
= 3p(win) + p(win)(1-2P(win))/P(win)
= p(win)*(3 + (1-2P(win))/P(win))
= p(win)*(1+P(win))/P(win)
g > (1 + P(win))/(1 – P(win)), so the model checked to ensure that this condition was
not broken.
However, this formulation gave too few draws for the worst managers, and too many
for the best, and so was abandoned in favour of the formulation described in the text.
It also produced a quirky result from the optimisation procedure, as during the
inspired period at the start of a manager’s tenure, the probability of losing a game
would be very small, often below 2%, for the world class managers. So the model
would find it optimal to sack the manager if he lost his first game, since it was almost
impossible to sack a world class manager this way. This was clearly unrealistic
behaviour that was an artefact of the model.
30
JUDGE INSTITUTE OF MANAGEMENT
WP Altered States: the Impact of the New Economy on Sustainable Development in Southeast
1/2001 Asia (B Ahlering and C Smallman)
WP A Customer-Driven Model of Job Design: Towards a General Theory
2/2001 (R F Conti and M Warner)
WP The Timing and Terms of Mergers, Stock Offers and Cash Offers
3/2001 (B Lambert)
WP Against Learning
4/2001 (C Grey)
WP Extremes in Operational Risk
5/2001 (E A Medova and M N Kyriacou)
WP Mergers and Acquisitions: The Influence of Methods of Payment on Bidder’s Share Price
6/2001 (R Chatterjee and A Kuenzi)
NUMBER TITLE AND AUTHOR
For details of further past papers in our series, please contact us, and we will be happy to send you our
booklet.
ORDER FORM
Please send this form, together with draft, money order or cheque (made payable to
THE JUDGE INSTITUTE OF MANAGEMENT) to:
Publications Secretary
Judge Institute of Management
University of Cambridge
Trumpington Street
CAMBRIDGE CB2 1AG
UK
Tel: + 44 1223 339636
Fax: + 44 1223 339701
E-mail: [email protected]
--------------------------------------------------------------------------------------------------------------------
Name:
Address:
List paper number(s) required below, together with the author’s name(s):
WP Number Author(s)
Price per publication - £5.50 (th is includes postage and packing for one working paper.
Please note that for substantial orders (5 working papers or more) and overseas orders
p & p will be billed separately.
Total Payment enclosed:
£