1
00:00:13,790 --> 00:00:17,870
Welcome back to the Oxford Mathematics Public Lecture Home Edition.

2
00:00:17,870 --> 00:00:26,150
This is the fourth event. My name is and I go, really? And I'm in charge of external relations for the Mathematical Institute as usual.

3
00:00:26,150 --> 00:00:30,140
Special thanks to a sponsor execs Market Exchange Market,

4
00:00:30,140 --> 00:00:37,010
a leading quantitative driven electronic market maker with offices in London, Singapore and New York.

5
00:00:37,010 --> 00:00:42,230
Their ongoing support is crucial in providing you quality content.

6
00:00:42,230 --> 00:00:50,030
It has now been more than five months since the lockdown started. The news is still dominated by the current coronavirus crisis.

7
00:00:50,030 --> 00:00:58,400
Scientist and mathematician are busier than ever working on a possible solution to the many problem that it has created.

8
00:00:58,400 --> 00:01:02,400
But if you are like me, you must be a bit exhausted by all of it.

9
00:01:02,400 --> 00:01:10,490
So at Oxford Mathematics, we decided that it was time for a bit of fantasy fantasy football, to be precise.

10
00:01:10,490 --> 00:01:14,870
A couple of weeks ago, it was announced that one of the post-docs, Joshua Boyle,

11
00:01:14,870 --> 00:01:24,710
was the winner of the 2020 Fantasy Premier League competition, an event in which more than eight million people around the world compete.

12
00:01:24,710 --> 00:01:33,680
In his day job, Josh studied mathematical biology in Oxford, and he has kindly agreed to give us a public lecture about fantasy football.

13
00:01:33,680 --> 00:01:37,430
What is it? An arcane mathematics help win today?

14
00:01:37,430 --> 00:01:41,060
Let's find out if you have questions for Josh.

15
00:01:41,060 --> 00:01:46,250
Send them to us through social media. Josh will collate them and post the answers.

16
00:01:46,250 --> 00:01:52,770
So thank you very much, Josh, for doing this. Please stop now.

17
00:01:52,770 --> 00:01:58,830
Thanks, Helen. So my name is Joshua Bell, and I'm a researcher, not in fantasy football,

18
00:01:58,830 --> 00:02:04,440
but in mathematical biology, which is something that I'm going to sneak into this talk as well.

19
00:02:04,440 --> 00:02:11,400
And so I work in the Wolfson Centre for Mathematical Biology at the Oxford University Mathematical Institute.

20
00:02:11,400 --> 00:02:17,010
And as I say, my my main area of focus is on cancer modelling.

21
00:02:17,010 --> 00:02:21,000
And but today, I'm not really going to talk to you about that cancer.

22
00:02:21,000 --> 00:02:31,830
Talk to you about whether maths can tell us how to win fantasy football. And the answer to that and one slide is not really, but also sort of yes.

23
00:02:31,830 --> 00:02:36,370
And so thank you all for watching. It's been a lot of fun.

24
00:02:36,370 --> 00:02:45,570
And don't worry, obviously you get more for your money than that, and I will start by just kind of introducing myself a bit.

25
00:02:45,570 --> 00:02:53,490
So as I say, I'm not a football expert, I'm a mathematician and specifically a researcher in mathematical biology.

26
00:02:53,490 --> 00:03:00,150
And as I say, I'm developing mathematical tools to apply to the problem of cancer.

27
00:03:00,150 --> 00:03:02,940
So in particular, I use something called mathematical modelling,

28
00:03:02,940 --> 00:03:11,790
which I'm going to talk to you about in a bit more detail and just generally trying to use maths to gain insights and test ideas.

29
00:03:11,790 --> 00:03:17,010
And yeah, that's that's kind of what this process of mathematical modelling is.

30
00:03:17,010 --> 00:03:23,430
So as far as giving a talk about fantasy football goes, my football expertise is a bit more limited.

31
00:03:23,430 --> 00:03:29,730
And so you might question my footballing judgement straightaway. I mean, at Switch Town Football Club supporter.

32
00:03:29,730 --> 00:03:38,310
And for those of you that are more focussed on less football inclined, they are the best football club in the world, as far as I'm aware.

33
00:03:38,310 --> 00:03:42,750
And unfortunately, I can't remember ever scoring anything apart from an own goal.

34
00:03:42,750 --> 00:03:49,680
But I do occasionally come up with the good football pub, although I can assure you there won't be any endless talk, at least not intentionally.

35
00:03:49,680 --> 00:03:56,190
Maybe if I think of one off the top of my head. And so why am I giving this [INAUDIBLE] about fantasy football?

36
00:03:56,190 --> 00:04:05,550
Well, because the Premier League made a GIF about me. So there is a competition which they've won every year called Fancy Premier League,

37
00:04:05,550 --> 00:04:13,050
and which this year about seven and a half million people competed and I was lucky enough to win it.

38
00:04:13,050 --> 00:04:21,690
So I'm going to talk to you about what that means a bit more and and talk to you about the question which everyone asked,

39
00:04:21,690 --> 00:04:27,450
which is, Well, does that have anything to do with you being a mathematician? But first things first.

40
00:04:27,450 --> 00:04:35,190
What is fantasy football? So I try to explain this to some mathematicians recently, and it took about 20 minutes.

41
00:04:35,190 --> 00:04:38,430
It wasn't a good explanation, and at the end, someone said, Hey,

42
00:04:38,430 --> 00:04:44,340
you could have just played that one minute video the fantasy Premier League on Twitter, and that would have said it.

43
00:04:44,340 --> 00:04:46,500
So I thought I'd do exactly that.

44
00:04:46,500 --> 00:04:55,710
So the idea of fantasy football is that you'd start off with a budget of £100million and you choose a team of 15 players using that.

45
00:04:55,710 --> 00:05:00,540
And from these specific positions, you're only allowed three players from any club.

46
00:05:00,540 --> 00:05:07,980
So you can't just load up with Liverpool or Man City or something. And every week you choose 11 players that you think will do the best.

47
00:05:07,980 --> 00:05:12,570
And basically, when they go and play the actual football games, you're virtual,

48
00:05:12,570 --> 00:05:21,450
players get points for their performance on the pitch and you get one player to be your captain who scored twice as many points.

49
00:05:21,450 --> 00:05:26,740
That's it in a nutshell, and I'll go over some more rules later when they become relevant.

50
00:05:26,740 --> 00:05:35,730
Don't worry. So the question that everybody keeps asking me once it finished was, did you use maths to?

51
00:05:35,730 --> 00:05:40,830
And I think that's a really interesting question to kind of pick apart a little bit first

52
00:05:40,830 --> 00:05:46,620
because it definitely seems to be true that being a strategic thinker is a benefit,

53
00:05:46,620 --> 00:05:52,450
and mathematicians like to think of ourselves as being quite good strategic thinkers, I think.

54
00:05:52,450 --> 00:06:00,390
So just as an example here, this this headline is from The Guardian talking about Magnus Carlsen,

55
00:06:00,390 --> 00:06:10,230
who's the world chess champion and who also plays Fantasy Premier League, and he was topping the rankings back in December.

56
00:06:10,230 --> 00:06:14,820
I think he finished ninth or 10th, maybe 11th in the end.

57
00:06:14,820 --> 00:06:20,580
But anyway, there seems to be some link between strategic thinking and doing well at this game.

58
00:06:20,580 --> 00:06:25,200
But I really want to pick into this question of did you use maths?

59
00:06:25,200 --> 00:06:30,330
Why? Why is that a question that I've been asked by lots of non mathematicians?

60
00:06:30,330 --> 00:06:34,860
Why do we think that maths could help us to understand the game of fantasy football?

61
00:06:34,860 --> 00:06:39,770
Are we expecting that we can write down some equation which tells us exactly what to do?

62
00:06:39,770 --> 00:06:46,680
And you know, that doesn't seem at first glance to have anything to do with choosing a fantasy football team.

63
00:06:46,680 --> 00:06:54,000
So I guess as a starting point to answer that question, we need to think about what it is that mathematicians do.

64
00:06:54,000 --> 00:07:00,390
And believe it or not, we we do do things. So I work in mathematical modelling.

65
00:07:00,390 --> 00:07:07,440
When I tell people that I work in modelling, they normally say, Well, somebody with your good looks and fashion sense that doesn't surprise us,

66
00:07:07,440 --> 00:07:12,030
and I don't have to explain to them that know that that's a different kind of mathematical model.

67
00:07:12,030 --> 00:07:18,600
It's not just a particularly good looking mathematician like Albert. And but it's it's a tool that we use.

68
00:07:18,600 --> 00:07:29,640
So in particular, I want to think about mathematical models as being a way of of exploring and understanding very complex systems,

69
00:07:29,640 --> 00:07:35,790
which is a very general definition that I just made up and have have.

70
00:07:35,790 --> 00:07:45,780
Cruelly quoted to Bobby Robson, I'm sure he was very good at mathematical modelling, but as far as I'm aware, he didn't actually say something.

71
00:07:45,780 --> 00:07:49,980
But what John Kerry did say is that people don't believe that maths is simple.

72
00:07:49,980 --> 00:07:53,490
It's just because they don't realise how complicated life is.

73
00:07:53,490 --> 00:08:01,860
So I think this is a really great quote for understanding mathematical modelling because to lots of people, maths is something that's very complex.

74
00:08:01,860 --> 00:08:09,270
And but really, when you compare the complexity of of solving some equations and the complexity of

75
00:08:09,270 --> 00:08:15,060
learning things in a maths class with trying to describe how the real world works,

76
00:08:15,060 --> 00:08:19,020
you realise that actually it's it's a simplification.

77
00:08:19,020 --> 00:08:29,100
And really what we try to do is is translate complicated things into maths so that we can then solve them in this kind of simple language.

78
00:08:29,100 --> 00:08:35,820
So in terms of a practical example of that, and I promise I will start talking about fantasy football eventually,

79
00:08:35,820 --> 00:08:43,740
but I want to think about the big question, the really big question in my research, which is how should we treat cancer?

80
00:08:43,740 --> 00:08:51,900
So that's a really complicated question. It's got a lot of a lot of problems just in the framing of the question.

81
00:08:51,900 --> 00:08:57,690
Cancer isn't one disease. There's not one particular approach which is going to be best.

82
00:08:57,690 --> 00:09:05,400
But what we aim to do is as scientists, first and foremost, is to try and try and break that down into smaller,

83
00:09:05,400 --> 00:09:11,850
more meaningful questions so we can break that down and break it down and break it down again.

84
00:09:11,850 --> 00:09:18,240
And don't worry if this is an overwhelming amount of information because it's supposed to be.

85
00:09:18,240 --> 00:09:24,750
These questions aren't easy. And the message I want you to take from this is that the aim of science is to break down these big

86
00:09:24,750 --> 00:09:32,070
questions until we get questions that we can actually answer until we get hypotheses that we can test.

87
00:09:32,070 --> 00:09:41,980
And that's science in a nutshell. So why am I telling you all of this in the fantasy football team and see football team fantasy football talk?

88
00:09:41,980 --> 00:09:46,300
Well, because fantasy football is another complex system.

89
00:09:46,300 --> 00:09:53,500
It's something where you know, it's it's very easy to ask the question Well, how do we win fantasy football?

90
00:09:53,500 --> 00:09:58,420
But actually trying to tackle that question is very difficult.

91
00:09:58,420 --> 00:10:02,260
You need to break it down into smaller parts. Who should I pick in my team?

92
00:10:02,260 --> 00:10:06,160
Who should I choose to be my captain? That's caused me double points.

93
00:10:06,160 --> 00:10:10,850
And then we can repeat this process until we get to some things we can actually test.

94
00:10:10,850 --> 00:10:15,550
You know, who should I put my team? I should choose players that have got good fixtures.

95
00:10:15,550 --> 00:10:20,980
That may be true. That may not be true, but it's a hypothesis that we can go out and investigate.

96
00:10:20,980 --> 00:10:26,140
And the way that scientists have traditionally done this is by going and doing experiments.

97
00:10:26,140 --> 00:10:34,390
Once they've got a hypothesis and they look at the experiments and they see whether whether it stands up to scrutiny or not.

98
00:10:34,390 --> 00:10:40,750
That's not the only way that we can explore these hypotheses, and that's what we try to do with mathematical modelling.

99
00:10:40,750 --> 00:10:50,940
The aim of the model is is to help us to explore these problems and to look at these hypotheses and see what more information we can get from them.

100
00:10:50,940 --> 00:10:58,470
So one of the best ways to think about about modelling is about reducing complexity.

101
00:10:58,470 --> 00:11:07,390
We don't want to try and build a model of the entire fantasy football game because it's very complicated.

102
00:11:07,390 --> 00:11:11,910
There's a lot of moving parts. We don't know how many points players will score.

103
00:11:11,910 --> 00:11:16,080
People might be injured or they might move team on.

104
00:11:16,080 --> 00:11:20,820
Any number of things could happen. And then during the game, anything might happen.

105
00:11:20,820 --> 00:11:30,000
So as with when we're applying research in a kind of more potentially more useful context,

106
00:11:30,000 --> 00:11:35,850
what we want to do is break, break things down and find a model to tackle a simple problem.

107
00:11:35,850 --> 00:11:41,340
So one way of thinking this is to try and think about what you do with your hypothesis.

108
00:11:41,340 --> 00:11:48,180
So this is an example that I'm kind of taking very generically from a research paper,

109
00:11:48,180 --> 00:11:53,940
which we recently did and which I've mainly put in here to convince you that I don't spend all of my time doing fantasy football.

110
00:11:53,940 --> 00:11:56,000
I do actually do some real research.

111
00:11:56,000 --> 00:12:05,490
And so you might start with a hypothesis like cells within an you might get pushed around and you might find some experimental data about that.

112
00:12:05,490 --> 00:12:11,940
So what you want your mathematical model to do is to take that data and build on it.

113
00:12:11,940 --> 00:12:16,740
You want to see if you can get an understanding of why you see the data that you see

114
00:12:16,740 --> 00:12:22,440
by taking in predictable rules that describe the system that you're interested in.

115
00:12:22,440 --> 00:12:28,620
So in this case, I want to simulate a tumour and simulate the forces that are at work within that.

116
00:12:28,620 --> 00:12:33,420
And then I can go and explore that and see what happens when I change things in my model that

117
00:12:33,420 --> 00:12:38,880
I couldn't necessarily do in an experiment and that might then generate a new hypothesis.

118
00:12:38,880 --> 00:12:45,030
So instead of saying, well, cells get pushed around, we can maybe say, well, with our model,

119
00:12:45,030 --> 00:12:52,110
we can predict what how the tumour is behaving by looking at the way that cells pushed around.

120
00:12:52,110 --> 00:13:03,300
So that hypothesis can then feed back into to our collaborators who work in the lab, who might then say, Oh,

121
00:13:03,300 --> 00:13:09,840
well, let's let's do the experiments, test that, let's see if we can, if we can push this and we get a cycle.

122
00:13:09,840 --> 00:13:20,770
So. Really, what we want to do or what I want to do is get these these kind of two different ideas through to the.

123
00:13:20,770 --> 00:13:26,740
Scientific thinking is about trying to reduce complex problems down to testable statements,

124
00:13:26,740 --> 00:13:33,280
whereas mathematical models are about taking taking those these simple statements and

125
00:13:33,280 --> 00:13:39,110
exploring them and seeing what the consequences are of the assumptions that you're making.

126
00:13:39,110 --> 00:13:47,420
So now we can talk about fantasy football, because in fantasy football, as we've just seen, we can have some hypotheses.

127
00:13:47,420 --> 00:13:52,460
So people kept saying to me, what you know, how?

128
00:13:52,460 --> 00:13:55,490
What was your strategy? What were you doing?

129
00:13:55,490 --> 00:14:02,600
And I didn't write these rules down at the start of the season and try to play the game, according to these rules.

130
00:14:02,600 --> 00:14:09,290
But in hindsight, I think these are the kinds of things that I was thinking about to make my decisions.

131
00:14:09,290 --> 00:14:16,970
So in particular, things like only having a very few expensive players, but sticking with them for the whole season.

132
00:14:16,970 --> 00:14:22,790
So as a general rule, players that score more points are more expensive in the game.

133
00:14:22,790 --> 00:14:29,830
So you'd like to have a team made just of expensive players, but obviously you can't afford to do that.

134
00:14:29,830 --> 00:14:39,060
I was also trying to give players a chance and hold on to them for, you know, longer than possibly most fantasy football managers.

135
00:14:39,060 --> 00:14:46,380
But when I was making transfers, I was using those transfers to get rid of players that I didn't want in my team,

136
00:14:46,380 --> 00:14:54,340
as opposed to just bringing in people that have scored really well, but one in my team and thinking, I need to get that player.

137
00:14:54,340 --> 00:14:59,920
And also had a very good practise of, if in doubt, cut Mo Salah, and of course,

138
00:14:59,920 --> 00:15:05,590
the most important hypothesis of all was never to pick him up from Norwich, which is such time finance.

139
00:15:05,590 --> 00:15:13,750
I can tell you it's very good advice. So the question really is I've got two sets of hypotheses here.

140
00:15:13,750 --> 00:15:19,540
Do these tell us what the right moves to make up? So what do we mean by the right moves?

141
00:15:19,540 --> 00:15:24,890
So I scored just over 2500 points over the course of the season.

142
00:15:24,890 --> 00:15:34,660
And but this this talking Kiwi on Twitter has managed to look back at the data and find a set of transfers you could make,

143
00:15:34,660 --> 00:15:39,250
which would score four thousand eight hundred eighty nine points, which is a lot.

144
00:15:39,250 --> 00:15:45,720
And no fantasy football manager, to my knowledge, has ever come anywhere close to doing that.

145
00:15:45,720 --> 00:15:51,460
So does that mean that those moves were the right moves to make and the, you know,

146
00:15:51,460 --> 00:15:57,790
the things that we should all have been doing last season were these exact transfers that this person has found.

147
00:15:57,790 --> 00:16:10,150
Well, the problem here is that all of the moves that this Kiwi has found are made in hindsight, so it's easy to look back and say, well,

148
00:16:10,150 --> 00:16:15,940
transferring out this expensive player that we expect to do well for this cheap defender

149
00:16:15,940 --> 00:16:22,360
who never plays but came on for that one week and happened to score three goals.

150
00:16:22,360 --> 00:16:28,150
Making those moves in real life is very risky and and it's very unlikely to pay off.

151
00:16:28,150 --> 00:16:35,680
But of course, in hindsight, we can see that they did pay off. So this is like trying to pick a sort of lottery numbers in hindsight.

152
00:16:35,680 --> 00:16:39,790
We don't really want to say that these moves, but the right moves to make,

153
00:16:39,790 --> 00:16:44,260
we want to think more about general strategies that we can then apply in the future.

154
00:16:44,260 --> 00:16:52,570
So that's the question I want to try and think about is can we use modelling to try and compare more general strategies,

155
00:16:52,570 --> 00:16:59,560
these broad approaches that you might make when making your decisions in Fantasy Premier League?

156
00:16:59,560 --> 00:17:07,660
And can we explore those with mathematics in the same way that we explore making similar decisions in cancer research,

157
00:17:07,660 --> 00:17:13,570
cyber physics or whatever other area of applied maths that might be?

158
00:17:13,570 --> 00:17:21,680
So the first question is, well, what should we model which of these hypotheses that we can think about should we test?

159
00:17:21,680 --> 00:17:27,700
And we have a very sophisticated way of making those kinds of decisions, which Twitter.

160
00:17:27,700 --> 00:17:35,290
So I went on Twitter the other day and I asked the wonderful fancy family community that's there.

161
00:17:35,290 --> 00:17:42,700
Which of these problems they worry about the most. And top of the list by quite a long way was captaincy choice.

162
00:17:42,700 --> 00:17:51,100
So each week I got to pick one player, just one player from my team, and I will get double those points.

163
00:17:51,100 --> 00:17:55,540
So here is my team from some points in the middle of last season,

164
00:17:55,540 --> 00:18:02,140
and most of my players didn't do very well, except for Mo Salah, who didn't score 32 points.

165
00:18:02,140 --> 00:18:06,110
He scored 16 points. And so I don't know what that is.

166
00:18:06,110 --> 00:18:12,570
A couple of goals and maybe an assist. And unfortunately, I've chosen him as my captain.

167
00:18:12,570 --> 00:18:17,020
And so instead of getting 16 points, I got thirty two,

168
00:18:17,020 --> 00:18:26,500
which is great because if I'd chosen different players like Kevin De Bruyne and instead of having those extra 16 points coming into my team,

169
00:18:26,500 --> 00:18:31,030
I would only have had an extra two. So this is a really tricky decision,

170
00:18:31,030 --> 00:18:38,800
which obviously people in the FPL world are worried about because it can really make or break your season can make a big difference.

171
00:18:38,800 --> 00:18:45,190
And it's not an easy thing to choose. There's a lot of different ways that you might think about choosing your captain.

172
00:18:45,190 --> 00:18:54,640
So what we want to do is think about some strategies and then come up with a way of modelling to compare them against each other.

173
00:18:54,640 --> 00:19:03,520
So one one comparison, which often gets thrown about in the world of FPL, is the idea of form versus fixture.

174
00:19:03,520 --> 00:19:06,700
Should I take a player who is in great form?

175
00:19:06,700 --> 00:19:12,340
The last three weeks that scored every week they have in the time of their life, they're playing really well.

176
00:19:12,340 --> 00:19:15,220
Or should I pick a player who's got an easy fixture?

177
00:19:15,220 --> 00:19:20,920
Because although that player, you know, although Mo Salah may have scored in every one of his last three weeks,

178
00:19:20,920 --> 00:19:25,030
now he's playing Man City away and it's a really tough match.

179
00:19:25,030 --> 00:19:29,350
We're not really expecting him to do that brilliantly. But at the same time,

180
00:19:29,350 --> 00:19:36,760
perhaps some other player who hasn't had a good run of luck recently has got a really easy fixture that might be playing Norwich home or something,

181
00:19:36,760 --> 00:19:41,440
and you expect to score lots of goals and that's a real dilemma.

182
00:19:41,440 --> 00:19:48,520
People don't know the answer. People genuinely have big arguments about this, so let's try and formalise it.

183
00:19:48,520 --> 00:19:53,020
Let's try and use some modelling to try and understand it.

184
00:19:53,020 --> 00:19:58,150
Have a way of doing this might just be well, I never captain a player who's playing away.

185
00:19:58,150 --> 00:20:02,530
I only have some people that are at home. It's not a good thing or a bad thing.

186
00:20:02,530 --> 00:20:07,720
And notice that these things aren't exclusive, you can be in good form and be playing at home.

187
00:20:07,720 --> 00:20:11,980
You can be in good form and be playing away. Which one should I think about first?

188
00:20:11,980 --> 00:20:16,390
Which one's most important? These are all the kinds of questions that we can ask.

189
00:20:16,390 --> 00:20:19,090
You might even just say, Well, I don't want to think about it.

190
00:20:19,090 --> 00:20:24,160
I'm just going to pick the same caps and I'm going to set it at the start of the season and forget about it.

191
00:20:24,160 --> 00:20:31,870
And whatever happens happens whoever that play. So we want to build some sort of model to answer these questions.

192
00:20:31,870 --> 00:20:40,510
And the way that I'm going to think about doing that today is by essentially simulating a lot of teams.

193
00:20:40,510 --> 00:20:53,380
So what I want to do is essentially make a load of of fake teams and pretend that they were all starting from the start of last season

194
00:20:53,380 --> 00:21:03,800
and let them play and let them choose their captain based on whichever strategy it might be and see how many points they would have got.

195
00:21:03,800 --> 00:21:12,980
So one way of doing this, and I have to invest that I'm going to talk to you about and it's to simplify the problem,

196
00:21:12,980 --> 00:21:20,750
so we don't want to worry about transfers, we don't want to worry about choosing our entire team of 11 people.

197
00:21:20,750 --> 00:21:26,600
We don't want to worry about how we allocate our budget. We're only interested in who we make the captain.

198
00:21:26,600 --> 00:21:35,450
So most teams have somewhere between, say, two and five players that they would probably be willing to make their captain in any given week.

199
00:21:35,450 --> 00:21:41,480
So really, we're not interested in all of the other players. We just want to look at this small pool of people.

200
00:21:41,480 --> 00:21:47,120
The players are choosing from Fantasy Premier League players as opposed to football players.

201
00:21:47,120 --> 00:21:52,280
These managers are choosing from in order to make them their captain.

202
00:21:52,280 --> 00:22:00,410
So I picked up a list of people here picks essentially every player in the game who costs more than nine million pounds,

203
00:22:00,410 --> 00:22:07,290
plus a couple of other players less than nine million pounds, but a very popular options.

204
00:22:07,290 --> 00:22:12,230
And that's going to be my my pool of possible players so I can have in my team.

205
00:22:12,230 --> 00:22:15,980
And then I'm going to simulate a team I'm going to pick, say,

206
00:22:15,980 --> 00:22:24,890
two to four of those players at random and say that four team number one, those are the two to four players that I've got.

207
00:22:24,890 --> 00:22:28,640
And then I'm going to choose a strategy for that team to stick to it for the season.

208
00:22:28,640 --> 00:22:34,370
Do I make transfers based on whether they're at home or why, based on who's in good form?

209
00:22:34,370 --> 00:22:40,670
And every week I'm going to look at the data that's available up to that point and make that decision.

210
00:22:40,670 --> 00:22:49,820
And then I'm going to do that 10000 times and we'll see over the course of those 10000 iterations which of those strategies as the most

211
00:22:49,820 --> 00:22:57,440
successful reason we can do that is that we're now controlling for all of these complicating factors that we don't have to worry about.

212
00:22:57,440 --> 00:23:04,190
We know that all of these teams were made in the same way and we know exactly what factors they're taking into that decision.

213
00:23:04,190 --> 00:23:09,770
We don't have to worry about whether it's true that in this occasion,

214
00:23:09,770 --> 00:23:18,020
we should have captained a defender who happened to score a hat trick because that's that's not the type of answer that we want to look at.

215
00:23:18,020 --> 00:23:25,010
We want to look at whether the general strategy gives you more points or not. So we've simplified the problem down.

216
00:23:25,010 --> 00:23:29,870
So now let's answer the question from my 10000 teams.

217
00:23:29,870 --> 00:23:38,330
I'm going to pick a random player from my from my captaincy choices that's either at home or is playing away.

218
00:23:38,330 --> 00:23:42,290
And over the course of of those 10000 teams,

219
00:23:42,290 --> 00:23:48,170
I can look at how many points that would get me during that season, and I get something that looks like this.

220
00:23:48,170 --> 00:23:53,390
So this is called a violin plot because it looks exactly like a violin.

221
00:23:53,390 --> 00:24:03,200
And what it does is it shows me the distribution of points for all of the teams that were playing at home or all of the teams that were playing away.

222
00:24:03,200 --> 00:24:07,700
So essentially, where this plotting is white,

223
00:24:07,700 --> 00:24:15,260
the show is that there are lots of teams that scored that number of points with that strategy, where it's very narrow.

224
00:24:15,260 --> 00:24:23,240
This means that there were a few teams that managed to score as many as 500 points with the strategy of choosing an away player.

225
00:24:23,240 --> 00:24:29,360
And by comparing these side by side, you can see quite happily that if that's what you're making,

226
00:24:29,360 --> 00:24:34,640
your decisions based on, choosing players that are at home is better than choosing players that are way.

227
00:24:34,640 --> 00:24:36,710
It's not always gotten to be true.

228
00:24:36,710 --> 00:24:44,240
You know, perhaps you happen to choose players that are away when they score points and you end up in this season's higher up here.

229
00:24:44,240 --> 00:24:52,940
Where is your friend who chooses players that are playing at home who just happen to have bad weeks end up down here?

230
00:24:52,940 --> 00:24:59,030
So it's not something which tells us exactly how to act, but it says that in general,

231
00:24:59,030 --> 00:25:06,700
if we choose players that are at home will do better than those of our rivals that choose players that are way.

232
00:25:06,700 --> 00:25:11,800
So perhaps the more contentious debate is form versus fiction.

233
00:25:11,800 --> 00:25:17,650
What should we do? Well, in exactly the same way we can measure based on player form.

234
00:25:17,650 --> 00:25:19,120
But what is platform?

235
00:25:19,120 --> 00:25:28,120
So I'm going to define it as just saying for that player, we look at the last four weeks and we look at how many points that scored.

236
00:25:28,120 --> 00:25:32,050
So if you've had a good previous four weeks, you'll have a high form.

237
00:25:32,050 --> 00:25:34,480
If you've had a bad week during that,

238
00:25:34,480 --> 00:25:42,790
it'll bring your form down and we're just going to choose the player who's got the best form in terms of looking at the fixtures.

239
00:25:42,790 --> 00:25:49,630
Well, fortunately, we can use something called FDR, but FDR, but this FDR over.

240
00:25:49,630 --> 00:25:56,890
So this is the fixture difficulty rating, and this is something that the Premier League thinks is so important that they make it available,

241
00:25:56,890 --> 00:26:00,940
on the other hand, the kind of front page of what this player is.

242
00:26:00,940 --> 00:26:05,980
So this is Jamie Vardy looking ahead for the season that's coming up,

243
00:26:05,980 --> 00:26:15,130
and his first fixture against West Brom has an FDR of two suggesting that they think that's going to be an easy fixture for Leicester.

244
00:26:15,130 --> 00:26:21,070
Whereas when we get to the 27th of September, they're going to be playing Man City away.

245
00:26:21,070 --> 00:26:26,650
We just got a maximum of five in the arts, the hardest as they can get.

246
00:26:26,650 --> 00:26:34,810
So we can again look at the average of the next few matches or in this case, because we're just interested in making them captain for one week.

247
00:26:34,810 --> 00:26:42,870
We'll just look at what they have to all of their matches and decide whether to make them cuts and based on the.

248
00:26:42,870 --> 00:26:48,990
So when we flip over to these violent plots again, we get quite a clear answer.

249
00:26:48,990 --> 00:26:56,460
When we're choosing players based on whether they're informal or not, we can expect to get around 400 points more captain,

250
00:26:56,460 --> 00:27:04,800
whereas if we choose and based on just looking at how easy that fixture is, this bump here rises up.

251
00:27:04,800 --> 00:27:09,210
So there's a lot of variance. And these again, you know, it's not a done deal.

252
00:27:09,210 --> 00:27:17,040
Sometimes there'll be a good player you want to captain, even though they have a hard fixture and that might work out better.

253
00:27:17,040 --> 00:27:27,430
But we can say is a model in general, and you're better off looking at the fixture difficulty than you are looking at the form for a captain.

254
00:27:27,430 --> 00:27:32,230
So it's not something we would expect, so it's not something that we would.

255
00:27:32,230 --> 00:27:39,340
Well, one way that we might think about this is by considering something called the gambler's fallacy.

256
00:27:39,340 --> 00:27:46,860
So. A good way of addressing this is to think about something like dice rolls.

257
00:27:46,860 --> 00:27:51,180
So if I roll the dice 10 times, I can count how many times so I can have one.

258
00:27:51,180 --> 00:27:58,200
How many times do I get a two? Etc., etc. And I did that and I got this.

259
00:27:58,200 --> 00:28:06,900
I didn't pull any buttons at all. My dice is clearly bad. It doesn't give me any one's fine, but it gives me lots of threes and fights.

260
00:28:06,900 --> 00:28:12,350
So if I said I want to predict what the next I thought was.

261
00:28:12,350 --> 00:28:17,530
Based on just looking at this, I might say, well, I think it's good of all the three or five.

262
00:28:17,530 --> 00:28:23,020
It's definitely not going to be a one. And that's that's what the gambler's fallacy is, in a nutshell,

263
00:28:23,020 --> 00:28:30,280
is saying a look at these past events, which are random and I'll make my decision based on those.

264
00:28:30,280 --> 00:28:32,920
My coin has come up heads 10 times in a row.

265
00:28:32,920 --> 00:28:40,210
Therefore, it's more likely to be heads or, hey, I'm due a tale I haven't had a tail in while must be a tails.

266
00:28:40,210 --> 00:28:50,240
But actually, all of these things are random. The probability of of something random happening doesn't change based on on you.

267
00:28:50,240 --> 00:28:57,910
Yeah, based on the roll of the dice, you've just had a flip of the coin that you've just had these things that are independent.

268
00:28:57,910 --> 00:29:06,130
So if I roll more, more dice roll coin, but you won't get any numbers if I roll more dice.

269
00:29:06,130 --> 00:29:11,320
You can see up to $100. I ended up with a lot more ones. Great.

270
00:29:11,320 --> 00:29:16,930
So this purple line here is what I expect the distribution of dice rolls to be.

271
00:29:16,930 --> 00:29:22,390
I expect that I should be getting all of these things equally.

272
00:29:22,390 --> 00:29:25,180
So what happens if I roll more?

273
00:29:25,180 --> 00:29:35,570
Well, I get closer and closer to getting the real distribution after doing a dice all 100000 times, and that's pretty much a straight line to.

274
00:29:35,570 --> 00:29:40,640
So what does this tell us about cabinet choices?

275
00:29:40,640 --> 00:29:47,610
Well. When something's random, the past performance doesn't necessarily predict the future performance.

276
00:29:47,610 --> 00:29:53,010
So this graph here is showing Jamie Vardy again, and we want to know.

277
00:29:53,010 --> 00:29:57,240
Over the course of the last season, how often did he score?

278
00:29:57,240 --> 00:30:01,590
Zero points, one point two points? So you can see this.

279
00:30:01,590 --> 00:30:07,890
This peak here shows that about 35 percent of the time he only scored two points or less.

280
00:30:07,890 --> 00:30:12,850
But sometimes he scored as many as 16 20 points.

281
00:30:12,850 --> 00:30:16,370
So. Can we think of this as like dice roll?

282
00:30:16,370 --> 00:30:21,770
Can we think of what the distribution would be that this would come from and what we can?

283
00:30:21,770 --> 00:30:27,890
We might do something squiggly like this and say, well, it's far more likely to get a two than a four.

284
00:30:27,890 --> 00:30:33,110
So we'll put a big peak there. But this feels a lot like we're over forgetting something here.

285
00:30:33,110 --> 00:30:36,740
We probably want some smooth approximation to this.

286
00:30:36,740 --> 00:30:44,300
And there are there are ways that we might be able to predict mathematically exactly what distribution we expect this to be from.

287
00:30:44,300 --> 00:30:46,040
We're not going to talk about this today.

288
00:30:46,040 --> 00:30:55,910
Well, I'm going to do is I'm just going to say if there is some distribution here where week to week, we expect Jamie Vardy to score this many points.

289
00:30:55,910 --> 00:30:57,800
So most often [INAUDIBLE] be well.

290
00:30:57,800 --> 00:31:06,140
This bump is around three or four points, but occasionally [INAUDIBLE] be out here in the tale, scoring 19 20 points.

291
00:31:06,140 --> 00:31:12,830
If if his score from week to week isn't correlated is is truly random,

292
00:31:12,830 --> 00:31:20,690
then we don't know whether he's going to score highly in any given week and choosing him as my captain based on that.

293
00:31:20,690 --> 00:31:25,070
Well, I don't know whether he's going to be in the tail or in this bump.

294
00:31:25,070 --> 00:31:29,450
Well, this is a bit of an oversimplification because it may not be true that it's random.

295
00:31:29,450 --> 00:31:33,860
It may be that if Jamie Vardy scores a hat trick, one week is full of confidence.

296
00:31:33,860 --> 00:31:42,980
He wants to go out and score again. The next week takes more shots. And there are obviously a lot more statistics that you can go into.

297
00:31:42,980 --> 00:31:51,380
And but in terms of an explanation for why, for why the difficulty of the fixture might be more important than the form,

298
00:31:51,380 --> 00:31:59,120
this is quite a good one because a player who has been in bad form is more likely to regain that form by playing weaker opposition

299
00:31:59,120 --> 00:32:11,060
and then than perhaps a player who is capable of scoring highly or capable of scoring poorly does against any opposition.

300
00:32:11,060 --> 00:32:20,310
So. What about comparing all of these strategies side by side and how do we how do we do Typekit basically?

301
00:32:20,310 --> 00:32:25,770
So these two on the left are our form vs. fiction story.

302
00:32:25,770 --> 00:32:30,870
And again, we've got the home vs. away story, but I've also included three more strategies in here.

303
00:32:30,870 --> 00:32:34,290
Do I just pick a random captain who might be Kevin Devono?

304
00:32:34,290 --> 00:32:39,450
In which case I'll score 500 points, but might be someone else, in which case I'll school less?

305
00:32:39,450 --> 00:32:45,420
Do I just pick a random captain each week and just change it based on, you know, flipping a coin?

306
00:32:45,420 --> 00:32:51,300
Or should I pick the most expensive player that I've got in my team and hope that that means they'll score well?

307
00:32:51,300 --> 00:32:57,930
All of these strategies give us different expectations of how many points we'll get.

308
00:32:57,930 --> 00:33:02,370
But it's not. Yeah, it's not a done deal.

309
00:33:02,370 --> 00:33:04,080
You can't say for certain.

310
00:33:04,080 --> 00:33:17,120
By following the strategy, I will do better than following Strategy B. But as we've said, these strategies aren't necessarily unnecessarily.

311
00:33:17,120 --> 00:33:22,760
Able to block each other, right? I I can have a player with good fixture.

312
00:33:22,760 --> 00:33:28,310
Maybe I've got two players with the same and same difficulty of fixture.

313
00:33:28,310 --> 00:33:34,940
Well, then should I look at their form as a tie breaker? Should I look if one of those players is playing at home or not?

314
00:33:34,940 --> 00:33:45,890
So again, we can use simulated results to to prioritise what happens if I pick the player that's got the best fixture difficulty and then I compare,

315
00:33:45,890 --> 00:33:50,150
do I tiebreak using that form or do I type using my at home?

316
00:33:50,150 --> 00:33:54,470
Turns out it doesn't make any difference. Same for tiebreakers.

317
00:33:54,470 --> 00:34:00,080
If I start off with the player with the best form and then use fixed difficulty as a tiebreaker.

318
00:34:00,080 --> 00:34:06,590
In fact, none of the tiebreakers seem to make an awful lot of difference, which may be says to us as fantasy football managers,

319
00:34:06,590 --> 00:34:12,170
we shouldn't worry too much about the finer details and just focus on getting in players for the easy

320
00:34:12,170 --> 00:34:19,130
fixtures and not worry too much about whether they're at home or away or whether they're in form.

321
00:34:19,130 --> 00:34:24,020
But as you can see, when I compare all of these side by side,

322
00:34:24,020 --> 00:34:29,960
you get a slight bump for starting off by looking at that fixture difficulty, but it's not a lot.

323
00:34:29,960 --> 00:34:37,850
There's still a lot of variation and here that I'm going to talk about that later. So that's that's our first model.

324
00:34:37,850 --> 00:34:42,680
But now I want to go back to this poll and think about looking at a different hypothesis.

325
00:34:42,680 --> 00:34:51,880
So. What we want to do is try and build on the type of model that we've just made and see whether we can answer any more complex questions.

326
00:34:51,880 --> 00:35:00,250
So the most complex question on here is transfer strategy. How do I decide who to bring into my team or who not to bring into my team?

327
00:35:00,250 --> 00:35:06,160
Can I build on my model in order to try and answer that question?

328
00:35:06,160 --> 00:35:10,360
Well, let's think about simplifying.

329
00:35:10,360 --> 00:35:16,780
We started off just by thinking about two to five players, so our next problem is how do we pick a starting eleven?

330
00:35:16,780 --> 00:35:22,240
How do we actually choose a team to start with because we can't make any transfers if we don't have a team?

331
00:35:22,240 --> 00:35:30,160
And that's a hard problem. It's mainly a hard problem because we have lots of these restrictions, so we need 11 players.

332
00:35:30,160 --> 00:35:38,410
We can't have any more than three from a team. So it's not as simple as just saying who does best and we have a budget to stick with them as well.

333
00:35:38,410 --> 00:35:42,610
And ultimately, we're going to be able to transfer those players in and out.

334
00:35:42,610 --> 00:35:49,000
So the first step to this to building up to this problem is to think about a set and forget team again.

335
00:35:49,000 --> 00:35:55,990
We're going to forget about transfers for a moment and just say, how do I pick 11 players within that budget?

336
00:35:55,990 --> 00:36:04,420
And I expect to do well. Well, this is quite a well known problem within months is the exact problem.

337
00:36:04,420 --> 00:36:08,410
So why are of thinking about this is to say we've got a knapsack?

338
00:36:08,410 --> 00:36:15,550
Well, what if the type of practises and it can hold up to 15 kilograms in weight?

339
00:36:15,550 --> 00:36:22,150
And I've got some items that I can take with me and they all way a different amount and they're all worth different amounts.

340
00:36:22,150 --> 00:36:32,060
So Lockwood, one pound, one kilo. The big bet would Gold Bar were three pounds, two kilos, probably not real gold.

341
00:36:32,060 --> 00:36:37,850
But then, you know, we can have community shields, which are worth a little bit more and at the top of the scale,

342
00:36:37,850 --> 00:36:43,610
most valuable is the FA Cup, which weighs a lot but also gets you a lot of value.

343
00:36:43,610 --> 00:36:50,480
So the way that the knapsack problem works is it says, how do we maximise the value of what we take with us?

344
00:36:50,480 --> 00:36:59,120
What do we put in our bag? Do we want to have a few heavy things that are worth lots of money, but we can only take maybe one or two items?

345
00:36:59,120 --> 00:37:03,920
Or do we want lots of light things stuff in as many wooden blocks as possible?

346
00:37:03,920 --> 00:37:10,760
And well, an obvious way to think about that is to say, what's the best value for money?

347
00:37:10,760 --> 00:37:19,250
You know, are wooden block is gets us £1 for every kilogram and our Community Shield gets us £1 for every kilogram.

348
00:37:19,250 --> 00:37:28,290
So it doesn't really matter if we have 15 kilos worth of community shields or wood blocks, it's still going to be worth £15.

349
00:37:28,290 --> 00:37:35,730
So when we then look at the value for money, for the other things and see the best value for money, that was our FFA Cup.

350
00:37:35,730 --> 00:37:41,820
So perhaps we got to get as many FA Cups as we can and then move down the list.

351
00:37:41,820 --> 00:37:48,120
Well, it's kind of a natural way of thinking about this problem, but it doesn't really work for fantasy football.

352
00:37:48,120 --> 00:37:56,550
The problem being that player prices vary over the course of the season, so the more people that buy a player, the more expensive they will become.

353
00:37:56,550 --> 00:38:02,400
And if you own them, you as a manager are able to sell them to get yourself a bigger budget.

354
00:38:02,400 --> 00:38:08,100
So you've essentially got a knapsack that's that's changing size over the course of the season.

355
00:38:08,100 --> 00:38:13,110
And at the same time, we don't know how many players, how many points each player is going to score.

356
00:38:13,110 --> 00:38:21,570
So we don't actually know what the value of these places makes it very difficult to say, well, who gets the most points per million pounds?

357
00:38:21,570 --> 00:38:28,960
Maybe we can look at previous seasons, but that won't necessarily tell us what's going to happen this season.

358
00:38:28,960 --> 00:38:37,510
There's also another problem, which is that although although expensive plans do well,

359
00:38:37,510 --> 00:38:44,680
our players come in different positions, so an expensive defender is going to be cheaper than a cheap forward.

360
00:38:44,680 --> 00:38:50,410
Most defenders cost around five or six million pounds inside the game and six million pounds.

361
00:38:50,410 --> 00:38:53,410
You're expecting to score a lot of points for a defender,

362
00:38:53,410 --> 00:38:59,770
whereas if you have a forward worth six million pounds, you'll maybe not expect them to get many points at all.

363
00:38:59,770 --> 00:39:06,650
So should I stock up on expensive defenders rather than expensive forwards?

364
00:39:06,650 --> 00:39:16,600
Cheap defenders. And that's the kind of problem that most fantasy football managers have right now trying to choose a team for the next season.

365
00:39:16,600 --> 00:39:23,050
So rather than try and think about points per million or trying to use price as a way of telling us the

366
00:39:23,050 --> 00:39:30,850
value because our expensive defenders will show up as being less valuable than our cheap forwards,

367
00:39:30,850 --> 00:39:38,440
I'm going to simplify the game what we're doing, we're always trying to simplify the problem and find find a way of attacking it.

368
00:39:38,440 --> 00:39:44,110
So I'm going to give every player a new value to try and bring them all onto the same scale.

369
00:39:44,110 --> 00:39:48,970
And the way I'm going to do that is I'm going to arbitrarily label them not one,

370
00:39:48,970 --> 00:39:55,420
two or three and the players that are worth not are likely to be the cheapest players in that position.

371
00:39:55,420 --> 00:39:58,360
So off to the table here with the prices that I've used.

372
00:39:58,360 --> 00:40:06,820
If you're kind of particularly interested in doing this type of calculation yourself, these are the assumptions I made.

373
00:40:06,820 --> 00:40:10,750
So you can see here the premium price bracket, the players.

374
00:40:10,750 --> 00:40:16,060
I expect to get really good value for money or a lot of points.

375
00:40:16,060 --> 00:40:27,520
And these are the players I expect to be getting a lot of points and a very different price points, depending on their position.

376
00:40:27,520 --> 00:40:35,950
So a premium goalkeeper might cost £6million, but the premium midfielder is more like 10 million.

377
00:40:35,950 --> 00:40:41,280
So. That basically means I can now just forget about cost for a minute.

378
00:40:41,280 --> 00:40:46,500
I'm not going to worry about my £100 budget, my 100 million pound budget.

379
00:40:46,500 --> 00:40:50,130
I'm just going to worry about this No. Zero one two three.

380
00:40:50,130 --> 00:41:00,060
That's like the weight in the knapsack problem. And I'm going to choose 11 players with a combined value of up to 21 y 21.

381
00:41:00,060 --> 00:41:07,330
Completely arbitrary number that I just made up. So. Models are full of assumptions that we have to make.

382
00:41:07,330 --> 00:41:15,940
And my assumption is that that probably gives me about the right number of the right balance of premium players to cheat players.

383
00:41:15,940 --> 00:41:24,520
But it is something that I'm assuming, and I'm going to be very interested in looking at different thresholds of that different sizes of my knapsack.

384
00:41:24,520 --> 00:41:32,440
It's just a simple way of being able to restrict my initial thing because it means if I want to have lots of players from that premium price bracket,

385
00:41:32,440 --> 00:41:38,410
I'm going to need to fill up the rest of my team with players that aren't likely to score many points at all.

386
00:41:38,410 --> 00:41:46,750
Or equally, I could say I'm not going to have any of these premium players, but I'm going to make everyone in the team kind of middle of the road.

387
00:41:46,750 --> 00:41:56,230
And so we can again do our 10000 simulated teams and choose these these starting values at random.

388
00:41:56,230 --> 00:42:02,230
See what the people who have lots of expensive players score compared to the people that time.

389
00:42:02,230 --> 00:42:08,090
And you'll be very surprised to hear that having more expensive players gives you more points.

390
00:42:08,090 --> 00:42:12,130
And so that's a good thing from a modelling point of view.

391
00:42:12,130 --> 00:42:16,810
We've got a result which coincides with our intuition.

392
00:42:16,810 --> 00:42:21,940
What most players want to do is try and put in the expensive plans that they think

393
00:42:21,940 --> 00:42:27,520
will score big points and then fill up the rest of their team with whoever they can.

394
00:42:27,520 --> 00:42:35,800
And in general, having having a more expansive team in terms of the value that I've assigned is better.

395
00:42:35,800 --> 00:42:41,500
But we can also say, Well, let's compare, for instance, all of the teams in this pink bar,

396
00:42:41,500 --> 00:42:50,200
all the four teams who cost 20 of my arbitrary value and we can look within there and say,

397
00:42:50,200 --> 00:42:55,960
well, how many of those players come from that top premium price bracket?

398
00:42:55,960 --> 00:43:01,990
And so that's what we do within that within those teams, which cost 20.

399
00:43:01,990 --> 00:43:06,250
We can see that those which have a higher number of top price players do better.

400
00:43:06,250 --> 00:43:12,250
So those teams, which had six of those high value three point players,

401
00:43:12,250 --> 00:43:17,020
they only have two points to play with to fill up the other five places in their team.

402
00:43:17,020 --> 00:43:23,270
And yet those teams still do better than the ones which have a much more balanced with maybe only one expensive player,

403
00:43:23,270 --> 00:43:30,370
but filling the rest of that team with one or two point plans. So that's kind of interesting.

404
00:43:30,370 --> 00:43:34,690
This is a good step towards the the way of building an initial team that we want to do.

405
00:43:34,690 --> 00:43:41,320
We're getting realistic initial teams, but we can start to look at some other questions of this model.

406
00:43:41,320 --> 00:43:46,810
So for each of these teams, I've chosen 11 players, but I've chosen them in a legal formation.

407
00:43:46,810 --> 00:43:51,700
One of the formations that you're allowed to choose. And I can compare those things.

408
00:43:51,700 --> 00:43:56,320
So do I have three defenders, four midfielders and three forwards.

409
00:43:56,320 --> 00:43:59,800
And what's imaginatively known as a 3-4-3 formation?

410
00:43:59,800 --> 00:44:06,880
Or do I have five defenders, four midfielders and one attacker, which you'll be surprised to hear is a five for one?

411
00:44:06,880 --> 00:44:11,140
And happily, it doesn't seem to make a huge difference what I do.

412
00:44:11,140 --> 00:44:14,770
There's very slight bias away from having five defenders,

413
00:44:14,770 --> 00:44:18,220
which is a point that you should remember because I'm going to come back to it in a

414
00:44:18,220 --> 00:44:25,420
moment because what we're really interested in doing is making this model more complex.

415
00:44:25,420 --> 00:44:30,100
We want to increase the model complexity so that we can think about transfers,

416
00:44:30,100 --> 00:44:34,360
but we can't really make transfers based on this because removing a player with a

417
00:44:34,360 --> 00:44:39,490
value of three is very different if they're a forward versus being a defender.

418
00:44:39,490 --> 00:44:46,630
We can't do that swap in the game because the defenders are much cheaper than the forwards and there are different positions.

419
00:44:46,630 --> 00:44:50,890
So we need to think about real prices and real transfers to build on our model,

420
00:44:50,890 --> 00:44:54,610
which means that we can't just use that simple way of choosing an initial team.

421
00:44:54,610 --> 00:45:03,100
We do need to think about the budget. So instead of allocating my twenty one arbitrary points amongst my 11 players,

422
00:45:03,100 --> 00:45:08,020
what I'm going to do is I'm going to choose a team initially completely random,

423
00:45:08,020 --> 00:45:14,200
so I'll set my formation three four three and I'll choose a random player for each of those positions.

424
00:45:14,200 --> 00:45:20,020
And if we're within the budget, I'm going to pick one of those players at random and I'm going to upgrade them.

425
00:45:20,020 --> 00:45:23,710
And all that I'm going to do is look at the players for the more expensive than them,

426
00:45:23,710 --> 00:45:30,270
and I'm going to spot them out so much more expensive and I'm going to keep doing that until I've got my team sorted.

427
00:45:30,270 --> 00:45:37,960
So this is called Pythagorean Theorem, because Pythagoras famously thought that more expensive football is probably better.

428
00:45:37,960 --> 00:45:43,350
And so we can use this model and move on.

429
00:45:43,350 --> 00:45:50,640
But hold on a moment. We haven't talked about transfers yet, but what we have done is give a different way of making an initial team.

430
00:45:50,640 --> 00:45:58,330
We've got a second model. And if we don't make any transfers, then that's exactly the same as the previous one,

431
00:45:58,330 --> 00:46:04,600
so we can compare these things together and see whether we get any difference in those models.

432
00:46:04,600 --> 00:46:10,900
So remember the formations that we looked at a moment ago? That's this box on the left here.

433
00:46:10,900 --> 00:46:17,410
But now if we look at the the new version, we can see much more of a difference in the formation.

434
00:46:17,410 --> 00:46:26,470
So in particular, think about the 5-4-1 formation. Lots of defenders, not very many and not very many attacking players.

435
00:46:26,470 --> 00:46:35,320
And yet, these teams are scoring much more than the five for one players were relative to the rest of the teams in our original model.

436
00:46:35,320 --> 00:46:42,280
So there's something funny going on there. There's also something funny going on with the the y axis.

437
00:46:42,280 --> 00:46:51,340
So you can see for our new model, our teams before doing any transfers or captains are scoring around 4500 points in a season,

438
00:46:51,340 --> 00:46:55,610
whereas in our original model, they were only scoring around twelve hundred.

439
00:46:55,610 --> 00:47:01,520
So we need to try and understand how our model assumptions are affecting the results that we're getting out.

440
00:47:01,520 --> 00:47:07,820
Well, in terms of the schools being higher, our new model actually uses the full budget.

441
00:47:07,820 --> 00:47:14,810
There's no guarantee in our previous model that we might have lots of cheap plans and be leaving lots of money on the bench,

442
00:47:14,810 --> 00:47:17,630
which is going to correlate with lots of points on the bench.

443
00:47:17,630 --> 00:47:24,110
Whereas our new model does doesn't do that, so we can expect to get slightly higher results.

444
00:47:24,110 --> 00:47:28,370
But we're also having a bit of bias because if you remember the way that I'm choosing

445
00:47:28,370 --> 00:47:35,060
plants in my new model and I just pick a player at random and I make them more expensive.

446
00:47:35,060 --> 00:47:42,260
Well, if I pick a defender at random who's a bottom of the line, four million pound defender and I upgrade them,

447
00:47:42,260 --> 00:47:48,140
there's only a few price points before they become a six million pound top of the line defender,

448
00:47:48,140 --> 00:47:54,440
whereas upgrading our bottom of the line £4million forward,

449
00:47:54,440 --> 00:47:59,900
they've got an awful long way to go before they become the most expensive player in the game at 12 million.

450
00:47:59,900 --> 00:48:05,060
So the way that I'm setting up my model is introducing a bias.

451
00:48:05,060 --> 00:48:13,850
I'm more likely to have premium, high point scoring defenders than I am to have premium point scoring midfielders and attackers.

452
00:48:13,850 --> 00:48:25,160
And I think that's why we see this, this bias here. But as I add more players into my defence, I see a step up in the points that I'm able to score.

453
00:48:25,160 --> 00:48:33,260
So in the most famous quotes in mathematical modelling, it's from George Best, who said that all models are wrong.

454
00:48:33,260 --> 00:48:42,170
Some models are useful, and that's a really important take home message is that when we're making these models, we're making simplifications.

455
00:48:42,170 --> 00:48:48,410
And those simplifications will mean that we're not we're not dealing with the exact same system.

456
00:48:48,410 --> 00:48:52,610
We are making things simpler, therefore we are making things wrong.

457
00:48:52,610 --> 00:48:55,950
But just because they're wrong doesn't mean that they're not useful.

458
00:48:55,950 --> 00:49:03,990
When we design a model, we design it to look at a specific question, and we don't need it to be perfectly accurate in all of the other areas.

459
00:49:03,990 --> 00:49:09,150
We just need to design it carefully enough to be able to address the question we're interested in.

460
00:49:09,150 --> 00:49:13,470
So with our new model, what's the question we're interested in?

461
00:49:13,470 --> 00:49:16,650
Well, we want to compare transfer strategies with each other.

462
00:49:16,650 --> 00:49:23,440
We want to know how I should be making decisions about who to bring in and who not to bring in.

463
00:49:23,440 --> 00:49:26,550
That's something that's really hard to do just from looking at the data.

464
00:49:26,550 --> 00:49:35,910
But in this mathematical modelling framework, because where we're simulating teams with a known strategy, we can start to compare that.

465
00:49:35,910 --> 00:49:41,460
So I'm going to make the assumption that I can make zero or one transfers every week,

466
00:49:41,460 --> 00:49:49,590
and there's some more complexities in the game where you can make more transfers, you can save transfers and use them in a subsequent week.

467
00:49:49,590 --> 00:49:54,000
I'm not going to worry about them for the moment, just going to say every week I can,

468
00:49:54,000 --> 00:50:00,160
I keep my team as it is or swap out one player for a different one.

469
00:50:00,160 --> 00:50:05,470
But how do we prioritise its transfers? That's the first question me to answer.

470
00:50:05,470 --> 00:50:13,870
Do I look at my team and find the player that I like the least and get rid of them and bring in whoever I can afford that is left outside?

471
00:50:13,870 --> 00:50:21,430
Well, do I look at my team and say, Well, do I look at the players that I don't have and say, which of those players do I most want?

472
00:50:21,430 --> 00:50:26,800
And then look at who in my team I have to get rid of in order to bring them in.

473
00:50:26,800 --> 00:50:31,790
And then when I'm raising my players or the players outside, how do I rate them?

474
00:50:31,790 --> 00:50:36,700
Also, look at the same things when I was investigating captaincy. Do they have good fixtures?

475
00:50:36,700 --> 00:50:44,020
Do they have home matches of any good form? And again, as before, we can compare these these strategies.

476
00:50:44,020 --> 00:50:52,390
So these first two violin plots where we value a player most if they have good fixtures coming up,

477
00:50:52,390 --> 00:50:58,420
the second two are if we value a player the most because they're in good form, scoring well.

478
00:50:58,420 --> 00:51:05,410
And finally, do we value them because they're playing at home and there's a clear winner in there,

479
00:51:05,410 --> 00:51:11,080
which is that we should make transfers based on the recent platform.

480
00:51:11,080 --> 00:51:16,480
So if you remember our captaincy model, this is completely the opposite conclusion.

481
00:51:16,480 --> 00:51:23,090
In that model, we said that the form wasn't very important and that we should just look at what the fixture is.

482
00:51:23,090 --> 00:51:27,170
Well, why is this? If you remember our gambler's fallacy.

483
00:51:27,170 --> 00:51:35,990
Example for picking a captain, we're trying to choose a player that's going to do well that specific week and

484
00:51:35,990 --> 00:51:40,490
our best chance of maximising that is to choose someone who has an easy fixture,

485
00:51:40,490 --> 00:51:46,640
good opportunities to score. But when we're making a transfer, we're looking at a longer time period.

486
00:51:46,640 --> 00:51:53,690
We want those players to sit in our team for a longer period of time and during that longer period of time, if they were plan,

487
00:51:53,690 --> 00:52:03,440
that is in good form and is able to to score well, they're more likely to to hit those the tail of that distribution.

488
00:52:03,440 --> 00:52:12,710
So think about this Jamie Vardy example. If we just have him for one week, we're more likely to hit a two point seven to twenty point two,

489
00:52:12,710 --> 00:52:19,700
whereas if we have him for the whole season, we know that we're occasionally going to get high scores in this title.

490
00:52:19,700 --> 00:52:28,250
So we don't necessarily want to bring a player in just because just because they we expect them to do well in one particular game.

491
00:52:28,250 --> 00:52:36,100
But having players that score well over the long term is something that we should be looking to get from our transfers.

492
00:52:36,100 --> 00:52:42,640
So in terms of which transfer strategy as best, we think a lot about form versus fixture.

493
00:52:42,640 --> 00:52:47,980
But actually the other question we raised is a bit more a bit more interesting.

494
00:52:47,980 --> 00:52:54,100
Should I make my transfers based on bringing a player in or based on getting a player out?

495
00:52:54,100 --> 00:53:00,490
And I think a lot of fantasy football managers see a player that they don't have and do really well,

496
00:53:00,490 --> 00:53:07,090
and they think I have to earn that and then they tear apart their fantasy football team in order to get that player in.

497
00:53:07,090 --> 00:53:16,480
And it doesn't pay off, whereas actually, although a very expensive player may have scored very well that we don't own.

498
00:53:16,480 --> 00:53:21,910
We're probably better off sticking to our guns with our expensive players and getting rid of one of these

499
00:53:21,910 --> 00:53:27,970
other players that isn't returning the points and replacing them with someone who we can't afford,

500
00:53:27,970 --> 00:53:34,360
even if they're not the person that just scored a hat trick. We still want to replace them with the best player we can,

501
00:53:34,360 --> 00:53:41,160
but our priority should be on looking at getting rid of those players that aren't performing.

502
00:53:41,160 --> 00:53:48,930
So what have we found out? Well, hopefully we found out from a scientific and mathematical point of view,

503
00:53:48,930 --> 00:53:56,280
and we can think of science as being this process of reducing complex problems down to statements that we can test,

504
00:53:56,280 --> 00:54:03,570
and we can think of mathematical modelling as being a tool which we can use to help explore these statements.

505
00:54:03,570 --> 00:54:09,960
And in particular, I want you to go away with the idea that maths and science aren't just subjects that you study at school.

506
00:54:09,960 --> 00:54:14,100
The ways of thinking about the world. Ways of tackling complex problems.

507
00:54:14,100 --> 00:54:22,740
Whether those problems are fluid mechanics. And whether a how a plane can fly best or Formula One car can drive best.

508
00:54:22,740 --> 00:54:28,620
Whether those problems are to do with with cancer research or modelling the spread of coronavirus,

509
00:54:28,620 --> 00:54:32,970
or whether those problems are to do with choosing your fantasy football team,

510
00:54:32,970 --> 00:54:38,850
maths and science a mindset that you can take to the world to tackle problems and to break those problems

511
00:54:38,850 --> 00:54:45,960
down to ideas that you can test also want you to remember the mathematical models aren't perfect.

512
00:54:45,960 --> 00:54:53,190
They can't tell us exactly what's happening because by definition, a mathematical model is a simplification of the world.

513
00:54:53,190 --> 00:54:58,170
But that doesn't mean that they're not useful. They may be wrong, but they may be useful.

514
00:54:58,170 --> 00:55:08,700
Remember, George Box mathematical models can give us a really deep insight into how a system works, which parts of that system are important,

515
00:55:08,700 --> 00:55:21,240
and which parts we should focus our efforts on, whether that's making transfers or whether that's prioritising how we treat the disease.

516
00:55:21,240 --> 00:55:26,970
So what do we take away from this in terms of fantasy football? Well, did I do the right thing?

517
00:55:26,970 --> 00:55:32,130
I have some hypotheses. Did I win because I was following them?

518
00:55:32,130 --> 00:55:37,200
Well, in terms of my new hypotheses and definitely going to keep focussing on getting rid of

519
00:55:37,200 --> 00:55:42,840
underperforming players rather than bringing in players who have just done well recently.

520
00:55:42,840 --> 00:55:46,590
But I'm going to start thinking a lot more about fixtures over form when I'm setting

521
00:55:46,590 --> 00:55:51,840
my captain and more about four major fixtures when I'm making my transfers.

522
00:55:51,840 --> 00:56:00,630
It's not to say that I'm going to ignore the fixtures, but I'm definitely going to take these insights into account when I'm making my prioritisation.

523
00:56:00,630 --> 00:56:06,870
Probably most importantly, I'm still not going to pick a lot from Norwich, and that's still a good hypothesis to have,

524
00:56:06,870 --> 00:56:16,230
even even when they're no longer in the Premier League and just in general to take it as gospel truth.

525
00:56:16,230 --> 00:56:20,910
So can you tell us how to win fantasy football? Well, not really.

526
00:56:20,910 --> 00:56:26,520
But also sort of it can tell us which strategies are most likely to be successful.

527
00:56:26,520 --> 00:56:32,850
But all of the graphs that we've seen have got this huge variation where by following the same strategy,

528
00:56:32,850 --> 00:56:41,730
you can end up with a low number of points or a high number of points. And that variance is it's part of the game.

529
00:56:41,730 --> 00:56:50,240
It's there for a few different reasons. Just because you're making transfers based on statistics doesn't mean that you understand football.

530
00:56:50,240 --> 00:56:54,780
In order to make sensible decisions, you need to think about plans that are going to do well.

531
00:56:54,780 --> 00:57:03,620
Plans that you think looked tired, plans that you think always perform well at home, they always perform well on a Friday night.

532
00:57:03,620 --> 00:57:10,890
They always perform well when it's raining. You bet. All of that football knowledge is what goes into reducing that variance.

533
00:57:10,890 --> 00:57:19,770
But at the same time, there's a lot of luck involved. And you may be making perfectly good decisions, but it just didn't pay off.

534
00:57:19,770 --> 00:57:25,980
You know, that player did everything they could, but they had a million shots on goal and they just still got sites.

535
00:57:25,980 --> 00:57:32,640
They were just unlucky. And if your players are unlucky, then as a football manager, you're unlucky.

536
00:57:32,640 --> 00:57:39,960
And so remember, the time variance is always going to be there and remember that it's a game and go and enjoy it.

537
00:57:39,960 --> 00:57:48,540
And in the first group that I haven't butchered and other than the addition of fantasy football, it's a serious business.

538
00:57:48,540 --> 00:57:54,100
Take it seriously. Think about the maths. Thank you very much for listening.

539
00:57:54,100 --> 00:58:00,510
I just want to say a huge thanks to a bust of Anand, who I've never met and never talked to.

540
00:58:00,510 --> 00:58:07,500
But who does have all of this data available on GitHub? And and that's the dataset that I've been looking at for this talk.

541
00:58:07,500 --> 00:58:11,220
So thanks very much to him for compiling all of that.

542
00:58:11,220 --> 00:58:17,010
If you want to have a goal of beating me and beating the other mathematicians that are watching this talk,

543
00:58:17,010 --> 00:58:23,040
you have a fantasy football players that are watching this talk and hop on to the Premier League website and make a team.

544
00:58:23,040 --> 00:58:32,040
You've got until until an hour before the first Kick-Off, which is it's either Saturday, the 12th for might be an ugly game.

545
00:58:32,040 --> 00:58:36,090
Think it's Saturday? And, yeah, make yourself a team.

546
00:58:36,090 --> 00:58:40,260
Put in this code here to KB K2 and come and join us.

547
00:58:40,260 --> 00:59:04,339
Come, come and play God. Thank you very much for listening.