1
00:00:17,470 --> 00:00:19,400
So thanks to everybody for showing up today.

2
00:00:19,420 --> 00:00:25,480
I must say I'm a little intimidated because I'm not used to speaking in the same series with Roger Penrose and Stephen Hawking.

3
00:00:25,870 --> 00:00:33,160
So but I'll do my best. I'm I must say, John, you did a great job in that introduction.

4
00:00:33,760 --> 00:00:39,160
It's a little embarrassing, in a way. I sometimes wonder if I have a sort of Don Quixote tendency.

5
00:00:40,360 --> 00:00:44,950
But, you know, now I'm trying to do something useful with my life. So I'm going to tell you about that part.

6
00:00:45,640 --> 00:00:53,040
And, you know, when you think about the world and what we do as scientists in particular,

7
00:00:53,050 --> 00:00:58,240
well, any kind of science, I would argue we're contributing to the collective awareness.

8
00:00:59,700 --> 00:01:05,709
You know, models that we as a culture and society use to understand our world, to think about the cause,

9
00:01:05,710 --> 00:01:10,540
effect relationships, to understand what our options are and where they will lead us.

10
00:01:11,620 --> 00:01:15,700
And you might think about three levels of awareness that we have.

11
00:01:16,960 --> 00:01:19,630
One starts with understanding the external environment.

12
00:01:20,140 --> 00:01:29,550
This goes all the way back, really, for millennia to people like Ptolemaic, Ptolemy and Archimedes and people like that.

13
00:01:29,560 --> 00:01:40,420
But when Newton solidified the laws of gravitation and mechanics and we now have the ability to,

14
00:01:40,420 --> 00:01:42,880
for example, predict the motions of planets quite well,

15
00:01:43,540 --> 00:01:49,540
and this becomes a practical thing if you think about a problem like avoiding asteroid collisions,

16
00:01:50,080 --> 00:01:53,080
which we know asteroids have hit the earth many times.

17
00:01:53,080 --> 00:01:58,840
They've done enormous damage comparable to what we might expect from a nuclear war.

18
00:01:59,380 --> 00:02:05,280
And so it's a rather serious challenge, actually. And I thought this was something we had looked till I looked into it.

19
00:02:05,290 --> 00:02:07,930
I was going to use this as an example of something we really had nailed.

20
00:02:08,410 --> 00:02:13,630
And then I started Googling around and I found the NASA's Asteroid Grand Challenge,

21
00:02:14,080 --> 00:02:21,940
and they were quite proud of the fact that we've identified 95% of the asteroids that are greater than one kilometre in radius.

22
00:02:22,540 --> 00:02:25,810
One kilometre radius asteroid would really clobber the earth.

23
00:02:26,080 --> 00:02:35,410
So there's still 5% of them that we don't know are out there. And and if you get down to 300 metre asteroids, we've only identified 10%.

24
00:02:35,770 --> 00:02:41,400
Those could be pretty nasty, too. So even with something like that, we're still actually trying to do things.

25
00:02:41,460 --> 00:02:51,700
Interestingly, they're using crowdsourcing and relying on volunteers and so on, because the issue is if we see the asteroid, say,

26
00:02:51,700 --> 00:02:58,090
ten years before it hits the earth, we can predict things pretty well, but we have to actually see it and we're not looking actively.

27
00:02:59,920 --> 00:03:03,370
Another good example would be weather prediction there.

28
00:03:03,400 --> 00:03:12,650
I think that's a success story. You know, we've been actually predicting the weather for more than a century, initially through subjective methods.

29
00:03:15,490 --> 00:03:23,680
Around the turn of the 19th to 20th century, people started to be able to make weather maps and they began just looking at

30
00:03:23,680 --> 00:03:27,340
these weather maps and seeing patterns subjectively and what was happening.

31
00:03:27,670 --> 00:03:36,070
And professional weather forecasters emerged who would have a catalogue of past weather maps and they would look at the weather map for the day.

32
00:03:37,690 --> 00:03:42,250
You'll remember there were computers back then. They go, Oh, this reminds me of 1925.

33
00:03:42,640 --> 00:03:49,020
And they'd flip through their book and find an example of a weather map that matched sort of today's weather pattern.

34
00:03:49,390 --> 00:03:54,760
Go well, the wind's 20 degrees off and it's five degrees more like your finger.

35
00:03:54,760 --> 00:04:00,910
And then they would make a weather forecast that would be published in the newspaper and people would go on picnics and frequently get rained on.

36
00:04:02,590 --> 00:04:10,390
Now, you know, starting in the fifties, in 1951, Norman was looking for peacetime applications of the digital computer,

37
00:04:10,780 --> 00:04:14,559
and he teamed up with the best meteorologists of the day. In particular.

38
00:04:14,560 --> 00:04:21,100
In the end, a guy named Jules Charney led the charge to they worked hard,

39
00:04:21,100 --> 00:04:25,300
mathematically created a simple one dimensional partial differential equation,

40
00:04:25,750 --> 00:04:33,160
which they simulated on the computer with the help of John Van Norman's wife, who actually did the programming to put it on the computer.

41
00:04:33,640 --> 00:04:38,090
They made a forecast for a couple of weather patterns 12 hours ahead.

42
00:04:38,120 --> 00:04:43,180
And the forecast. Each of them took them 24 hours. So there were well below real time at that point.

43
00:04:44,590 --> 00:04:49,060
But, you know, those forecasts came out not too bad. And so they declared victory.

44
00:04:49,690 --> 00:04:53,860
Billions of dollars got pumped into improving physical weather forecasts.

45
00:04:54,340 --> 00:04:59,469
And I was actually around at the tipping point because in 1980 I was an intern at the

46
00:04:59,470 --> 00:05:05,910
National Centre for Atmospheric Research with working with Edward Lorenz of Chaos Fame.

47
00:05:06,430 --> 00:05:12,400
And and I asked him, Ed, you know, how how are you guys doing relative to this objective weather forecasters?

48
00:05:12,880 --> 00:05:16,300
And it was a very you know, he's a very mild mannered guy, said, well.

49
00:05:16,980 --> 00:05:21,510
We're not doing too bad and we're going to start beating them real soon.

50
00:05:22,200 --> 00:05:25,529
And he was right. 1980 was a tipping point.

51
00:05:25,530 --> 00:05:30,990
But no, that took 30 years to just break even with these guys with their libraries of weather maps.

52
00:05:31,260 --> 00:05:42,240
Now, since 1980, weather forecasts are vastly better than they were in 1980, but your picnic is much less likely to get rained on if the weather.

53
00:05:42,420 --> 00:05:47,180
Weather Forecasting Service says that it won't happen. No.

54
00:05:47,190 --> 00:05:48,540
So weather prediction we're pretty good at.

55
00:05:48,900 --> 00:05:55,170
Now, we're also starting to think about our effect on the environment and climate change is a great example of that.

56
00:05:55,830 --> 00:06:01,049
When you think about it, you know, climate change modellers get a lot of flak, but I think they're doing a pretty good job.

57
00:06:01,050 --> 00:06:10,230
They're telling us something that's rather subtle that's going to happen 50 years from now that will have very profound consequences on the earth,

58
00:06:10,650 --> 00:06:16,020
and even giving us pretty quantitative information about what we have to do to keep it from happening.

59
00:06:17,910 --> 00:06:22,379
So now note that climate change grew out of weather prediction.

60
00:06:22,380 --> 00:06:29,160
They were able to build on a very solid foundation. But we are understanding increasingly more and more about our effect on the environment.

61
00:06:29,550 --> 00:06:38,250
But I think the really hard challenge is to understand our collective effect on ourselves as our world becomes increasingly more artificial.

62
00:06:38,760 --> 00:06:41,339
As you know, there are various estimates,

63
00:06:41,340 --> 00:06:51,780
but many people estimate that we spend about 40% of the energy that's used to fix carbon on the Earth is ultimately going to support humans.

64
00:06:52,140 --> 00:06:59,850
So we're placing a huge load on the planet. And, you know, if you just think about your day, you mostly live in a pretty artificial environment.

65
00:07:00,870 --> 00:07:04,560
So somehow we need to start understanding better our collective effect on ourselves.

66
00:07:04,920 --> 00:07:11,400
Which brings us back into social science and the challenge of understanding things in a quantitative way.

67
00:07:11,730 --> 00:07:15,660
In some areas, like traffic forecasting, for example, we've gotten pretty good.

68
00:07:16,020 --> 00:07:22,770
You know, if I'm these days, if I'm driving back from somewhere at rush hour,

69
00:07:22,770 --> 00:07:30,120
I'll pull out my iPhone and I'll check to see what the traffic's doing and it can reroute me if traffic's messed up.

70
00:07:30,330 --> 00:07:37,860
It does a pretty good job of predicting things like that. But now what about the bigger, bigger things like business cycles, for example?

71
00:07:38,400 --> 00:07:41,970
Well, so here is an example of a plot of a business cycle.

72
00:07:42,300 --> 00:07:45,960
And what you can see is various measures. And I'll come back to some in a minute.

73
00:07:46,260 --> 00:07:52,260
Things like GDP and unemployment belts up and down by large factors as a function of time.

74
00:07:52,620 --> 00:08:06,210
And you can see that we have some big events like this is the crisis of 2008, which the Dallas Fed estimated cost the US between six and $15 trillion.

75
00:08:06,630 --> 00:08:11,370
That's quite a lot of money. That's like a half a year of U.S. GDP.

76
00:08:11,380 --> 00:08:15,180
And for the world, the answer is certainly a lot more than that.

77
00:08:15,570 --> 00:08:19,950
At least many months of GDP, probably a half a year, it suggests.

78
00:08:20,220 --> 00:08:26,040
If that hadn't happened, we could have all gone on vacation for six months and been about even with where we are.

79
00:08:27,150 --> 00:08:31,740
Somehow we didn't get to do that. But so this is an expensive thing.

80
00:08:31,740 --> 00:08:35,520
It costs us as a society a huge amount to have something like that.

81
00:08:35,730 --> 00:08:38,940
And it's not just the rich fat cats that suffer.

82
00:08:39,540 --> 00:08:48,630
It hits the poor the most of all. You know, at this point, unemployment for Greeks under 25 is at 80%.

83
00:08:49,350 --> 00:08:52,899
It's pretty horrendous. It's having a huge toll on Greek society.

84
00:08:52,900 --> 00:08:54,600
You have some very close Greek friends.

85
00:08:54,900 --> 00:09:00,460
They say it's depressing to be in Athens these days because people are just so dispirited about what's going on.

86
00:09:00,930 --> 00:09:07,440
And so here are some examples of what what crises do to unemployment from some arbitrary starting point.

87
00:09:07,770 --> 00:09:17,640
You can see it's not that uncommon for unemployment to swing down to as much as 20% lower than where it was before the crisis started.

88
00:09:17,650 --> 00:09:25,020
So these things hurt. And I would argue I would say there's even a pretty good argument that the Brexit vote might have gone the other

89
00:09:25,020 --> 00:09:30,690
way if we hadn't had the crisis of 2008 and that Donald Trump would not have been nominated for president.

90
00:09:32,980 --> 00:09:37,010
Okay. Now, 2008 wasn't the only crisis.

91
00:09:37,030 --> 00:09:42,300
I already showed you several on this plot here. Financial panics are endemic.

92
00:09:42,310 --> 00:09:47,110
I. This really hit me. It's a long story, so I'm going to give you the short version.

93
00:09:47,110 --> 00:09:53,139
But I happened to be in the Treasury building. I was lobbying for something involving the Office of Financial Research,

94
00:09:53,140 --> 00:09:59,530
which I played a small role in helping start the remarkable grassroots movement of various scientists and

95
00:09:59,530 --> 00:10:05,919
Wall Street practitioners who were shocked by the fact that when Lehman Brothers was about to fail in 2008,

96
00:10:05,920 --> 00:10:13,960
we didn't have a clue who their counterparties were and what effect it would have on the financial system if if we didn't prop them up.

97
00:10:14,560 --> 00:10:22,180
And so much of legislation got the text for legislation got developed by a committee of which I played a small part.

98
00:10:22,440 --> 00:10:27,820
It's about 20 pages. It actually went verbatim into the Dodd-Frank bill and got enacted.

99
00:10:28,270 --> 00:10:34,010
And it's an author's to collect all the data together and allow us to really think systematically about markets.

100
00:10:34,030 --> 00:10:35,410
Now, I was in the Treasury building,

101
00:10:35,650 --> 00:10:44,200
actually lobby the Deputy Secretary of the Treasury to to appoint somebody who had a strong science background to the post.

102
00:10:44,830 --> 00:10:48,130
I was not successful, but that's another story.

103
00:10:49,690 --> 00:10:55,929
But as I was leading, I started to walk out of the building and I noticed through these portraits of the past secretaries of the Treasury.

104
00:10:55,930 --> 00:11:00,129
So I started, you know, looking at them and reading the little inscriptions below.

105
00:11:00,130 --> 00:11:03,810
And it's kind of as I was going backwards in time, boring, boring, boring.

106
00:11:04,060 --> 00:11:08,500
But then I got to the ones that were actually dead when the inscriptions had been written,

107
00:11:08,500 --> 00:11:11,920
and then the inscriptions got much more truthful and much more interesting.

108
00:11:12,430 --> 00:11:20,079
And I noticed that about every other one would say either did a great job of dealing with the

109
00:11:20,080 --> 00:11:26,110
financial panic of the year or did a terrible job of dealing with a financial panic of X year.

110
00:11:26,530 --> 00:11:32,290
And after I got out of there, I actually looked it up and I made a list of these, the financial panics.

111
00:11:32,560 --> 00:11:38,650
As you can see, there's a lot of them. There were routine events back in the old days now.

112
00:11:39,790 --> 00:11:43,990
So and I can't resist throwing the some of these pictures in just for the graphics.

113
00:11:45,070 --> 00:11:55,030
Cartoonists have been good for a while. A particularly important panic was the panic of 1907 that was triggered by events in copper markets.

114
00:11:55,750 --> 00:12:03,730
It looked like the whole economy was on the verge of collapse, and JPMorgan rounded up all the big players on Wall Street,

115
00:12:04,000 --> 00:12:10,720
put them in his office and said and basically gave them orders about what to do to prop up the financial market.

116
00:12:11,140 --> 00:12:16,959
This ended up being hugely controversial because people were understandably suspicious of a guy like Jp morgan.

117
00:12:16,960 --> 00:12:24,910
Did he profit from it? He actually later felt quite, let's say, frustrated because he felt like he'd done the world a good turn.

118
00:12:25,150 --> 00:12:33,280
But this led finally for the US to start a central bank, the Fed in 1913.

119
00:12:34,450 --> 00:12:41,589
And by the way, since then you'll notice if you go back and look at my list panics up until the Fed was

120
00:12:41,590 --> 00:12:47,650
formed came about every little less than eight years on average since the Fed was formed.

121
00:12:47,920 --> 00:12:56,620
Well, we had a doozy, obviously, in 1929 and another big one up here, but they've been somehow bigger but less frequent.

122
00:12:56,620 --> 00:13:00,969
So it seems that it has made a difference, but bigger but less frequent.

123
00:13:00,970 --> 00:13:07,840
They're still happening. We still don't really understand when as they're coming on us, we don't understand they're coming.

124
00:13:08,170 --> 00:13:12,250
And and it's still debated how we should deal with them once we're there.

125
00:13:12,460 --> 00:13:15,010
Witness the big debates about austerity.

126
00:13:15,550 --> 00:13:25,300
Now, given this, you might think just if you go back and do the numbers, if we're talking, say, $50 trillion that a crisis can cost us.

127
00:13:26,140 --> 00:13:29,020
Well, how much research should we be doing to deal with that?

128
00:13:29,860 --> 00:13:36,850
Well, just to take some numbers, the National Science Foundation for Economics and social science in the United States is $30 million a year,

129
00:13:37,150 --> 00:13:42,130
not much the US Federal Reserve and it's like 12 branches.

130
00:13:42,490 --> 00:13:48,160
It spends a total of about $500 million. Some fraction of that goes for research.

131
00:13:48,490 --> 00:13:54,490
And I made a ballpark estimate of how much money gets paid out to academic economists throughout the world.

132
00:13:55,210 --> 00:13:56,650
Maybe it's $1,000,000,000.

133
00:13:57,830 --> 00:14:06,649
So, you know, that's not much when you think that the mistakes that are happening can cost us $50 trillion, you know, $50 trillion.

134
00:14:06,650 --> 00:14:11,300
That's five times ten to the 13 versus ten to the nine.

135
00:14:11,570 --> 00:14:16,340
So we're talking about the order of 10,000 or 50,000 difference.

136
00:14:16,580 --> 00:14:24,710
So even if there's low odds of of making things more successful, it looks like investing more money in economics should be a really good investment.

137
00:14:25,730 --> 00:14:31,280
I'll note just as a reference point, $1,000,000,000 is about what the US budget is for polar research.

138
00:14:31,680 --> 00:14:35,270
Now I'm a great advocate of polar research. I don't want them to cut that one little bit.

139
00:14:35,720 --> 00:14:39,110
But I it does seem like a dichotomy to me.

140
00:14:39,650 --> 00:14:44,990
Now. No, but you ought to ask, of course, would this more funding be effective?

141
00:14:46,070 --> 00:14:52,190
And there are several arguments that you might make to say that it's not maybe the economy is inherently unpredictable.

142
00:14:52,340 --> 00:14:58,340
For example, maybe there's fundamental limits like market efficiency that say it can't happen.

143
00:14:58,610 --> 00:15:03,050
Or maybe social science is a quixotic quest in and of itself.

144
00:15:03,080 --> 00:15:09,400
Maybe people can't be described by law as people like George Soros, who ultimately pays my salary,

145
00:15:09,410 --> 00:15:15,140
have suggested that that's true, though he's still paying my salary, so he must not think it's a complete lost cause.

146
00:15:16,640 --> 00:15:24,230
And now maybe there's another effect, though, which is maybe the culture of economics doesn't scale with more funding.

147
00:15:25,910 --> 00:15:29,660
And I'll note that they are unlike, say, meteorology,

148
00:15:29,660 --> 00:15:36,559
where you can bet every budget cycle they're there in Congress saying we need more money, we need to spur computers,

149
00:15:36,560 --> 00:15:39,440
we need better satellites, we need better measuring instruments,

150
00:15:39,440 --> 00:15:43,820
we need more money because we really we're always pushing the limits of what big science can do.

151
00:15:44,030 --> 00:15:50,659
Economics is not doing that. Now, I just want to note that these arguments are only valid if they're really strong,

152
00:15:50,660 --> 00:15:53,850
because remember, there's that factor of 10,000 or 50,000 there.

153
00:15:54,530 --> 00:16:02,600
So to to think about this question, let's step back now and let's think about the standard template for an economic model.

154
00:16:02,630 --> 00:16:06,470
How do economic models work? So I'm going to teach you economics in about 3 minutes.

155
00:16:06,890 --> 00:16:15,650
A challenging task, but it's only made a little easy and that economic models all follow more or less this template.

156
00:16:16,670 --> 00:16:21,410
So what do economists do? They assume that agents have preferences and beliefs.

157
00:16:22,280 --> 00:16:25,370
The preferences might come in the form of a utility function.

158
00:16:25,670 --> 00:16:29,660
The beliefs might come in in the form of something called rational expectations.

159
00:16:29,900 --> 00:16:33,650
So they they optimise that utility function perfectly.

160
00:16:34,370 --> 00:16:42,860
Then they just assume there's a fixed point equilibrium where agents will maximise their preferences according to their beliefs.

161
00:16:43,580 --> 00:16:46,810
And so then they compute what everybody's actions are.

162
00:16:46,820 --> 00:16:54,770
And then from there you can compute prices and quantities and all the other things that are come out in the economic model.

163
00:16:55,220 --> 00:17:05,600
And so this programs really hit its stride back in the fifties, maybe reached its peak about 1980.

164
00:17:06,260 --> 00:17:13,460
And and then since then, the program in economics has been to modify these assumptions one at a time.

165
00:17:13,850 --> 00:17:21,620
So they, for example, in the old days, they would assume utility and perfect rationality, that everybody would know everything.

166
00:17:22,280 --> 00:17:30,400
They say, no, okay, maybe not everybody knows everything. Maybe there's asymmetric information, maybe there's institutional constraints.

167
00:17:30,410 --> 00:17:36,650
So there are some things that people can't do. Maybe only some of the agents are rational and a few of the other ones are rational.

168
00:17:36,800 --> 00:17:46,040
So they they play with these assumptions one at a time. But the thing that is invariant and they always, almost always do when I say they,

169
00:17:46,040 --> 00:17:53,030
I'm referring to the sort of mainstream, because there are some rebels out there like me that are challenging this.

170
00:17:53,630 --> 00:17:59,030
But but the sticks point assumption is something that's invariant in this.

171
00:17:59,540 --> 00:18:03,680
So I'm going to examine that is the first thing I do in this lecture.

172
00:18:04,010 --> 00:18:08,270
And before I go forward, I just want to comment that there are many kinds of equilibrium.

173
00:18:10,040 --> 00:18:16,520
You know, I was educated as a physicist, so, you know, you immediately learn about forces balancing this trivial kind of equilibrium.

174
00:18:17,330 --> 00:18:21,200
Then you learn about thermodynamic equilibrium. It's a bit more subtle.

175
00:18:21,800 --> 00:18:28,459
Energy flows are balanced in a detailed way. And but in physics, we know what equilibrium is.

176
00:18:28,460 --> 00:18:35,840
We develop a framework outside of equilibrium, and then we can test to see whether equilibrium is an appropriate assumption or not.

177
00:18:36,200 --> 00:18:41,690
When it is appropriate, it's great because it really simplifies everything and you can do the calculations really well.

178
00:18:43,830 --> 00:18:48,060
I've already referred to dynamic equilibrium. That is the notion of a fixed point.

179
00:18:48,450 --> 00:18:53,280
So where things don't change, which underlies almost all math,

180
00:18:53,670 --> 00:18:59,400
economic theory and strategic equilibrium, which is specifically what they're talking about,

181
00:18:59,670 --> 00:19:09,000
where all the agents fully take each other into account and and do that in order to consider their reasoning and in fact, in economic models.

182
00:19:09,210 --> 00:19:12,540
One of the difficulties. Well, it's a good thing and a bad thing.

183
00:19:12,870 --> 00:19:19,230
Things are greatly simplified by assuming equilibrium off the bat, but you're left wondering whether the equilibrium assumption is valid.

184
00:19:20,640 --> 00:19:25,709
The pinnacle of economic theory that this has led us to, or what are called dynamic,

185
00:19:25,710 --> 00:19:32,730
stochastic general equilibrium models and what any Haldane has dubbed a rocking horse economy.

186
00:19:33,390 --> 00:19:37,260
And the reason I say that is, if you think about the analogy to physical equilibrium,

187
00:19:38,550 --> 00:19:43,140
imagine that you have a rocking horse and you bang that rocking horse with a hammer.

188
00:19:43,560 --> 00:19:50,010
So you knock it away from its equilibrium. So now it's rocking and it starts to damp down again.

189
00:19:50,490 --> 00:19:53,639
And then you hit it with a hammer again and it rocks again.

190
00:19:53,640 --> 00:19:56,760
So that's a little bit the assumption.

191
00:19:57,210 --> 00:20:04,170
The assumption is if it weren't for shocks, the economy is in equilibrium, but zuse throw shocks down from on high.

192
00:20:04,530 --> 00:20:09,810
The shocks could be technological shocks because some technology has gotten better.

193
00:20:09,990 --> 00:20:15,670
That could be behavioural shocks because people have changed their attitudes towards things like that.

194
00:20:15,690 --> 00:20:19,860
They become more impatient for some reason, so they're discounting differently than they used to.

195
00:20:20,040 --> 00:20:25,410
The shocks knock the world out of equilibrium, the world moves back towards equilibrium and the shocks hit again.

196
00:20:25,740 --> 00:20:32,850
And these models spend a lot of effort in diagnosing what these shocks are and and and

197
00:20:33,060 --> 00:20:38,850
trying to make some predictions now to see the kind of thing that they're talking about.

198
00:20:38,850 --> 00:20:43,350
Here's another picture of the business cycle. In this case.

199
00:20:43,350 --> 00:20:50,850
We're looking at the logarithm of gross national product and over about half the little more than half a century.

200
00:20:51,240 --> 00:20:54,870
And we're looking at the deviations around the mean trend, which is upward.

201
00:20:55,350 --> 00:20:57,570
So there's already an art in how you do that stuff.

202
00:20:57,960 --> 00:21:07,560
And so so under this view, all these fluctuations are driven by these shocks hitting the world, which is always trying to go back to that mean trend.

203
00:21:08,980 --> 00:21:13,240
No. I want to comment and I'll come back later to say some more about limits to prediction.

204
00:21:13,750 --> 00:21:20,200
But when you make a model like this, you've inherently said that the noise is doing a lot of the work,

205
00:21:20,560 --> 00:21:27,750
because you're saying that the main explanation for knocking for this whole movement in something like a real business cycle model,

206
00:21:27,760 --> 00:21:32,829
I find that a bit of a strange name is that is that these shocks are hitting

207
00:21:32,830 --> 00:21:37,750
and the shocks are something that economists kept on priori cannot predict.

208
00:21:39,040 --> 00:21:44,769
It's always worth remembering that, by the way, whenever somebody talks about noise, noise is typically not fundamental.

209
00:21:44,770 --> 00:21:49,510
I mean, the only fundamental source of noise that we know of is measurement in quantum mechanics.

210
00:21:49,720 --> 00:21:56,200
Everything else just amounts to our ignorance. And not that noise models aren't useful, but they're inherently self-limiting.

211
00:21:57,010 --> 00:22:01,610
So. So now why would you expect that the economy might not be in equilibrium?

212
00:22:01,630 --> 00:22:05,140
Well, first of all, beliefs and outcomes may not be consistent.

213
00:22:05,140 --> 00:22:07,450
And when you dig in to equilibrium theories,

214
00:22:07,450 --> 00:22:15,490
you see that consistency between beliefs and outcomes is a is a deeply important part of what is going on in equilibrium.

215
00:22:15,760 --> 00:22:20,710
And as I've already suggested, the strategy dynamics may not settle into a fixed point.

216
00:22:21,430 --> 00:22:27,880
So to illustrate this, I'm going to begin by thinking about the strategic context where people are reasoning,

217
00:22:28,210 --> 00:22:35,950
and I'm going to use something called a Keynes beauty contest. Now in Keynes's Day, papers would publish.

218
00:22:36,160 --> 00:22:40,870
This is an example of a page of a newspaper from the time they would publish pictures of women.

219
00:22:41,180 --> 00:22:44,980
I admit that's a very sexist thing to do, that I don't think they ever did this with men.

220
00:22:45,400 --> 00:22:53,559
And they would they would then ask people to rate how attractive they thought these women were and people would then rate them.

221
00:22:53,560 --> 00:23:02,350
And then if you voted for the choice that was most popular, your name could be drawn out of a hat and you might win a vacation or something like that.

222
00:23:03,100 --> 00:23:10,990
And but Keynes pointed out that what matters isn't there's no inherent measure of which women are the most beautiful.

223
00:23:11,350 --> 00:23:18,190
The the criterion is, is to guess what other people will think is most beautiful.

224
00:23:18,490 --> 00:23:22,330
And he pointed out that the stock market is very much that kind of thing.

225
00:23:22,330 --> 00:23:31,870
It's not necessarily what makes prices go up isn't what company is the best company, it's what company the people think is the best company.

226
00:23:32,260 --> 00:23:37,450
So that was the context of his doing this. But I'll now give you an example of a different kind.

227
00:23:37,450 --> 00:23:38,260
In that spirit.

228
00:23:39,520 --> 00:23:47,920
You're asked to name a number between zero and 100, and the winner is going to be the one whose guess is two thirds of the average guess.

229
00:23:48,640 --> 00:23:54,580
So I want you to think about this for a few seconds to think what would your guess be?

230
00:23:55,210 --> 00:24:00,460
Now, I've done this in the past by polling people and I've discovered that it's a bit of a dangerous

231
00:24:00,460 --> 00:24:05,110
thing to do because it can be a little embarrassing once I explain what's happening.

232
00:24:05,290 --> 00:24:10,750
So I don't want to embarrass anybody. Also, I can't make my snotty remarks because I really wouldn't want to offend anybody.

233
00:24:10,900 --> 00:24:16,480
So just think about it. Formulate your guess. I want everybody to pick a number, you know, how are you going to do it?

234
00:24:17,290 --> 00:24:21,970
All right. Now, let me show you what about 30,000 Danes did.

235
00:24:22,270 --> 00:24:25,900
Now, this was published in the newspaper, so they had more chance to think about it than you did.

236
00:24:26,560 --> 00:24:33,700
And and you can see that, you know, the guess is there are people who actually guessed 100.

237
00:24:35,770 --> 00:24:41,090
Now, let's go back and think about you wonder a little bit, you know, okay.

238
00:24:41,110 --> 00:24:45,640
I mean, this kind of might explain why people are actually going to vote for Donald Trump.

239
00:24:48,700 --> 00:24:52,630
Okay. Let's suppose all of you but one, just 100.

240
00:24:53,200 --> 00:24:56,860
Okay. If everybody goes to 100, then what would the winning guess be?

241
00:24:57,040 --> 00:25:00,460
What, 66? Right, two thirds of 100.

242
00:25:01,090 --> 00:25:04,780
So obviously, anything greater than 100 is a loser.

243
00:25:05,230 --> 00:25:08,590
So all those guys that there's a lot of quite a few of them,

244
00:25:08,950 --> 00:25:16,149
either they didn't understand the game or they really weren't thinking, now let's suppose then you go, okay, I'll get 66.

245
00:25:16,150 --> 00:25:20,560
But you go, well, 66 is pretty high because other people aren't total dummies either.

246
00:25:20,890 --> 00:25:26,020
And why would they all get to 100? Well, okay, so maybe they would guess 50.

247
00:25:26,620 --> 00:25:31,150
And now let's take two thirds of that. So that would be about a third.

248
00:25:32,110 --> 00:25:36,969
So you see there's actually a peak there, but then you might go, well, okay,

249
00:25:36,970 --> 00:25:42,150
suppose people get stopped then I'm going to suppose I'm going to reason one more all two time.

250
00:25:42,160 --> 00:25:48,040
That's the typical gas. And so in fact, that's the winner around around just above 20.

251
00:25:48,820 --> 00:25:58,959
So but but note, if you were an economist and you were trained to think about rational expectations, you would have iterated this process, too,

252
00:25:58,960 --> 00:26:05,770
and you would have guessed the Nash equilibrium, which is zero, because that's where you end up if you assume that everybody is really smart.

253
00:26:05,770 --> 00:26:11,440
If we were all smart and we all knew that everybody else was. Mart, we would just go to zero and it would be a fixed point equilibrium.

254
00:26:11,740 --> 00:26:21,710
Everything would be peachy. It would be easy. So my snotty remark is, okay, these people are obviously idiots and these guys are idiots for months.

255
00:26:22,570 --> 00:26:26,500
So, all right. My snotty remark. Okay.

256
00:26:26,500 --> 00:26:33,070
So when is this assumption justified? So I'm going to show you an example of some work we did to test this in the context of game theory.

257
00:26:33,520 --> 00:26:38,480
The players can make one of several possible moves. So a game is something where players make moves.

258
00:26:38,500 --> 00:26:46,060
So it's like the games you're familiar with. At each turn, players receive payments according to the set of moves of all players.

259
00:26:46,090 --> 00:26:52,510
In other words, if on that turn, Martin makes a move and I make a move and a list makes a move,

260
00:26:52,600 --> 00:26:56,190
and we make move number, you know, listen, makes move number one. I make number two.

261
00:26:56,470 --> 00:26:58,060
Martin makes move number three.

262
00:26:58,360 --> 00:27:06,850
We then look up in the payoff matrix what are payment is and each of us get some payoff that may be quite specific to each of us.

263
00:27:07,360 --> 00:27:13,110
And so that payoff matrix defines the game. Then we're going to play that game repeatedly.

264
00:27:13,110 --> 00:27:16,049
So we're going to play it again and again so we have a chance to learn from it.

265
00:27:16,050 --> 00:27:23,940
So and by the way, I think if we played that Keane's beauty contest game repeatedly, we would have seen a progression towards zero.

266
00:27:25,500 --> 00:27:29,640
And we're good, but we're going to make the players learn their strategies, unlike the.

267
00:27:31,600 --> 00:27:34,510
Well, I guess I had excuse me this on my next slide.

268
00:27:38,000 --> 00:27:44,209
There is there is a standard theory by Nash which says that there's always a fixed point in a game like that.

269
00:27:44,210 --> 00:27:48,230
There's always an equilibrium like the zero equilibrium in the beauty contest.

270
00:27:48,530 --> 00:27:53,690
As long as players can choose moves and introduce an element of randomness to keep the other players guessing,

271
00:27:54,200 --> 00:27:56,660
as long as they can do that, there always exists to fix point.

272
00:27:56,990 --> 00:28:02,510
But Nash didn't prove it stable so we don't know that that's behaviourally what people are actually going to do.

273
00:28:02,810 --> 00:28:07,220
And the experiment I just showed you suggest that that's not what people typically do.

274
00:28:07,940 --> 00:28:13,520
But what do they do now? Oh, okay. Let me give you a little bit of intuition.

275
00:28:15,210 --> 00:28:20,010
So the first thought I have on this is simple games should be easier to learn than difficult games.

276
00:28:20,520 --> 00:28:26,130
So I'm reminded of when I was about nine or ten on the playground,

277
00:28:26,550 --> 00:28:31,770
we started playing this game called Tic TAC tOe, which I think in Britain you called nuts and crosses.

278
00:28:32,400 --> 00:28:40,500
So you make a little three by three grid, you put four zero, you try and get a, you know, a diagonal, horizontal or vertical line of connected things.

279
00:28:40,500 --> 00:28:41,640
Whoever does that first wins.

280
00:28:42,360 --> 00:28:50,190
So at some point, I figured out that I could be zero and I could always block the other guy from doing it, so I could always at least get a draw.

281
00:28:51,030 --> 00:28:57,600
And so for about a day, maybe I felt like a real hero because I was just cleaning up.

282
00:28:58,320 --> 00:29:02,070
And then all my friends figured it out and the game became completely boring.

283
00:29:02,070 --> 00:29:06,880
We quit playing it. So that's a good example of tic tac toe is a simple game.

284
00:29:06,900 --> 00:29:12,310
Children learn the Nash equilibrium. Now, what about more complicated games?

285
00:29:12,360 --> 00:29:15,810
Say, take chess. So now I'm going to tell you a little joke.

286
00:29:16,050 --> 00:29:21,320
How did two economists play chess? Well, you know, so there's a random element to chess.

287
00:29:21,710 --> 00:29:25,850
You begin by choosing white or black, and then they choose white or black,

288
00:29:25,850 --> 00:29:31,040
and then they argue about whether white or black wins under the Nash equilibrium of the game.

289
00:29:31,820 --> 00:29:36,200
Now, why is the joke a little silly? Well, because we don't know what the Nash equilibrium of the game is.

290
00:29:37,520 --> 00:29:45,679
It is a deterministic game once you've chosen white or black. No, I have to tell the corollary, the joke, because I'm a physicist,

291
00:29:45,680 --> 00:29:50,240
so it's rude to tell jokes about economists without telling a joke on your own tribe.

292
00:29:51,530 --> 00:29:58,850
So how did two physicists play chess? Well, they get quite drunk and they start moving the pieces around randomly.

293
00:29:59,360 --> 00:30:05,690
And they argue about whether the game is more or less interesting under scaling of the size of the board and the number of pieces.

294
00:30:06,590 --> 00:30:11,570
So you'll see from what I'm about to present that I'm a physicist now.

295
00:30:11,750 --> 00:30:17,030
So what we did is we said we wanted to understand generic behaviour and I did want to make sure

296
00:30:17,030 --> 00:30:21,500
that my comment was registered that there's a we're not the first to think about this issue.

297
00:30:21,530 --> 00:30:23,899
Lots of people have worried about this from the beginning.

298
00:30:23,900 --> 00:30:31,490
In fact, von Neumann himself was quite critical about game theory for its emphasis on Nash equilibrium, and in particular at Oxford.

299
00:30:31,490 --> 00:30:36,650
We have a couple of people who have been pioneers in this Paten, Yang and Vince Crawford.

300
00:30:36,950 --> 00:30:43,430
So other people have thought about this, but we took a different approach and we really wanted to understand what's generically true.

301
00:30:44,570 --> 00:30:46,310
So we constructed games at random.

302
00:30:46,880 --> 00:30:53,450
So remember, there's this payoff matrix that determines how much you get paid under the collection of moves that get made.

303
00:30:53,480 --> 00:30:56,510
We put random numbers into that payoff matrix.

304
00:30:56,930 --> 00:31:00,860
Once we put those random numbers in, we keep the payoff matrix fixed.

305
00:31:02,540 --> 00:31:06,470
And then we have the players. We give the players a learning algorithm and we have them play.

306
00:31:07,160 --> 00:31:14,360
Now our goal is a bit like we want to characterise the behaviour a priori as much as we can.

307
00:31:14,720 --> 00:31:15,260
In other words,

308
00:31:15,410 --> 00:31:22,610
we want to know are there some rules of thumb that might tell you in advance whether a game's likely to converge to an equilibrium or not?

309
00:31:23,090 --> 00:31:26,330
An analogy can be made to the Reynolds number and fluid turbulence,

310
00:31:26,780 --> 00:31:34,370
which is a number the ratio of the shear stress to viscosity nine dimensional number that numbers around one.

311
00:31:34,640 --> 00:31:39,050
Then I pretty much know the fluids laminar that is the dynamics sits at a fixed point.

312
00:31:39,890 --> 00:31:42,320
On the other hand, if that number is 10,000,

313
00:31:42,620 --> 00:31:49,130
then I have a quite I'm virtually certain that the fluid flow is going to be highly turbulent and chaotic.

314
00:31:49,760 --> 00:31:53,240
So we're looking for something a little like that now.

315
00:31:53,960 --> 00:31:59,480
So we actually did something. We didn't just populate these payoff matrices completely at random.

316
00:31:59,870 --> 00:32:01,040
We choose payoffs.

317
00:32:01,040 --> 00:32:08,150
So they were normally distributed and we constrain the playoffs and we began by studying two players and then we'll go to more players later.

318
00:32:08,600 --> 00:32:18,050
So when we have two players, we constrain them so that we have a parameter gamma that involves the relative payoff matrices of the two players.

319
00:32:18,290 --> 00:32:22,220
If game is one, the payoff matrices of the players are identical.

320
00:32:22,520 --> 00:32:25,940
So if I'm playing Martin, his payoff matrix and mine are exactly the same.

321
00:32:26,270 --> 00:32:33,770
If I win, he wins. If I'm a zero, if I win, he might well, he might lose completely uncorrelated.

322
00:32:33,980 --> 00:32:37,880
And if karma's minus one, we have a zero sum game. If I win, he loses.

323
00:32:38,720 --> 00:32:40,730
So that's an important parameter we discovered.

324
00:32:41,420 --> 00:32:47,900
We also chose a learning algorithm in which players learn their strategies using an algorithm called experience weighted attraction.

325
00:32:48,410 --> 00:32:53,120
We chose this. Under this algorithm, players choose their strategies probabilistically.

326
00:32:53,540 --> 00:32:58,040
So if I was saying rock, paper, scissors, I would have a certain probability of playing rock,

327
00:32:58,070 --> 00:33:02,330
certain probability of paper, certain probability of playing scissors on any given turn.

328
00:33:02,990 --> 00:33:06,500
Those probabilities can change dynamically at time and time,

329
00:33:06,710 --> 00:33:11,630
but on every turn I'm going to roll some dice that are going to be balanced according to those probabilities,

330
00:33:12,110 --> 00:33:15,530
to determine what how likely I'm going to select something.

331
00:33:16,070 --> 00:33:22,549
And we chose this because some economists had done experiments and shown that it's

332
00:33:22,550 --> 00:33:26,870
a pretty good model for how people actually behave in games in a laboratory.

333
00:33:27,830 --> 00:33:31,070
It was also analytically tractable for us to study it.

334
00:33:31,610 --> 00:33:37,669
Now, this this learning algorithm has a parameter alpha that determines memory because the

335
00:33:37,670 --> 00:33:43,220
probability of mimicking a given move is dependent on how well it's been working in the past.

336
00:33:43,820 --> 00:33:49,250
If Alpha Zero, the past can be infinitely long as I just accumulate more and more data.

337
00:33:49,460 --> 00:33:54,200
And something that happened years ago is equivalent to something that that happened the last turn.

338
00:33:54,860 --> 00:34:01,670
If Alpha is greater than zero, then I pay more attention to things that happened the last turn than things that happened in the past.

339
00:34:02,540 --> 00:34:06,890
Now, what did we find? And I should say one more comment that's important.

340
00:34:07,520 --> 00:34:12,910
Being physicists, we chose the limit where there's lots of possible moves.

341
00:34:12,920 --> 00:34:14,840
We looked at really complicated games.

342
00:34:15,200 --> 00:34:23,510
Now we did that because that allowed us to go into the thermodynamic limit where we can actually compute stuff using cool analytic techniques.

343
00:34:24,380 --> 00:34:27,950
So that my collaborator, Tobi Scala, is an expert in.

344
00:34:28,460 --> 00:34:33,680
So what did we find? What we found is what happens really depends on the parameters.

345
00:34:34,070 --> 00:34:40,220
This parameter alpha that determines memory of the strategy is important and the parameter.

346
00:34:41,130 --> 00:34:49,770
Gamma is also important. And there's a big region here where you just converge to a unique fixed point that's close to a Nash equilibrium.

347
00:34:49,980 --> 00:34:53,550
So that's very much like what the economists are typically assuming.

348
00:34:53,850 --> 00:34:57,330
This is the zone where traditional economics is okay.

349
00:34:58,410 --> 00:35:02,460
But then we found this chaotic zone over here that I'll tell you more about in a minute.

350
00:35:03,300 --> 00:35:08,280
And then there was so this happens, by the way, when Gamma is less than zero.

351
00:35:08,280 --> 00:35:13,500
So the games are competitive in the sense that if I win, Martin tends to lose.

352
00:35:15,360 --> 00:35:19,890
And so we saw these chaotic cases here where there was no convergence to a fixed point.

353
00:35:20,190 --> 00:35:28,470
When Gamma is greater than zero, we saw multiple fixed points, which in the large state spaces we were looking at where each player had 50 moves.

354
00:35:28,740 --> 00:35:34,260
There could be hundreds of these fix points. And which one you converge to depends on exactly where you start.

355
00:35:34,680 --> 00:35:40,740
So there's multiple basins of attraction in the dynamical system that corresponds to the learning dynamics.

356
00:35:40,980 --> 00:35:46,469
Now, what's going on in these chaotic cases when the cases so we saw lots of things.

357
00:35:46,470 --> 00:35:54,030
So we saw, for example, now here I'm going to pick three out of the 50 strategies or I should call them three out of pure strategies.

358
00:35:54,030 --> 00:35:57,240
This is another term or three out of the 50 possible moves.

359
00:35:58,080 --> 00:36:01,530
So this is the probability of making that move as a function of time.

360
00:36:02,730 --> 00:36:08,090
And what you can see in this case that corresponds to these parameters, there's just an oscillation.

361
00:36:08,100 --> 00:36:11,940
So the players are just going around around in a loop. History is repeating itself.

362
00:36:12,540 --> 00:36:19,530
And in the state space where you take a couple of the moves and you just project things out of that axis,

363
00:36:19,920 --> 00:36:25,080
you see an attractor whose dimension is essentially one, and they just go around on this loop.

364
00:36:25,740 --> 00:36:34,319
In other cases, you see things like this where you see what we're able to do to tell for using various techniques is a low dimensional,

365
00:36:34,320 --> 00:36:42,059
strange attractor. And in other cases, the dimension of that attractor gets bigger and bigger.

366
00:36:42,060 --> 00:36:47,490
It's still a strange attractor, but the dimension gets bigger and the behaviour gets more and more random.

367
00:36:47,970 --> 00:36:53,440
Here it starts to just look like a ball of yarn. It's to all intents and purposes, a random process.

368
00:36:54,000 --> 00:36:59,490
And though there are some interesting things going on here, if you look at one of the strategies here,

369
00:36:59,940 --> 00:37:03,420
the strategy is sometimes played with the probability of approaching one.

370
00:37:03,750 --> 00:37:10,770
And notice this is a logarithmic scale. And another points, the probability of the strategy being used can be ten to the minus,

371
00:37:10,770 --> 00:37:15,180
20 virtually never used, and then it can come back and become very popular again.

372
00:37:15,390 --> 00:37:24,810
So there's an automatic generation of fads of different strategies being used in this world, which has a certain realism.

373
00:37:25,260 --> 00:37:28,410
Now you can ask what happens when you go to more players?

374
00:37:28,890 --> 00:37:34,260
What changes if instead of two player games you have multiplayer games?

375
00:37:34,470 --> 00:37:42,420
Now let me first make the remark that we were able to analytically calculate this blue line to very high approximation.

376
00:37:42,990 --> 00:37:51,030
So we were quite proud of that. And it takes 20 pages of hard math to do that, but it matched pretty well with it with what we did in the simulations.

377
00:37:51,750 --> 00:37:54,750
Now, what about when we have more than two players?

378
00:37:54,780 --> 00:37:57,780
Well, when you have more than two players, by the time you get up to, say,

379
00:37:57,780 --> 00:38:03,430
ten players, it becomes really hard to simulate this for more than two moves.

380
00:38:03,450 --> 00:38:06,560
So we can't even go into that high move zone again.

381
00:38:06,570 --> 00:38:11,610
So here we really had to rely on our theory. We tested our theory as best we could.

382
00:38:12,000 --> 00:38:20,280
We pushed the theory out there, and we showed that the area of the chaotic zone increases is the square root of the number of players.

383
00:38:21,090 --> 00:38:30,030
And so, in fact, what happens is, as you add more players, this boundary just translates itself over like this.

384
00:38:30,480 --> 00:38:37,110
And chaos takes more and more of the parameter space, as does this complicated region.

385
00:38:37,290 --> 00:38:43,590
And in the limit where you have lots of players, it's you never see the classic Nash fixed equilibrium.

386
00:38:44,100 --> 00:38:45,989
Now, if you're an economist,

387
00:38:45,990 --> 00:38:53,850
this should make you a little nervous because this is suggesting that when there's a lot of players in a complicated game,

388
00:38:54,210 --> 00:38:57,360
there may be it may be the generic case.

389
00:38:57,630 --> 00:39:03,930
I would argue you have to justify that you're going to use a model that just has a fixed point, because what it means,

390
00:39:03,930 --> 00:39:09,060
if you assume a fixed point in something complicated, like I would say the stock market is a good example.

391
00:39:09,360 --> 00:39:17,640
You're you're approximating this big ball of yarn with this complicated dynamics of the strategies by a point sitting in the middle of this.

392
00:39:18,270 --> 00:39:24,420
Now it's a point with some noise. And as I said, this is for many practical purposes, a random process.

393
00:39:24,720 --> 00:39:30,630
But I think there can be key differences, like as I actually showed in one of my earlier papers,

394
00:39:30,810 --> 00:39:37,140
you can sometimes take advantage of this when there's chaos to make a prediction about the endogenous dynamics.

395
00:39:37,800 --> 00:39:41,100
And and I think it can be insightful as to the type of the model.

396
00:39:41,280 --> 00:39:47,040
And by the way, as a side effect we saw here, you see the change in the total payoff as a function of time.

397
00:39:47,970 --> 00:39:51,050
You see these periods where things are pretty quiet.

398
00:39:51,300 --> 00:39:53,340
You see periods where things are quite noisy.

399
00:39:53,520 --> 00:39:59,220
This is very reminiscent of what you see if you look at returns in the stock market, something called clustered volatility.

400
00:39:59,430 --> 00:40:02,430
This was a surprise, not something we expected to see.

401
00:40:02,970 --> 00:40:06,210
Now I have to add some caveats and I'm worried about running out of time here.

402
00:40:06,540 --> 00:40:09,179
But what happens with other learning algorithms? Well,

403
00:40:09,180 --> 00:40:18,600
we've explored some like we've explored some versions of the kind of level K learning that you you saw people using in that beauty contest example.

404
00:40:20,430 --> 00:40:25,380
We've, you know, you can you can ask whether the ensemble of games we're choosing is representative.

405
00:40:26,160 --> 00:40:29,790
You can ask about games with just a few actions, like two by two games.

406
00:40:30,370 --> 00:40:34,979
Marco Pagano, my graduate student here in math, has been looking at that,

407
00:40:34,980 --> 00:40:43,469
has actually made a a classification of what happens under all two by two games and showed that while chaos is now much less likely,

408
00:40:43,470 --> 00:40:47,370
there is still a zone where chaos happens even in two by two games.

409
00:40:48,340 --> 00:40:52,660
No. So how do you cope with this kind of nonlinear non-equilibrium behaviour?

410
00:40:52,840 --> 00:40:56,050
If you throw out those six point models, what do you do instead?

411
00:40:56,620 --> 00:41:01,479
So I would argue that you have to you have to go to simulation,

412
00:41:01,480 --> 00:41:06,670
because if you want to make any kind of realistic models, you end up with nonlinear equations.

413
00:41:06,970 --> 00:41:12,430
And nonlinear equations are typically not solvable via any other means than simulation.

414
00:41:12,880 --> 00:41:18,700
And just as a note, computation has revolutionised the physical and natural sciences.

415
00:41:19,780 --> 00:41:24,040
I think it accounts for most of the progress we've made over the last 50 or 60 years.

416
00:41:24,670 --> 00:41:32,800
And this was really driven home to me, this concept, the importance of nonlinear dynamics or nonlinear equations in general.

417
00:41:33,730 --> 00:41:38,440
When when I was a graduate student, Stan Hull came to the physics department at UC Santa Cruz,

418
00:41:38,440 --> 00:41:44,020
where I was studying and gave a lecture in which he began by apologising for using the term nonlinear.

419
00:41:44,830 --> 00:41:53,440
And he said, I'm apologising because calling an equation nonlinear is calling is like calling an animal a non elephant.

420
00:41:53,920 --> 00:42:00,310
Because most equations are nonlinear and we've been blinded to that fact by the by.

421
00:42:00,340 --> 00:42:04,749
Because up until now Ulam was one of the pi.

422
00:42:04,750 --> 00:42:09,729
Up until recently, all we could solve were linear equations.

423
00:42:09,730 --> 00:42:14,320
So we were like the mathematician sitting under the lamppost looking for our keys,

424
00:42:14,650 --> 00:42:20,050
because that was the only place there was any light, even though we knew we've dropped the key somewhere else.

425
00:42:20,680 --> 00:42:27,430
And so now I was actually one of the first people to use computers to look at nonlinear.

426
00:42:27,640 --> 00:42:33,879
Nonlinear systems was the very first example of simulating something on a computer.

427
00:42:33,880 --> 00:42:38,950
In 1945 was the Fermi past ultimate experiment in which they put masses with

428
00:42:38,950 --> 00:42:45,820
coupled springs on the computer and they expected just to see ergodic behaviour.

429
00:42:46,210 --> 00:42:50,410
They were just trying to confirm something that was standard in statistical mechanics.

430
00:42:50,590 --> 00:42:55,239
They thought the energy would be equally partitioned among the modes of oscillation and instead

431
00:42:55,240 --> 00:42:59,350
they saw solitary waves propagating back and forth the opposite of what they predicted.

432
00:42:59,740 --> 00:43:07,090
And so no social science, I would argue, has not benefited as physical and natural science have.

433
00:43:07,450 --> 00:43:13,840
I think there's two reasons for that. One is the elementary processes are not well understood, so it can be harder to put things on the computer.

434
00:43:14,110 --> 00:43:20,980
And secondly, there is different cultural attitudes. That has to do with, I think, symmetry breaking in the history of economics.

435
00:43:21,280 --> 00:43:28,150
But I'll maybe come back and say a little bit about that at the end. Now, I want to give you another example from my work of how you can deal with.

436
00:43:30,240 --> 00:43:35,910
Out of equilibrium behaviour that relates to something called leverage cycles and leverage cycles,

437
00:43:35,910 --> 00:43:41,800
as we discovered in 2008, are one of the most important examples of the dynamics of systemic risk.

438
00:43:42,480 --> 00:43:49,390
They were foreseen before 2008 by people like Hyman Minsky who wrote about them in words.

439
00:43:49,410 --> 00:43:56,100
He didn't have a mathematical model, and basic story is that during calm times, people get complacent.

440
00:43:56,400 --> 00:43:59,670
They compete for profits. Leverage creeps up.

441
00:44:00,030 --> 00:44:02,790
I'll explain what leverage is in a minute. But people take more risks.

442
00:44:03,180 --> 00:44:09,270
And then as leverage gets higher and higher, eventually some negative shocks are going to hit and they're going to cause a crash.

443
00:44:09,960 --> 00:44:15,660
JANJUA Two couples who I've collaborated with on many papers also had a theory of leverage cycles.

444
00:44:15,660 --> 00:44:22,140
He actually coined the term involving heterogeneous investors that optimists use more leverage.

445
00:44:22,350 --> 00:44:25,020
Bad news comes along that wipes out the optimists.

446
00:44:25,320 --> 00:44:32,550
And because the optimists were baked into prices before then, there's a big downward spike because they get removed from the picture.

447
00:44:33,300 --> 00:44:37,680
Now I'm going to show an even simpler model than either of these two.

448
00:44:38,490 --> 00:44:44,100
That was developed with Christoph Almond's, who was my graduate student here and is now a postdoc at the LSC.

449
00:44:46,450 --> 00:44:52,210
So it involves leverage. Now, what is that? That involves buying assets with borrowed money.

450
00:44:52,780 --> 00:45:00,640
And the leverage is a number that's the ratio, the value of those risky assets to how much money you actually put down.

451
00:45:00,670 --> 00:45:05,740
So if you buy a house and you make a down payment, that's equal to 20% of the price of the house.

452
00:45:06,100 --> 00:45:14,650
That implies a 5 to 1 leverage because the value of the house is at the moment where you purchase it 5 to 1, the ratio is 5 to 1.

453
00:45:14,800 --> 00:45:19,970
Now, of course, after that, the value of the House may change and your leverage may go up and down.

454
00:45:20,260 --> 00:45:21,940
And it turns out that's very important.

455
00:45:23,630 --> 00:45:32,450
Our model is a model that it's targeted at explaining rather ambitiously the great moderation and the subsequent crisis.

456
00:45:32,870 --> 00:45:41,120
What happened historically is that stock market prices went up over a period of a decade or so.

457
00:45:42,890 --> 00:45:50,510
Stock market volatility went down and leverage for assets of US broker dealers,

458
00:45:51,230 --> 00:45:58,820
speculative traders went up and actually skyrocketed in 2008, then crash down sent volatility really high.

459
00:45:59,060 --> 00:46:04,310
The stock market came down with it. So can we explain this kind of behaviour from a simple model?

460
00:46:05,000 --> 00:46:10,549
Now I have to say before I present this model, we begin by making a rather complicated model.

461
00:46:10,550 --> 00:46:16,940
We had a bunch of banks and a bunch of assets and the banks were managing things and initially we didn't have any risk control.

462
00:46:17,270 --> 00:46:23,120
And we said, what happens when we throw in Basel two the regulation that banks are supposed to follow?

463
00:46:23,960 --> 00:46:30,110
And we saw the whole economy going up and down and this big oscillation and we said, Holy [INAUDIBLE], what's going on?

464
00:46:30,590 --> 00:46:36,650
And we began simplifying the model and we simplified it down to the model I'm going to show you here.

465
00:46:37,460 --> 00:46:44,300
So when we strip things down to the essence, all we needed, it turned out, were two agents, a bank,

466
00:46:44,300 --> 00:46:49,400
which you can think of as any agent that using leverage and a fundamentalist who is

467
00:46:49,400 --> 00:46:54,710
an agent who isn't and who buys undervalued assets and sells overvalued assets.

468
00:46:55,990 --> 00:47:00,280
We assume there's one risky asset plus cash and we make four assumptions.

469
00:47:00,280 --> 00:47:07,930
We assume the bank uses an exponential moving average of historical volatility to estimate the expected volatility going forward.

470
00:47:08,440 --> 00:47:14,590
So what does that mean? That the bank if you go back to the kind of picture I showed you a moment ago,

471
00:47:14,920 --> 00:47:22,719
the banks who are leveraging are looking at this time series and they're using a historical average of what this has

472
00:47:22,720 --> 00:47:29,260
done over the last couple of years or so in order to decide how volatile things are likely to be going forward.

473
00:47:29,560 --> 00:47:33,070
That's actually a very reasonable thing to do because volatility is pretty persistent.

474
00:47:33,580 --> 00:47:36,940
And so the path of the future is in general, like the past.

475
00:47:37,360 --> 00:47:40,360
There can be spikes, as we see here, but most of the time it's a pretty good rule.

476
00:47:41,480 --> 00:47:47,440
And I stand on that assumption because I can just tell you, you survey what people do.

477
00:47:47,740 --> 00:47:49,480
They do that when I was a prediction company,

478
00:47:49,780 --> 00:47:58,210
the firm where Norman Packard and I formed after we didn't get rich playing roulette, we started an investment management firm.

479
00:47:58,780 --> 00:48:01,960
They don't throw you out for winning. And we did pretty well.

480
00:48:02,140 --> 00:48:07,760
And we did this, too, like everybody else. Basel two risk management.

481
00:48:07,770 --> 00:48:11,910
This is a standard thing. Actually this came in while we were a prediction company in the late nineties.

482
00:48:12,240 --> 00:48:15,400
You something called value at risk where you set the leverage target.

483
00:48:15,420 --> 00:48:20,790
I'll come back and explain what that is. We use a standard assumption about price formation supply equals demand,

484
00:48:21,030 --> 00:48:26,430
so we find prices that way and we assume this fundamentalist buys undervalued assets and vice versa.

485
00:48:26,730 --> 00:48:30,930
So the diagram of the model looks like this. We have a fund, we have a bank.

486
00:48:31,230 --> 00:48:36,510
The fund detects a mispricing, we just assume a fixed valuation.

487
00:48:37,230 --> 00:48:41,820
It updates the portfolio. Wait, there's a little shock that keeps knocking things out of equilibrium.

488
00:48:41,820 --> 00:48:47,130
So we're using a little bit. The same trick is that dynamic, stochastic general equilibrium model.

489
00:48:47,310 --> 00:48:51,960
But with the important thing, we don't enforce strategic equilibrium in this model.

490
00:48:52,920 --> 00:49:00,900
And. The bank similarly updates its risk estimate in force as a leverage constraint, and we form prices.

491
00:49:01,480 --> 00:49:07,320
And now let me mention a key fact, which is that for a passive investor with leverage greater than one like the bank.

492
00:49:07,740 --> 00:49:11,400
If prices drop, leverage goes up and vice versa.

493
00:49:11,910 --> 00:49:18,540
Now, so why is that? If you look at the formula for leverage with prices drop, it affects the risky assets.

494
00:49:18,960 --> 00:49:22,920
It affects the equity, which is the assets minus the liabilities.

495
00:49:23,190 --> 00:49:29,830
It affects these assets by the same amount. But because you're subtracting off liabilities in the denominator, this effect is bigger.

496
00:49:30,090 --> 00:49:36,570
And so you see an amplification. So so when when prices drop, leverage goes up.

497
00:49:36,960 --> 00:49:45,330
Now, if you're under a if you have to control your leverage, this becomes destabilising because now your leverage is higher than it should be.

498
00:49:45,840 --> 00:49:49,770
So you're above your target. So you have to sell assets to match your target.

499
00:49:50,100 --> 00:49:53,700
So notice your risk control is now destabilising the market.

500
00:49:54,150 --> 00:50:01,110
So although it's a sensible thing for you to do as an individual for the world as a whole, if if everybody is doing this,

501
00:50:01,110 --> 00:50:09,390
their price impact is destabilising thing because they're buying into rising markets and selling in into falling markets.

502
00:50:09,690 --> 00:50:16,980
And in fact, when we originally tried to make this model, we tried to make it with just banks and it just blew up because it's just destabilising.

503
00:50:17,310 --> 00:50:21,540
We had to put the fundamentalists in there even to stabilise things.

504
00:50:22,350 --> 00:50:28,230
Now just to explain value at risk, that is a.k.a. Basel two,

505
00:50:28,980 --> 00:50:39,030
value at risk is an and a little rule that mathematically looks like this whereby when volatility increases, your leverage decreases.

506
00:50:39,360 --> 00:50:48,630
And according to Basel two, it would follow this blue line where this B parameter here is minus a half and there's some limit, but up to the limit.

507
00:50:49,350 --> 00:50:55,980
If you set the limit to zero, you can see it's just your leverage is just going as one over the standard deviation of volatility.

508
00:50:56,640 --> 00:51:03,360
Very sensible thing to do from a macroprudential point of view, because if the world's riskier, you should take less leverage and take less risk.

509
00:51:04,710 --> 00:51:08,520
Supported and endorsed by the establishment.

510
00:51:08,760 --> 00:51:11,940
By the way, this dash green line would correspond to constant leverage.

511
00:51:12,450 --> 00:51:21,599
Now, what happens when we run this model? Well, in one regime, when we make the banking sector small, we see random oscillations.

512
00:51:21,600 --> 00:51:27,210
And I should mention this model is now a dynamical system. It's got it's about six dimensional.

513
00:51:28,590 --> 00:51:30,810
And it's it's stochastic in general.

514
00:51:31,200 --> 00:51:36,569
So here we're running it and you see that things are kind of randomly varying around some price, but not much is happening.

515
00:51:36,570 --> 00:51:42,240
The leverage is fluctuating a bit, but not much is happening. Now, suppose we make the banking sector bigger.

516
00:51:42,240 --> 00:51:49,170
And what does that mean? Well, when the banking sector is small, it means we don't give them much money or we don't let them take very much leverage.

517
00:51:49,530 --> 00:51:55,660
But if we let them do either one of those things more strongly, then we see something happening like this.

518
00:51:55,680 --> 00:52:03,660
This is that oscillation I already mentioned, where prices are slowly building and then crashing and slowly building and then crashing.

519
00:52:03,990 --> 00:52:12,420
And leverage is similarly slowly building and crashing. Now, interestingly, we can turn the noise completely off if we turn the noise off.

520
00:52:12,750 --> 00:52:17,820
Now we go to a fixed point in the small banking sector version, things just get flat.

521
00:52:18,330 --> 00:52:21,900
But for the large banking sector, we see an oscillation.

522
00:52:22,440 --> 00:52:28,409
And in fact, this oscillation we can show is a chaotic oscillation where things go around some loop

523
00:52:28,410 --> 00:52:34,170
in a space that involve price and who owns the asset and what the perceived risk is.

524
00:52:34,380 --> 00:52:37,710
And you can see there's some kind of complicated behaviour over in this part.

525
00:52:39,930 --> 00:52:47,070
But if you slice this thing, you'll see the characteristic fractal structure that you expect from a low dimensional, strange attractor.

526
00:52:47,460 --> 00:52:53,280
And by the way, there's a hyperbolic fixed point for those of you that have taken dynamical systems and know what that means.

527
00:52:53,550 --> 00:53:01,170
That's driven by this place where the leverage of is going to infinity, where things get really unstable in this model.

528
00:53:02,700 --> 00:53:06,720
Now let me just discuss a little bit what happens as you very parameters.

529
00:53:07,050 --> 00:53:12,060
There's two key parameters. One is alpha, which is controlling the risk of the individual agents.

530
00:53:12,330 --> 00:53:19,740
And B, which is controlling the risk policy that we let them follow because we're interested in, you know, what what are the alternatives?

531
00:53:20,340 --> 00:53:31,570
And the the face base looks like this. So Basel two corresponds to -0.5 for this parameter B, that makes the risk control more or less pro-cyclical.

532
00:53:31,590 --> 00:53:40,200
It's called and as long and this this is corresponds to varying alpha but I plot leverage instead because that really simplifies the plot.

533
00:53:41,070 --> 00:53:45,990
And what we see is there's a region here where things are nice and stable. This is where the fixed point equilibrium works.

534
00:53:46,590 --> 00:53:53,610
There's a region up here where things oscillate and the boundary between those involves a hop bifurcation.

535
00:53:53,610 --> 00:53:57,630
Things happen very suddenly. It's a flip bifurcation.

536
00:53:57,990 --> 00:54:03,630
Things happen very suddenly. You suddenly go from a fixed point to a finite amplitude oscillation.

537
00:54:03,960 --> 00:54:10,200
And as a regulator, you really better understand this because obviously the consequences could be serious.

538
00:54:10,500 --> 00:54:14,010
Now, does this explain the great moderation in the crisis?

539
00:54:14,340 --> 00:54:17,700
Well, I wouldn't want to argue that it is the explanation.

540
00:54:18,330 --> 00:54:21,330
But on the other hand, it could have caused it.

541
00:54:21,720 --> 00:54:31,020
And I forgot to say something essential back here, which is that when you look at this oscillation here, the time axis is not arbitrary.

542
00:54:31,050 --> 00:54:39,930
We've calibrated this model. These are measured in years. So these these run ups automatically come out with a period of around 15 years.

543
00:54:40,440 --> 00:54:46,350
And the levels of leverage that trigger this to happen are quite plausible for what happens in the market.

544
00:54:46,680 --> 00:54:50,100
So it is we get a reasonably good match with what's actually going on.

545
00:54:50,280 --> 00:54:56,670
And the model very clearly says the leverage was too high leading up in the period leading up to the crisis.

546
00:54:57,960 --> 00:55:06,360
So it could have caused the great moderation in the crisis. And and therefore, the housing bubble may have just been the spark that made this happen.

547
00:55:07,050 --> 00:55:10,140
Now, how can we do things more realistically?

548
00:55:10,440 --> 00:55:14,400
Well, we've been working on making what are called agent based models.

549
00:55:14,760 --> 00:55:18,899
What I just showed you was a very simple example of an agent based model in general.

550
00:55:18,900 --> 00:55:21,960
If you want to make things more realistic, you've got to do better than that.

551
00:55:22,140 --> 00:55:28,020
You're not going to get an accurate model of the economy with a with a model that just has two agents and one asset.

552
00:55:29,160 --> 00:55:30,150
We're not that lucky.

553
00:55:32,610 --> 00:55:40,950
I argue that to surmount the fundamental problem of macroeconomics, which is that if you think about the business cycle I showed you,

554
00:55:41,220 --> 00:55:45,930
there's only about 15 well observed business cycles in the history of civilisation.

555
00:55:47,430 --> 00:55:52,500
The ones in the distant past are probably not completely relevant to what's going on now.

556
00:55:52,920 --> 00:55:56,790
So we have very limited data to use to calibrate models.

557
00:55:57,000 --> 00:56:00,329
We have to, I would argue, make good fundamental models.

558
00:56:00,330 --> 00:56:03,930
We can't just use statistical models because the course of dimensionality blocks us.

559
00:56:04,470 --> 00:56:10,050
And and and we've got to use micro data to validate them.

560
00:56:10,410 --> 00:56:16,020
Now I'm out of time, so I'm not going to say much about this. But just to say, if you look at something like climate models,

561
00:56:16,350 --> 00:56:23,160
it's been key that they're not just looking at time series like global mean temperature and global carbon dioxide.

562
00:56:23,460 --> 00:56:27,840
They're looking at details about the weather patterns they predict.

563
00:56:27,840 --> 00:56:32,730
And they can show they can do things like join Tierra del Fuego together with Antarctica.

564
00:56:32,880 --> 00:56:38,760
There are simulations. The world suddenly gets vastly warmer than it is now because they block the Antarctic current.

565
00:56:38,970 --> 00:56:43,800
And you can look back in the in the record and see that that's in fact how it was done.

566
00:56:44,460 --> 00:56:52,170
Now, my vision of the economics of the future would be a simulation of the economies of the major countries of the world.

567
00:56:52,620 --> 00:56:56,130
We get the micro big data inputs directly off the Internet.

568
00:56:56,580 --> 00:57:05,910
We might even couple this to other social models. We'd have teams focusing on each components on households, firms, banks,

569
00:57:06,210 --> 00:57:12,810
etc. and they would be monitoring, using the Internet whether their models are really being followed.

570
00:57:13,380 --> 00:57:16,830
I'm going to skip over this because I want to wrap up pretty soon.

571
00:57:18,660 --> 00:57:21,450
Now, how well do I think we can do that?

572
00:57:21,840 --> 00:57:28,900
Well, we're not going to do that perfectly, because there are things like market efficiency, market efficiencies, a self-organising process.

573
00:57:28,920 --> 00:57:32,820
I'll say a bit in a moment, it limits the predictability of tradable assets.

574
00:57:33,120 --> 00:57:37,889
I've already alluded to the fact that I think in the real world we see something

575
00:57:37,890 --> 00:57:42,960
that's very close to turbulence in financial assets and in the economy.

576
00:57:43,140 --> 00:57:49,650
And in fact, there's a beautiful paper in nature from about 26 just pointing out that as you vary this,

577
00:57:50,040 --> 00:57:58,140
if you do a match up of the statistics of fluid turbulence to those of the financial system, the match up is quite accurate.

578
00:58:03,150 --> 00:58:07,710
I don't think we're going to be able to make detailed forecasts about the future of the economy.

579
00:58:08,010 --> 00:58:10,110
But I do think there's a lot of stuff we can predict,

580
00:58:10,560 --> 00:58:18,630
and that is there's a lot of information we can compress by making better models that we can use to formulate policies.

581
00:58:20,190 --> 00:58:23,730
It's back to this question of social science versus physics.

582
00:58:24,030 --> 00:58:28,470
You know, unlike atoms, people can't think. That does put.

583
00:58:29,040 --> 00:58:32,130
So people can think, sorry. Sometimes I think they can't.

584
00:58:32,550 --> 00:58:38,290
But yes, people can think. What scientific limits does this impose?

585
00:58:38,310 --> 00:58:43,320
Well, it does impose some limits. It's you know, economics is a [INAUDIBLE] of a lot harder than physics.

586
00:58:43,590 --> 00:58:49,079
I've done both, so I can vouch for that. But that doesn't mean everything's impossible.

587
00:58:49,080 --> 00:58:54,360
That doesn't mean you have to just throw up your hands. And I would point to the consistency of psychological experiments.

588
00:58:54,690 --> 00:58:56,579
There's a lot of psychological experiments.

589
00:58:56,580 --> 00:59:03,900
If you put people in a certain situation, you can make a pretty good prediction that 70% of them will do one thing and 30% will do another thing,

590
00:59:04,140 --> 00:59:11,520
and it holds up fairly consistently through time and so on with some variations, cultural variation and so on.

591
00:59:12,150 --> 00:59:17,580
I think big data empowers us to do incredible things that we couldn't do before.

592
00:59:18,060 --> 00:59:22,470
I already mentioned the example of traffic forecasting, which is a sort of trivial example.

593
00:59:24,360 --> 00:59:31,860
You know, it's a bit frustrating if you look at how we go about running the models that are actually run to understand economies,

594
00:59:32,250 --> 00:59:41,700
they're run through a methodology based on national accounting that was developed in the 1950s and has evolved only little since then.

595
00:59:42,090 --> 00:59:46,490
It involves surveys. It involves boiling everything down into aggregate data.

596
00:59:46,500 --> 00:59:53,580
It involves the numbers coming out often a year after they were actually the events were actually happening.

597
00:59:53,820 --> 01:00:01,020
I just think we can do infinitely better than that. Now, I want to say a bit about market efficiency and then conclude, you know,

598
01:00:02,190 --> 01:00:06,120
the hypothesis of market efficiency says that markets are inherently unpredictable.

599
01:00:06,480 --> 01:00:10,440
Now, when I heard about that, my immediate reaction is, okay, I've got to go try and predict markets.

600
01:00:10,890 --> 01:00:19,530
And, you know, and I think we and many others have since then showed that actually markets are somewhat predictable,

601
01:00:19,530 --> 01:00:24,660
predictable enough to make some somebody but even, you know, they're good firms like ours.

602
01:00:25,050 --> 01:00:29,250
We were right, maybe on 55% of our bets and wrong on 45% of them.

603
01:00:29,280 --> 01:00:32,850
So it's still pretty noisy. So I don't want to say that this is easy.

604
01:00:33,870 --> 01:00:42,810
The world is actually close to being market efficient. The question is, I argue, how much difference does it make?

605
01:00:43,680 --> 01:00:48,089
How much difference do these inefficiencies make? Well, they make an important difference.

606
01:00:48,090 --> 01:00:51,240
They put a lot of money in my pocket. Okay.

607
01:00:51,600 --> 01:00:55,770
But but in general, for society, is this something we ought to care about?

608
01:00:56,190 --> 01:00:59,400
And I argue that we should, because there's a catch.

609
01:00:59,790 --> 01:01:05,279
You can see automatically markets can't be perfectly efficient because market efficiency requires arbitrage,

610
01:01:05,280 --> 01:01:10,470
cause you've got people out there to exploit them, but there's no inefficiencies to exploit the law, go home.

611
01:01:10,740 --> 01:01:15,450
And so the market will be inefficient again. So there must always be some residual inefficiencies.

612
01:01:15,960 --> 01:01:24,210
And so the way I would put this is markets are efficient to first order, but necessarily inefficient at second order.

613
01:01:24,600 --> 01:01:31,710
And I think this has important consequences. I wrote a paper back in 2002 with a market ecology hypothesis,

614
01:01:32,850 --> 01:01:38,130
and the hypothesis is that we need to think about what the strategies that agents are following,

615
01:01:38,370 --> 01:01:44,370
what are these inefficiencies that they're feeding off of? How do these different kinds of agents interact with each other?

616
01:01:44,790 --> 01:01:56,189
And and and what happens? And I did a lot of stuff in that paper where I basically mapped the financial system into the theoretical framework

617
01:01:56,190 --> 01:02:03,690
that's used in ecology and showed that you got all the same things that ecologists have worried about in,

618
01:02:04,200 --> 01:02:05,760
in financial systems.

619
01:02:06,120 --> 01:02:17,760
And I'm working with illicit coinage in health and others to try and use these ideas to actually get better models for understanding financial risk.

620
01:02:18,030 --> 01:02:23,249
Because one of the key hypotheses here, the last one is that imbalances in the market,

621
01:02:23,250 --> 01:02:28,230
food web can drive crises when market ecology is get out of balance.

622
01:02:28,530 --> 01:02:29,940
Strange things can happen.

623
01:02:30,120 --> 01:02:39,930
And I would argue this is what happened in the crisis of 2008, when subprime mortgages and strange new derivatives got way out of hand.

624
01:02:40,650 --> 01:02:45,840
Now, there's a lot of challenges to doing this kind of thing, to making realistic models.

625
01:02:47,940 --> 01:02:54,090
There's a group of us here who are working on this, most of whom are in the Mathematical Institute.

626
01:02:55,380 --> 01:02:59,820
We're trying to create the key building blocks. We're trying to take advantage of the collective,

627
01:03:00,030 --> 01:03:06,360
collective wisdom that's been accumulated by the small cadre of people that are working on out of equilibrium

628
01:03:06,630 --> 01:03:14,220
agent based models to understand how to do simulations in economics make them really simple and easy to use.

629
01:03:14,430 --> 01:03:17,940
We have an open source project. Any of you can track what we're doing.

630
01:03:18,210 --> 01:03:21,210
Join it. All the codes available as it's being created.

631
01:03:21,480 --> 01:03:26,730
If you have programming skills, please, and your interest in this kind of thing, please contact us.

632
01:03:27,160 --> 01:03:33,399
And we'll sign you up. We're trying to create a library of standard models that anybody who wants to join can build

633
01:03:33,400 --> 01:03:38,860
on to deal with the problem that these kind of models actually are time intensive to create.

634
01:03:39,160 --> 01:03:47,139
And if you have to create everything from scratch, it's hard to make to accumulate progress because a typical graduate student will spend several

635
01:03:47,140 --> 01:03:52,060
years just building writing software to get to the point where they can do something useful.

636
01:03:52,270 --> 01:03:57,760
And we're focusing on three use cases at the moment housing markets, insurance industry and bank stress testing.

637
01:03:58,300 --> 01:04:01,300
Now, to conclude, I mean, I apologise.

638
01:04:01,300 --> 01:04:06,790
I haven't said anything in this lecture about long range evolution, about growth,

639
01:04:07,060 --> 01:04:13,420
innovation, the things that actually cause long term structural change in the economy.

640
01:04:13,870 --> 01:04:19,240
There just isn't enough time to talk about it here. Let me just say that's another component of the work I do.

641
01:04:20,740 --> 01:04:29,410
We've shown together with others like Francois Lafond, that actually a lot of these trends are surprisingly forecast able.

642
01:04:29,710 --> 01:04:33,820
There are laws like Moore's Law and Rights Law that make that true.

643
01:04:34,060 --> 01:04:39,070
So I think there is and this is another unexplored area in economics.

644
01:04:39,670 --> 01:04:47,980
Another key application that we desperately need are better economic models to steer us toward a sustainable economy.

645
01:04:49,090 --> 01:04:54,460
Current rates of growth are simply unsustainable. How are we ever going to do this without a huge crisis?

646
01:04:54,760 --> 01:05:03,070
I would argue our existing models are hopelessly flawed. And I think actually this is a mathematics lecture.

647
01:05:03,460 --> 01:05:10,150
I think what we're doing, I think the ecology of the way mathematics done is done, is evolving.

648
01:05:11,140 --> 01:05:14,530
I will stand who was one of the pioneers of that?

649
01:05:15,070 --> 01:05:19,450
As a graduate student, we were pioneers. We started doing what we called experimental economics.

650
01:05:19,570 --> 01:05:23,410
Sorry, experimental mathematics. We put equations on computers.

651
01:05:23,650 --> 01:05:32,650
We played around with them. We tried to map out the phenomenology to serve as a kind of scout for more focussed theory.

652
01:05:32,980 --> 01:05:38,380
And I think that's increasingly the way things are being done and will be done in the future.

653
01:05:38,710 --> 01:05:40,900
I think that will happen in economics as well.

654
01:05:41,170 --> 01:05:49,870
And simulation needs to play a key role in between data on one side and sharper small models on the other side.

655
01:05:50,080 --> 01:05:56,440
We need to be moving more towards an economic big science because while I'm a strong supporter of small science,

656
01:05:56,890 --> 01:06:04,660
you can't do everything with just small science. Imagine if we were trying to deal with global warming without large scale global circulation models.

657
01:06:04,870 --> 01:06:11,920
We just wouldn't be able to do it. Doing this is going to require a culture shift and it's a big job.

658
01:06:12,700 --> 01:06:16,269
And I appreciate you being patient and listening to me.

659
01:06:16,270 --> 01:06:18,480
I'm 10 minutes over time. Thank you.