1
00:00:08,680 --> 00:00:12,799
Speaker 1: Hello, and welcome to another episode of the Odd Lots Podcast.

2
00:00:12,840 --> 00:00:16,760
Speaker 1: I'm Joe Wisenthal and I'm Tracy Alloway. Tracy, I really

3
00:00:16,800 --> 00:00:21,000
Speaker 1: liked last week's episode with Andrew Lowe talking about quant

4
00:00:21,120 --> 00:00:25,320
Speaker 1: stuff and his sort of evolution of the efficient market

5
00:00:25,400 --> 00:00:28,920
Speaker 1: hypothesis and where that might go. Yeah, I did too.

6
00:00:29,000 --> 00:00:32,839
Speaker 1: I really like the ecosystem analogy, the idea that you

7
00:00:32,840 --> 00:00:35,640
Speaker 1: have all these different players with different motivations and they're

8
00:00:35,680 --> 00:00:40,200
Speaker 1: constantly evolving and adapting to the market. That point about

9
00:00:40,200 --> 00:00:43,000
Speaker 1: it adapting, and of course that's the name of his book,

10
00:00:43,080 --> 00:00:47,360
Speaker 1: Adaptive Market is the name of his hypothesis. Is really

11
00:00:47,520 --> 00:00:50,600
Speaker 1: key because one of the main points that he made

12
00:00:50,920 --> 00:00:54,000
Speaker 1: that I loved was this idea of hedge funds as

13
00:00:54,080 --> 00:00:57,920
Speaker 1: sort of the R and D laboratory for all of

14
00:00:57,960 --> 00:01:02,160
Speaker 1: the financial industry. Right, the funds are where innovative new

15
00:01:02,240 --> 00:01:05,720
Speaker 1: techniques get to be sort of hashed out without doing

16
00:01:06,040 --> 00:01:09,280
Speaker 1: usually too much damage, I guess, to the wider ecosyst

17
00:01:09,280 --> 00:01:13,680
Speaker 1: hopefully right, not always. There's certainly examples of hedge funds

18
00:01:13,680 --> 00:01:17,919
Speaker 1: actually having done major damage from time to time. But ideally,

19
00:01:18,160 --> 00:01:22,160
Speaker 1: you know what the evolution seems to be that some

20
00:01:22,200 --> 00:01:24,759
Speaker 1: new idea sort of starts in the hedge fund world

21
00:01:25,040 --> 00:01:27,600
Speaker 1: and eventually makes its way to the broader world. I

22
00:01:27,600 --> 00:01:29,759
Speaker 1: think the most obvious example of that that we could

23
00:01:29,760 --> 00:01:32,160
Speaker 1: say these days is a lot of this sort of

24
00:01:32,520 --> 00:01:36,360
Speaker 1: popular smart beta strategies e t f s that are

25
00:01:36,360 --> 00:01:41,119
Speaker 1: built on things like momentum or value or other factors,

26
00:01:41,440 --> 00:01:44,640
Speaker 1: sort of quantitative ideas that for many years were only

27
00:01:44,680 --> 00:01:48,920
Speaker 1: available to uh, you know, researchers at hedge funds. Right. So,

28
00:01:48,960 --> 00:01:53,160
Speaker 1: these sorts of quantitative investment or trading methods were usually

29
00:01:53,240 --> 00:01:56,040
Speaker 1: the purview of sophisticated hedge funds who had the time

30
00:01:56,080 --> 00:01:58,880
Speaker 1: and resources to develop them. And then you had a

31
00:01:58,920 --> 00:02:01,120
Speaker 1: bunch of e t f s who kind of caught

32
00:02:01,160 --> 00:02:03,800
Speaker 1: on and managed to replicate them. And now we can

33
00:02:03,840 --> 00:02:08,679
Speaker 1: all trade lighthedge funds for zero percent fees, right, exactly right,

34
00:02:09,040 --> 00:02:12,280
Speaker 1: And of course, once everyone can do it for very

35
00:02:12,320 --> 00:02:15,560
Speaker 1: few fees, I think it's safe to say those strategies

36
00:02:15,560 --> 00:02:19,840
Speaker 1: aren't going to produce the same returns, and hence the

37
00:02:19,919 --> 00:02:23,560
Speaker 1: market is forced to adapt again. Right. Presumably the hedge

38
00:02:23,560 --> 00:02:26,400
Speaker 1: funds are always trying to stay one step ahead as well, right,

39
00:02:27,120 --> 00:02:30,360
Speaker 1: exactly so, which raises the idea of like what will

40
00:02:30,400 --> 00:02:34,960
Speaker 1: be the next thing. If anyone can sort of invest

41
00:02:35,000 --> 00:02:38,720
Speaker 1: in a crude momentum strategy for virtually no fees, then

42
00:02:38,760 --> 00:02:41,440
Speaker 1: that requires the people on the cutting edge, the people

43
00:02:41,520 --> 00:02:44,160
Speaker 1: doing the R and D of this industry, to uh,

44
00:02:44,240 --> 00:02:46,000
Speaker 1: you know, figure out what the next big thing it's

45
00:02:46,000 --> 00:02:48,359
Speaker 1: gonna be. Do you know what the next big thing

46
00:02:48,400 --> 00:02:50,239
Speaker 1: is going to be? Joe? Can you share it with

47
00:02:50,680 --> 00:02:55,360
Speaker 1: your fellow partner at Odd Thoughts LLC? Sadly and unfortunately

48
00:02:56,000 --> 00:02:59,760
Speaker 1: to all of the odd Lots fan, I myself do

49
00:02:59,800 --> 00:03:03,480
Speaker 1: not know what the next big thing in quantitative strategy

50
00:03:03,639 --> 00:03:06,480
Speaker 1: or sort of advanced investing is going to be. But

51
00:03:06,560 --> 00:03:09,960
Speaker 1: I'm hoping that our guests on today's episode might be

52
00:03:10,040 --> 00:03:13,960
Speaker 1: able to shed some light. Who are there? Okay, so

53
00:03:14,040 --> 00:03:17,040
Speaker 1: today we're going to be talking to John Elberg. He

54
00:03:17,160 --> 00:03:21,240
Speaker 1: is the founder of Euclidean Technologies, a quant firm, as

55
00:03:21,240 --> 00:03:24,919
Speaker 1: well as Zach Lipton, a professor at Carnegie Mellon University

56
00:03:24,960 --> 00:03:28,519
Speaker 1: in the Business School and expert on machine learning. They

57
00:03:28,639 --> 00:03:33,480
Speaker 1: recently published a paper titled Improving factor based Quantitative Investing

58
00:03:34,000 --> 00:03:39,520
Speaker 1: by Forecasting company fundamentals. So what I think that means

59
00:03:39,560 --> 00:03:41,000
Speaker 1: and we'll talk to them, is, you know, we talk

60
00:03:41,080 --> 00:03:44,800
Speaker 1: all this stuff about price and computers and algorithms figuring

61
00:03:44,840 --> 00:03:47,880
Speaker 1: out what signal we can get from price, But maybe

62
00:03:48,480 --> 00:03:51,200
Speaker 1: the next generation can actually tell us something about the

63
00:03:51,240 --> 00:03:54,640
Speaker 1: fundamental workings of the company itself, and maybe this could

64
00:03:54,680 --> 00:03:57,880
Speaker 1: be sort of the next wave of where quant investing goes.

65
00:03:58,440 --> 00:04:02,080
Speaker 1: And this sounds absolutely senating, Joe, let's bring them on.

66
00:04:12,280 --> 00:04:15,240
Speaker 1: John and Zach, thank you very much for joining us.

67
00:04:15,920 --> 00:04:19,800
Speaker 1: Was that a reasonable characterization of sort of where your

68
00:04:20,160 --> 00:04:24,080
Speaker 1: paper and where your research is taking things? Yeah, I

69
00:04:24,120 --> 00:04:27,280
Speaker 1: think it is so. So. First of all, machine learning

70
00:04:27,320 --> 00:04:30,160
Speaker 1: has been kind of on a rocket ship of innovation

71
00:04:30,240 --> 00:04:33,159
Speaker 1: for the last ten years or so, and with the

72
00:04:33,240 --> 00:04:36,800
Speaker 1: advent of deep learning, you know, computers and machine learning

73
00:04:36,800 --> 00:04:39,120
Speaker 1: have been able to do things that you know, historically

74
00:04:39,120 --> 00:04:44,400
Speaker 1: have been very challenging, like image captioning and language translation.

75
00:04:45,080 --> 00:04:47,840
Speaker 1: So we Zach and I, you know, a couple of

76
00:04:47,880 --> 00:04:52,080
Speaker 1: years back, thought of the idea of collaborating to apply

77
00:04:52,560 --> 00:04:56,480
Speaker 1: deep learning to the problem of long term investing. So

78
00:04:56,520 --> 00:04:59,240
Speaker 1: how did you actually go about doing that and what

79
00:04:59,400 --> 00:05:03,680
Speaker 1: exactly do you mean by deep learning? That's exactly what

80
00:05:03,720 --> 00:05:07,159
Speaker 1: I wanted to know. To deep learning sort of the

81
00:05:07,200 --> 00:05:10,560
Speaker 1: rebranding of neural networks research to say I say I

82
00:05:10,600 --> 00:05:13,279
Speaker 1: had some data about a company, right like I had

83
00:05:13,680 --> 00:05:16,000
Speaker 1: machine learning. We call a vector of features. But what

84
00:05:16,240 --> 00:05:18,080
Speaker 1: we mean it's just like a list of attributes, each

85
00:05:18,120 --> 00:05:20,880
Speaker 1: of which is somehow like be made into a numerical quantity,

86
00:05:21,000 --> 00:05:24,240
Speaker 1: whether it's like their income, they're the number of assets whatever.

87
00:05:24,640 --> 00:05:28,039
Speaker 1: One way of deciding how to predict what the say,

88
00:05:28,120 --> 00:05:30,360
Speaker 1: what the price will be or something, as we say, well,

89
00:05:30,600 --> 00:05:33,400
Speaker 1: we're going to have this long vector of features, and

90
00:05:33,440 --> 00:05:36,200
Speaker 1: then we're going for every single company, uh, you know,

91
00:05:36,279 --> 00:05:38,640
Speaker 1: at every single time while this vector of features corresponding

92
00:05:38,640 --> 00:05:40,360
Speaker 1: to the state of the company at some period of time,

93
00:05:40,880 --> 00:05:44,640
Speaker 1: and then we'll have some target that we want to predict.

94
00:05:44,680 --> 00:05:47,600
Speaker 1: This could be a binary quantity like will the stock

95
00:05:47,640 --> 00:05:50,280
Speaker 1: go up or down in the next you know, time

96
00:05:50,400 --> 00:05:53,320
Speaker 1: unit of your choice, whether it's the next day or

97
00:05:53,320 --> 00:05:55,400
Speaker 1: in the next month or in the next year. Or

98
00:05:55,440 --> 00:05:58,599
Speaker 1: you could try to directly predict say the relative price,

99
00:05:58,680 --> 00:06:02,520
Speaker 1: so like you know, the percent improvement or decrease based

100
00:06:02,560 --> 00:06:05,440
Speaker 1: on sort of the available features. So one of the

101
00:06:05,480 --> 00:06:07,680
Speaker 1: simplest ways you can make a model is you say, hey,

102
00:06:07,720 --> 00:06:09,120
Speaker 1: I've got a bunch of features. I'm gonna do is

103
00:06:09,120 --> 00:06:11,479
Speaker 1: I'm gonna take a weighted some of these features the

104
00:06:11,520 --> 00:06:14,000
Speaker 1: way like you'd calculate a score to see, like what's

105
00:06:14,000 --> 00:06:16,240
Speaker 1: your risk of a heart disease. Maybe you take you know, well,

106
00:06:16,240 --> 00:06:20,120
Speaker 1: four times your cholesterol plus two times your age minus

107
00:06:20,120 --> 00:06:23,440
Speaker 1: one times you know, your amount of good cholesterol is

108
00:06:23,480 --> 00:06:25,800
Speaker 1: something like this if you come up with some formula

109
00:06:25,920 --> 00:06:28,800
Speaker 1: that's expressed simply as a weighted sum, so that would

110
00:06:28,800 --> 00:06:31,719
Speaker 1: be a linear model. Where deep learning make things different

111
00:06:31,839 --> 00:06:35,479
Speaker 1: is that you have many different layers of computation that

112
00:06:35,800 --> 00:06:38,720
Speaker 1: you basically are learning very complex patterns that maybe couldn't

113
00:06:38,760 --> 00:06:41,120
Speaker 1: be expressed as a as a weighted sum. So maybe

114
00:06:41,120 --> 00:06:44,960
Speaker 1: you're uncovering interactions between all of your features. Um. So,

115
00:06:45,000 --> 00:06:48,159
Speaker 1: for example, if you want to learn to recognize a

116
00:06:48,240 --> 00:06:50,839
Speaker 1: dog versus a cat in an image, there's no weighted

117
00:06:50,880 --> 00:06:53,000
Speaker 1: sum of pixel values it's actually going to tell you

118
00:06:53,000 --> 00:06:55,359
Speaker 1: this because it's just the patterns too complicated. So in

119
00:06:55,360 --> 00:06:58,239
Speaker 1: that case you need some some more like heavy duty machinery.

120
00:06:58,360 --> 00:07:00,800
Speaker 1: So what you do in deep learning essentially is that

121
00:07:00,839 --> 00:07:05,320
Speaker 1: you learn multiple successive transformations of your data such that

122
00:07:05,480 --> 00:07:09,640
Speaker 1: after applying many such transformations, you know, could be two,

123
00:07:09,760 --> 00:07:12,480
Speaker 1: four or five, ten, whatever, you come out at the

124
00:07:12,600 --> 00:07:15,240
Speaker 1: end of a representation of your data where you actually

125
00:07:15,280 --> 00:07:17,680
Speaker 1: can learn a very simple model on top of that.

126
00:07:17,920 --> 00:07:21,280
Speaker 1: So we sometimes call deep learning representation learning because it's

127
00:07:21,320 --> 00:07:24,200
Speaker 1: what we're doing is we're both learning how to feature

128
00:07:24,200 --> 00:07:26,600
Speaker 1: eye our data essentially, how to transform it and how

129
00:07:26,600 --> 00:07:29,360
Speaker 1: to classify it at the same time. So one of

130
00:07:29,400 --> 00:07:32,720
Speaker 1: the things in sort of traditional quantity, a lot of

131
00:07:32,800 --> 00:07:37,600
Speaker 1: quantitative investing focuses a lot on price and sort of

132
00:07:37,960 --> 00:07:43,240
Speaker 1: listening to your characterization. It seems like price and this

133
00:07:43,320 --> 00:07:46,200
Speaker 1: is relatively speaking, of course, price is a fairly you know,

134
00:07:46,240 --> 00:07:49,240
Speaker 1: it's sort of easy idea to capture. So you can

135
00:07:49,280 --> 00:07:53,720
Speaker 1: come up with some definition of what momentum is and

136
00:07:53,760 --> 00:07:56,640
Speaker 1: then sort of say, okay, these stocks are experiencing momentum

137
00:07:56,760 --> 00:08:00,600
Speaker 1: right now, or these stocks aren't, and then his history

138
00:08:00,720 --> 00:08:02,560
Speaker 1: tell us the stocks are going to do next if

139
00:08:02,560 --> 00:08:07,120
Speaker 1: they sort of meet these characterizations. Your paper really looks

140
00:08:07,200 --> 00:08:10,440
Speaker 1: at what can you do with this technology for sort

141
00:08:10,440 --> 00:08:14,080
Speaker 1: of looking at future fundamental so looking at the sort

142
00:08:14,120 --> 00:08:17,080
Speaker 1: of characteristics of the company and not just trying to

143
00:08:17,560 --> 00:08:21,440
Speaker 1: see where prices going, but where those characteristics are going,

144
00:08:21,480 --> 00:08:25,960
Speaker 1: so explain sort of what your research specifically attempts to uncover.

145
00:08:26,680 --> 00:08:30,160
Speaker 1: So one thing that deep learning allows a researcher to

146
00:08:30,240 --> 00:08:35,280
Speaker 1: do is look at kind of more raw features um.

147
00:08:35,320 --> 00:08:38,520
Speaker 1: Like Zach explained in the image case, you're looking at

148
00:08:38,640 --> 00:08:42,840
Speaker 1: raw pixels. Now, if you think about most quant funds

149
00:08:43,000 --> 00:08:46,240
Speaker 1: and most quant models, they the features that go into

150
00:08:46,280 --> 00:08:49,160
Speaker 1: the model are highly engineered, and they include things like

151
00:08:49,240 --> 00:08:53,800
Speaker 1: price and maybe book value, price divided by book value,

152
00:08:53,840 --> 00:08:57,520
Speaker 1: price divided by earnings, and then maybe some momentum features.

153
00:08:58,000 --> 00:09:01,000
Speaker 1: The interesting thing about deep learning is it allows you

154
00:09:01,080 --> 00:09:05,720
Speaker 1: to potentially let it uncover what the best features are.

155
00:09:05,760 --> 00:09:09,240
Speaker 1: If you over engineer features, you may not find the

156
00:09:09,240 --> 00:09:12,040
Speaker 1: ones that are best to predict what you're interested in predicting.

157
00:09:12,600 --> 00:09:16,360
Speaker 1: So that, you know, allows you to potentially find features

158
00:09:16,360 --> 00:09:20,199
Speaker 1: in the data that you wouldn't find through which traditional

159
00:09:20,240 --> 00:09:24,640
Speaker 1: feature engineering process. Yeah, and you know, to directly address

160
00:09:24,679 --> 00:09:27,080
Speaker 1: your question, your point is that the very most obvious

161
00:09:27,080 --> 00:09:28,920
Speaker 1: thing you could say, now, if I have this, I

162
00:09:28,960 --> 00:09:31,199
Speaker 1: have this learning machine, I have a bunch of features,

163
00:09:31,200 --> 00:09:32,720
Speaker 1: and I have to choose what am I going to predict?

164
00:09:33,080 --> 00:09:35,000
Speaker 1: The very most obvious thing to try to predict is

165
00:09:35,000 --> 00:09:38,040
Speaker 1: the price, because if you can actually do that perfectly,

166
00:09:38,360 --> 00:09:41,240
Speaker 1: then you're done, right. If if you actually know which

167
00:09:41,240 --> 00:09:43,280
Speaker 1: way the price is going to move in the next year,

168
00:09:43,440 --> 00:09:46,679
Speaker 1: then you can make the perfect choice. So the problem

169
00:09:46,760 --> 00:09:49,440
Speaker 1: is that that's that's not so easy because the markets

170
00:09:49,440 --> 00:09:53,120
Speaker 1: are quite capricious, right, Um, So one problem that we

171
00:09:53,200 --> 00:09:55,280
Speaker 1: found is we actually did these models where we were

172
00:09:55,280 --> 00:09:58,280
Speaker 1: trying to predict price directly. But among the other things

173
00:09:58,280 --> 00:10:00,520
Speaker 1: that you have is that one, it's hard to learn

174
00:10:00,640 --> 00:10:03,280
Speaker 1: models that do a good job of this that are

175
00:10:03,360 --> 00:10:06,480
Speaker 1: sort of robust across different time periods. So you might

176
00:10:06,559 --> 00:10:08,600
Speaker 1: have like, hey, I'm going to train on these like

177
00:10:08,720 --> 00:10:11,080
Speaker 1: decades of data and I'm going to try to directly

178
00:10:11,160 --> 00:10:13,800
Speaker 1: predict the price. But then I come into periods of

179
00:10:13,840 --> 00:10:16,760
Speaker 1: time where the markets behaving a little bit differently, and

180
00:10:16,800 --> 00:10:20,080
Speaker 1: we call this nonstationarity. Basically, like you're modeled, there's a

181
00:10:20,080 --> 00:10:23,280
Speaker 1: great job of uncovering the pattern that's present in the

182
00:10:23,440 --> 00:10:25,760
Speaker 1: data that you gave to the model, but that data

183
00:10:25,880 --> 00:10:28,079
Speaker 1: is anchored to some period of time, and the future

184
00:10:28,160 --> 00:10:30,840
Speaker 1: data that comes in, you know, the patterns changed a

185
00:10:30,840 --> 00:10:33,680
Speaker 1: little bit, and so the kind of like function that

186
00:10:33,679 --> 00:10:35,840
Speaker 1: you've learned no longer does a great job. So so

187
00:10:35,880 --> 00:10:38,120
Speaker 1: what we do instead of directly trying to predict the price,

188
00:10:38,480 --> 00:10:40,719
Speaker 1: the idea that we had was to think, well, this

189
00:10:40,880 --> 00:10:43,880
Speaker 1: core idea behind a factor model, generally right, is to

190
00:10:43,960 --> 00:10:45,840
Speaker 1: just say, hey, I'm going to sort all the stocks

191
00:10:46,320 --> 00:10:49,360
Speaker 1: according to some reason idea, Hey, the price of the

192
00:10:49,360 --> 00:10:52,440
Speaker 1: company should be tied to its income, any company, and

193
00:10:52,520 --> 00:10:55,040
Speaker 1: somehow it is justified by like it's the long term

194
00:10:55,120 --> 00:10:58,760
Speaker 1: discounted cash as well. Let's just say a factor strategy

195
00:10:58,800 --> 00:11:00,600
Speaker 1: just something very simple. It says, well, let's just look

196
00:11:00,600 --> 00:11:03,760
Speaker 1: at the current income divided by say the current price

197
00:11:03,880 --> 00:11:06,640
Speaker 1: or current income divided by the current you know, market

198
00:11:06,640 --> 00:11:10,880
Speaker 1: cap or enterprise value, some some notion of income and

199
00:11:10,920 --> 00:11:13,880
Speaker 1: some notion notion of financial performance, and divided by some

200
00:11:13,920 --> 00:11:16,640
Speaker 1: notion of company size and this, and then I'm going

201
00:11:16,679 --> 00:11:18,640
Speaker 1: to sort the stocks according to this. The ones that

202
00:11:18,720 --> 00:11:22,200
Speaker 1: come out highest are like most cheaply priced, so let's

203
00:11:22,200 --> 00:11:25,200
Speaker 1: buy those. So the ideas to say, hey, well, what

204
00:11:25,240 --> 00:11:27,800
Speaker 1: if I told you so. We actually know that this

205
00:11:27,840 --> 00:11:29,720
Speaker 1: does pretty well in back testing whether or not the

206
00:11:29,720 --> 00:11:32,600
Speaker 1: patterns will hold in the future. But you know, many

207
00:11:32,600 --> 00:11:34,199
Speaker 1: people have made a lot of money for many years,

208
00:11:34,360 --> 00:11:36,680
Speaker 1: so there's an idea of if you knew the income,

209
00:11:36,720 --> 00:11:39,520
Speaker 1: this is a good thing, a reasonable thing to try

210
00:11:39,559 --> 00:11:43,200
Speaker 1: to do. Our question that we asked, Unfortunately, um, John,

211
00:11:43,440 --> 00:11:46,840
Speaker 1: because he's actually in finance and I'm not, has this

212
00:11:47,240 --> 00:11:50,360
Speaker 1: really great set of like industry grade tools that unlike

213
00:11:50,360 --> 00:11:52,600
Speaker 1: most academic papers that look at like one stock over

214
00:11:52,640 --> 00:11:54,280
Speaker 1: a short period of time or something, we actually had,

215
00:11:54,360 --> 00:11:57,439
Speaker 1: you know, forty plus years of financial data and can

216
00:11:57,480 --> 00:12:00,920
Speaker 1: actually simulate like an applausible that guess what's going on.

217
00:12:01,320 --> 00:12:04,720
Speaker 1: We said, well, what if you did a factor model,

218
00:12:05,000 --> 00:12:08,120
Speaker 1: but someone gave you a crystal ball. So basically, instead

219
00:12:08,120 --> 00:12:11,240
Speaker 1: of dividing the current income divided by the current enterprise value,

220
00:12:11,600 --> 00:12:15,160
Speaker 1: someone gave you next year's income, and so you sorted

221
00:12:15,160 --> 00:12:18,080
Speaker 1: the stocks according to next year's income divided by the

222
00:12:18,120 --> 00:12:22,080
Speaker 1: current enterprise value something like this. So you're you're able

223
00:12:22,120 --> 00:12:24,560
Speaker 1: to peek into the future. You know how the company

224
00:12:24,600 --> 00:12:28,160
Speaker 1: will be performing next year, and you're saying, is how

225
00:12:28,240 --> 00:12:31,360
Speaker 1: is its next year's performance? Is that based on next

226
00:12:31,440 --> 00:12:34,280
Speaker 1: year's performance? Is its current price? Is it currently priced

227
00:12:34,320 --> 00:12:36,640
Speaker 1: cheaply or not? So it's what we call like a

228
00:12:36,720 --> 00:12:39,959
Speaker 1: clairvoyant factor model. Like you don't actually have such a

229
00:12:39,960 --> 00:12:41,880
Speaker 1: crystal ball, but if you, you know, give us some

230
00:12:41,960 --> 00:12:44,360
Speaker 1: license and you imagine that you did what would have

231
00:12:44,360 --> 00:12:46,320
Speaker 1: happened if you went back in history and you had

232
00:12:46,320 --> 00:12:49,080
Speaker 1: this crystal ball and you traded based on a clairvoyant

233
00:12:49,120 --> 00:12:51,800
Speaker 1: factor model, and it turns out that the clairvoyant factor

234
00:12:51,880 --> 00:12:55,000
Speaker 1: model just crushes it. So it does really, really well

235
00:12:55,040 --> 00:12:58,640
Speaker 1: and and not surprisingly, the more clairvoyant the model is.

236
00:12:58,679 --> 00:13:01,480
Speaker 1: So if it if it knows the performance of the

237
00:13:01,480 --> 00:13:05,440
Speaker 1: company six months out versus now, or twelve months out

238
00:13:05,520 --> 00:13:08,040
Speaker 1: versus six months out, it keeps getting better and better

239
00:13:08,080 --> 00:13:12,280
Speaker 1: and better. So what we decided was, well, maybe trying

240
00:13:12,280 --> 00:13:15,160
Speaker 1: to predict price directly as a bit you know, subject

241
00:13:15,520 --> 00:13:18,720
Speaker 1: to you know, a kind of fickle market, but the

242
00:13:18,760 --> 00:13:23,280
Speaker 1: patterns present in the fundamental reporting data itself is more stable.

243
00:13:23,840 --> 00:13:25,880
Speaker 1: So in our method what we do is instead of

244
00:13:25,920 --> 00:13:28,480
Speaker 1: just trying to predict a return, we try to predict

245
00:13:28,640 --> 00:13:32,680
Speaker 1: actually the fundamental reporting data itself, just so we're given

246
00:13:33,360 --> 00:13:36,920
Speaker 1: um these these features for like a trailing window of

247
00:13:36,920 --> 00:13:42,040
Speaker 1: of time corresponding to the company's like financial reporting, and

248
00:13:42,080 --> 00:13:44,199
Speaker 1: then we're trying to predict what they're going to report

249
00:13:44,240 --> 00:13:46,600
Speaker 1: next year. And then based on what they're going we

250
00:13:46,640 --> 00:13:49,040
Speaker 1: think they're going to report next year, we sort the

251
00:13:49,040 --> 00:13:52,920
Speaker 1: companies according to a value factor. So in essence, you

252
00:13:52,960 --> 00:13:56,760
Speaker 1: can pick out of that future prediction the components of

253
00:13:56,800 --> 00:13:59,960
Speaker 1: the factor model. Let let it whether it's a few

254
00:14:00,000 --> 00:14:04,480
Speaker 1: future predicted earnings, and you can take that out of

255
00:14:04,520 --> 00:14:09,079
Speaker 1: the future predicted fundamentals, divide that by current enterprise value

256
00:14:09,679 --> 00:14:12,160
Speaker 1: and and sort and then you have basically a factor

257
00:14:12,240 --> 00:14:17,719
Speaker 1: model which you are using. Instead of trailing twelve months earnings,

258
00:14:18,120 --> 00:14:21,560
Speaker 1: you're using the future predicted earnings by the deep learning

259
00:14:21,760 --> 00:14:24,680
Speaker 1: the deep neural network. So, as I understand it, the

260
00:14:25,400 --> 00:14:28,960
Speaker 1: deep learning or the neural networks are used primarily to

261
00:14:29,000 --> 00:14:34,960
Speaker 1: forecast the future fundamentals based on historic performance. Is that right,

262
00:14:35,800 --> 00:14:40,680
Speaker 1: historic fundamentals? Yeah, okay, So walk us through how you

263
00:14:40,720 --> 00:14:45,040
Speaker 1: actually develop an application that's able to do that, Like

264
00:14:45,320 --> 00:14:48,400
Speaker 1: what are those neural networks looking at and what sort

265
00:14:48,440 --> 00:14:51,360
Speaker 1: of information are they drawing in other than you know,

266
00:14:51,440 --> 00:14:56,880
Speaker 1: past predictive data to make those forecasts. There's two parts

267
00:14:56,880 --> 00:14:59,120
Speaker 1: of that. One is the data that we use, and

268
00:14:59,120 --> 00:15:02,360
Speaker 1: then two is the technology we use to build the

269
00:15:02,400 --> 00:15:06,720
Speaker 1: deep you know, neural network models. So on the data side,

270
00:15:06,960 --> 00:15:11,680
Speaker 1: what you use is historical fundamentals on all companies you

271
00:15:11,720 --> 00:15:13,880
Speaker 1: know that have ever you know, been listed in the

272
00:15:14,000 --> 00:15:17,720
Speaker 1: US for the past fifty years, and so what a

273
00:15:17,800 --> 00:15:21,880
Speaker 1: historical fundamentals mean? What it means earnings, book value, anything

274
00:15:21,880 --> 00:15:24,080
Speaker 1: you can find on an income statement and balance sheet

275
00:15:24,120 --> 00:15:27,880
Speaker 1: going back in time. In addition to fundamentals, we also

276
00:15:28,600 --> 00:15:32,080
Speaker 1: use as inputs to the to the model, you know,

277
00:15:32,240 --> 00:15:37,080
Speaker 1: momentum over you know, one month, six months, twelve months.

278
00:15:37,120 --> 00:15:38,920
Speaker 1: So then you know, if you think of it as

279
00:15:39,000 --> 00:15:43,040
Speaker 1: like a big you know, spreadsheet table where each row

280
00:15:43,600 --> 00:15:47,760
Speaker 1: is a point in time for a specific company, and

281
00:15:47,800 --> 00:15:52,120
Speaker 1: then you can think of sequences going back through time.

282
00:15:52,400 --> 00:15:57,400
Speaker 1: You know IBM in March of and then all of

283
00:15:57,440 --> 00:16:00,720
Speaker 1: its fundamentals in one row, plus it's moment at them,

284
00:16:00,760 --> 00:16:03,920
Speaker 1: and then that going back five years and time. So

285
00:16:03,920 --> 00:16:07,880
Speaker 1: those sequences, both the fundamentals and the momentum are fed

286
00:16:08,080 --> 00:16:12,680
Speaker 1: into a neural network and uh and and all of

287
00:16:12,720 --> 00:16:15,840
Speaker 1: those sequences for all companies and all time are fed

288
00:16:15,880 --> 00:16:19,840
Speaker 1: into a neural network and are trained to predict what

289
00:16:20,000 --> 00:16:22,800
Speaker 1: the fundamentals will be, you know, one time step out

290
00:16:22,840 --> 00:16:36,440
Speaker 1: in the future. So just to sort of summarize it

291
00:16:36,480 --> 00:16:39,360
Speaker 1: all up, you know, it's like, if you have all

292
00:16:39,400 --> 00:16:43,440
Speaker 1: these strategies, if you have all these funds chasing things

293
00:16:43,480 --> 00:16:49,040
Speaker 1: like earnings, quality, earnings, growth, momentum, all kinds of stuff

294
00:16:49,080 --> 00:16:53,600
Speaker 1: like that, your goal is to anticipate today when those

295
00:16:53,640 --> 00:16:56,160
Speaker 1: funds are going to be buying tomorrow. Is that a

296
00:16:56,160 --> 00:16:58,320
Speaker 1: fair way to characterize it. I think that's a fair

297
00:16:58,440 --> 00:17:01,280
Speaker 1: way to characterize is it. I think what we're really

298
00:17:01,320 --> 00:17:03,720
Speaker 1: just doing is trying to build a better a better

299
00:17:03,760 --> 00:17:06,840
Speaker 1: factor model, A better factor model in the sense that

300
00:17:06,880 --> 00:17:09,560
Speaker 1: you know, as Zach explained, if you had a clairvoyant

301
00:17:09,560 --> 00:17:13,679
Speaker 1: model where you actually knew what future fundamentals were and

302
00:17:13,760 --> 00:17:16,680
Speaker 1: could plug that into a factor model, you do substantially

303
00:17:16,760 --> 00:17:21,399
Speaker 1: better than what you could achieve with a value factor model. Today,

304
00:17:21,840 --> 00:17:25,040
Speaker 1: we're not like directly considering the psychology of the other

305
00:17:25,040 --> 00:17:28,440
Speaker 1: players in the market in this particular approach, right, No, sure,

306
00:17:28,920 --> 00:17:32,920
Speaker 1: but it's essentially saying, like, maybe the way to characterize

307
00:17:32,920 --> 00:17:36,600
Speaker 1: it is, if you want to invest on some fundamental

308
00:17:36,680 --> 00:17:40,680
Speaker 1: factor like earnings quality or earnings growth, bottom line, is

309
00:17:41,320 --> 00:17:44,679
Speaker 1: better to look at future twelve month results rather than

310
00:17:44,720 --> 00:17:48,840
Speaker 1: trailing twelve months. You look at the trailing, but you're

311
00:17:48,960 --> 00:17:51,959
Speaker 1: trying to predict the future. Like, so those two components, right,

312
00:17:52,000 --> 00:17:54,720
Speaker 1: you could say, like one is we have the component

313
00:17:54,800 --> 00:17:58,080
Speaker 1: that is trying to predict the future fundamentals. You know,

314
00:17:58,119 --> 00:18:01,520
Speaker 1: imagine that I came for the future, and I got

315
00:18:01,520 --> 00:18:03,919
Speaker 1: out of my time machine, and I gave you the

316
00:18:03,960 --> 00:18:07,320
Speaker 1: earnings reports from the future. Right, So so the first

317
00:18:07,320 --> 00:18:09,520
Speaker 1: thing you need is how do I get an approximate

318
00:18:09,600 --> 00:18:12,200
Speaker 1: time machine, right, which in our case is a predictive

319
00:18:12,200 --> 00:18:14,440
Speaker 1: model that has a good guess about what the future

320
00:18:14,440 --> 00:18:17,040
Speaker 1: will look like. The second thing is you still need

321
00:18:17,080 --> 00:18:20,080
Speaker 1: a way of executing on the strategy ones I. You

322
00:18:20,119 --> 00:18:23,560
Speaker 1: still need a way to decide which stocks to buy, right, So,

323
00:18:24,119 --> 00:18:27,520
Speaker 1: based on based on this future information, Like, it's possible

324
00:18:27,560 --> 00:18:29,359
Speaker 1: that if I if I come from the future and

325
00:18:29,400 --> 00:18:32,160
Speaker 1: I give you the earnings report, and I tell you

326
00:18:32,200 --> 00:18:35,199
Speaker 1: what the future income will be, what, it's possible that

327
00:18:35,200 --> 00:18:36,600
Speaker 1: the income is going to go up with the stock

328
00:18:36,640 --> 00:18:38,960
Speaker 1: price is going to go down, you know, like say

329
00:18:39,119 --> 00:18:41,720
Speaker 1: it's an Apple and like they made a lot more money,

330
00:18:41,760 --> 00:18:44,600
Speaker 1: but it was also like announced that they had a

331
00:18:44,640 --> 00:18:48,920
Speaker 1: major plant failure in the iPhone fourteen or whatever they're

332
00:18:49,000 --> 00:18:52,040
Speaker 1: up to is going to be delayed. So these two

333
00:18:52,040 --> 00:18:54,640
Speaker 1: components are are a little bit modular, Like we could

334
00:18:54,640 --> 00:18:57,320
Speaker 1: come up with m. John I think is more the

335
00:18:57,359 --> 00:18:59,680
Speaker 1: domain expert, so I'm I'm more the machine learning guy.

336
00:18:59,760 --> 00:19:02,119
Speaker 1: Like I'm sure John could come up with you know,

337
00:19:02,400 --> 00:19:04,920
Speaker 1: a million other ways that you might imagine that someone

338
00:19:04,960 --> 00:19:08,000
Speaker 1: would try to execute on this information. In our case,

339
00:19:08,040 --> 00:19:10,320
Speaker 1: what we're doing is we've adapted a factor model to

340
00:19:10,560 --> 00:19:14,840
Speaker 1: work with this kind of future guess. So one other example,

341
00:19:15,400 --> 00:19:17,439
Speaker 1: so so again, in our case, what we're doing is

342
00:19:17,480 --> 00:19:20,879
Speaker 1: taking the predicted future fundamentals and feeding that into a

343
00:19:20,960 --> 00:19:25,440
Speaker 1: value factor model. But you could imagine using let's say

344
00:19:25,600 --> 00:19:29,480
Speaker 1: the deep neural networks said, you know, a company is

345
00:19:29,480 --> 00:19:33,080
Speaker 1: going to do a hundred million, but consensus estimates in

346
00:19:33,080 --> 00:19:36,560
Speaker 1: in in earnings. Let's say, but consensus estimates said it's

347
00:19:36,560 --> 00:19:41,359
Speaker 1: gonna do seventy five million in in earnings. Well, you

348
00:19:41,400 --> 00:19:43,640
Speaker 1: know that might be you could you can imagine devising

349
00:19:43,640 --> 00:19:46,280
Speaker 1: a strategy around that where you'd want to go, you know,

350
00:19:46,359 --> 00:19:49,919
Speaker 1: bet on those guys and ones where consensus estimates are

351
00:19:49,960 --> 00:19:53,120
Speaker 1: above what the deep neural network is predicting, you'd want

352
00:19:53,119 --> 00:19:56,679
Speaker 1: to bet against. Right, assuming the current price is pricing

353
00:19:56,760 --> 00:19:59,520
Speaker 1: and that that's a really you know, John, you shouldnt

354
00:19:59,520 --> 00:20:03,040
Speaker 1: give away so goods, that that's a really good idea.

355
00:20:03,920 --> 00:20:08,800
Speaker 1: So are these kinds of machine learning driven predictive models

356
00:20:09,080 --> 00:20:11,320
Speaker 1: the future of investing? You think is that the way

357
00:20:11,320 --> 00:20:14,199
Speaker 1: that we're heading. I think what this paper showed is

358
00:20:14,240 --> 00:20:17,560
Speaker 1: that there's a lot of potential in using deep learning

359
00:20:17,680 --> 00:20:21,400
Speaker 1: to long term investing. I think that there's been some

360
00:20:21,480 --> 00:20:25,520
Speaker 1: debate about whether, you know, deep learning, which requires a

361
00:20:25,560 --> 00:20:29,760
Speaker 1: lot of data um to to to build successful models,

362
00:20:30,400 --> 00:20:34,320
Speaker 1: um whether in finance there's enough data, or whether you

363
00:20:34,359 --> 00:20:38,680
Speaker 1: even need this these kinds of complex models in finance,

364
00:20:38,720 --> 00:20:40,960
Speaker 1: I mean a lot of quant people feel, you know,

365
00:20:41,119 --> 00:20:44,679
Speaker 1: linear simple factor models are the best route to go,

366
00:20:45,520 --> 00:20:47,919
Speaker 1: And I think what we showed here is that if

367
00:20:47,920 --> 00:20:51,080
Speaker 1: you're trying to predict price changes, that might be true.

368
00:20:51,760 --> 00:20:55,800
Speaker 1: But if you decompose the problem into first trying to

369
00:20:55,840 --> 00:20:59,040
Speaker 1: predict fundamentals and then later you know, through a factor

370
00:20:59,119 --> 00:21:01,920
Speaker 1: model or some other method, trying to use those predicted

371
00:21:01,920 --> 00:21:06,720
Speaker 1: fundamentals to predict price, deep learning has a lot of

372
00:21:06,720 --> 00:21:12,200
Speaker 1: potential and does does substantially better at predicting future fundamentals

373
00:21:12,200 --> 00:21:14,280
Speaker 1: than than what you could do with a linear model.

374
00:21:14,600 --> 00:21:17,560
Speaker 1: There's a sort of a technical reason to recommend the

375
00:21:17,600 --> 00:21:20,760
Speaker 1: way we've cast a problem also without going too far

376
00:21:20,760 --> 00:21:24,520
Speaker 1: into the weeds. Basically, uh, you think really really powerful

377
00:21:24,600 --> 00:21:27,639
Speaker 1: machine learning models, deep neural networks. The thing that you

378
00:21:27,720 --> 00:21:29,879
Speaker 1: worried about is John was talking about how people people

379
00:21:29,960 --> 00:21:32,240
Speaker 1: agonize over what can you bring us to bear on

380
00:21:32,280 --> 00:21:34,520
Speaker 1: long term investing because you don't have as much data

381
00:21:34,960 --> 00:21:38,119
Speaker 1: right as if you were looking at the you know,

382
00:21:38,400 --> 00:21:42,679
Speaker 1: micro second kind of trade frequency, then you'd have, you know,

383
00:21:42,960 --> 00:21:45,399
Speaker 1: trillions of trade examples or something you get on. But

384
00:21:45,800 --> 00:21:47,760
Speaker 1: if you if you're looking at you know, your your

385
00:21:47,760 --> 00:21:49,879
Speaker 1: time tick is I have a data point you know,

386
00:21:49,960 --> 00:21:53,000
Speaker 1: once per month or once per year suddenly, and I

387
00:21:53,040 --> 00:21:56,439
Speaker 1: only have thousands of stocks, not millions of stocks. You

388
00:21:56,480 --> 00:22:00,199
Speaker 1: don't have such a huge amount of data. Um, So

389
00:22:00,840 --> 00:22:03,080
Speaker 1: what you worry about is that a model given given

390
00:22:03,119 --> 00:22:05,919
Speaker 1: a super powerful model, like a super overpowered model, and

391
00:22:05,960 --> 00:22:08,399
Speaker 1: then not too much data, that there's a propensity for

392
00:22:08,440 --> 00:22:10,840
Speaker 1: the models to do what we call overfitting, which is

393
00:22:10,880 --> 00:22:13,000
Speaker 1: the model basically it does a really good job of

394
00:22:13,040 --> 00:22:16,600
Speaker 1: memorizing the training data it's seen, but it learns kind

395
00:22:16,600 --> 00:22:19,639
Speaker 1: of a spurious pattern that doesn't generalize to future data

396
00:22:19,680 --> 00:22:23,040
Speaker 1: that it hasn't seen. So one cool thing about the

397
00:22:23,080 --> 00:22:25,800
Speaker 1: way that we're casting the problem is that we're not

398
00:22:25,880 --> 00:22:28,600
Speaker 1: just trying to predict the factor of interest. We're actually

399
00:22:28,600 --> 00:22:31,119
Speaker 1: trying to predict all the factors in the future. And

400
00:22:31,160 --> 00:22:35,159
Speaker 1: this means that the model has to simultaneously get the

401
00:22:35,200 --> 00:22:37,679
Speaker 1: income right, and get the assets right, and get the

402
00:22:37,720 --> 00:22:40,600
Speaker 1: debt right, and get all these different factors that are available.

403
00:22:40,600 --> 00:22:43,239
Speaker 1: So John was a fifteen target factors that we have

404
00:22:43,359 --> 00:22:46,960
Speaker 1: that we're trying to predict. So so in this case,

405
00:22:47,200 --> 00:22:48,680
Speaker 1: this sort of like this is this is what we

406
00:22:48,720 --> 00:22:51,960
Speaker 1: call multitask learning and the machine learning literature. And one

407
00:22:52,080 --> 00:22:55,320
Speaker 1: nice effect of multitask learning as that has a generalization

408
00:22:56,119 --> 00:22:58,520
Speaker 1: effect in that it's it's harder to fit a spurious

409
00:22:58,560 --> 00:23:00,960
Speaker 1: hypothesis because you have to come up with a representation

410
00:23:01,080 --> 00:23:03,679
Speaker 1: that is good for task one and also good for

411
00:23:03,720 --> 00:23:06,199
Speaker 1: task two, and also good for tast three. And the

412
00:23:06,240 --> 00:23:10,040
Speaker 1: probability that you come up with a pattern that's that's

413
00:23:10,080 --> 00:23:12,479
Speaker 1: good for solving all of these tasks that is not

414
00:23:12,600 --> 00:23:15,399
Speaker 1: the true pattern is much smaller than if you're only

415
00:23:15,520 --> 00:23:18,960
Speaker 1: like trying to solve one task, where it's easier to

416
00:23:19,000 --> 00:23:21,280
Speaker 1: just kind of memorize those data points. So we have

417
00:23:21,320 --> 00:23:24,960
Speaker 1: like essentially sixteen times as much training data and in

418
00:23:25,040 --> 00:23:29,400
Speaker 1: some relevant sense, so I have to ask in the

419
00:23:29,560 --> 00:23:35,879
Speaker 1: abstract of your paper or in the intro, you say that, um,

420
00:23:36,040 --> 00:23:40,399
Speaker 1: with this approach, you can improve your annual returns pretty

421
00:23:40,400 --> 00:23:44,240
Speaker 1: substantially over a standard factor model. In a bad test.

422
00:23:44,280 --> 00:23:48,280
Speaker 1: Seventeen point one percent versus a fourteen point four percent,

423
00:23:48,880 --> 00:23:52,119
Speaker 1: just pretty big beat. But as we know, and as

424
00:23:52,160 --> 00:23:54,000
Speaker 1: there's a lot of people pointed out, there's a lot

425
00:23:54,040 --> 00:23:57,840
Speaker 1: of strategies that seem to work in academic papers and

426
00:23:57,880 --> 00:24:01,639
Speaker 1: then when they're put into pract is, they don't seem

427
00:24:01,880 --> 00:24:06,080
Speaker 1: the results don't seem to arrive as easily. John, in

428
00:24:06,200 --> 00:24:09,800
Speaker 1: your firm, are you seeing the results of your research

429
00:24:10,280 --> 00:24:12,600
Speaker 1: that on paper look very compelling actually play out in

430
00:24:12,600 --> 00:24:17,280
Speaker 1: the market. So so this this paper, uh, we we

431
00:24:17,359 --> 00:24:21,280
Speaker 1: have not put this model to test, so to speak,

432
00:24:21,760 --> 00:24:25,280
Speaker 1: in in a fund yet, but you know, we're very

433
00:24:25,320 --> 00:24:28,560
Speaker 1: interested in in in doing that. I will I will

434
00:24:28,600 --> 00:24:31,879
Speaker 1: add though here that many of the back tests that

435
00:24:31,920 --> 00:24:36,520
Speaker 1: are done in the industry are done where you just run,

436
00:24:36,800 --> 00:24:39,080
Speaker 1: you know, thousands of back tests on a data set

437
00:24:39,119 --> 00:24:42,000
Speaker 1: over some time period ten years, twenty years, thirty years,

438
00:24:42,920 --> 00:24:46,080
Speaker 1: and there's no out of sample testing, meaning that they

439
00:24:46,080 --> 00:24:48,600
Speaker 1: don't then take that and then apply it to a

440
00:24:48,640 --> 00:24:52,440
Speaker 1: new data set. One thing that machine learning. One technique

441
00:24:52,560 --> 00:24:57,560
Speaker 1: that is used in machine learning to prevent overfitting, and

442
00:24:57,920 --> 00:25:01,080
Speaker 1: that we do here is we train them or we

443
00:25:01,200 --> 00:25:04,760
Speaker 1: build the model on one data set and then test

444
00:25:04,840 --> 00:25:07,640
Speaker 1: it at a sample on another on another data set

445
00:25:07,720 --> 00:25:10,520
Speaker 1: during a different time period, and the results we present

446
00:25:10,600 --> 00:25:14,800
Speaker 1: there are at a sample out of sample, always being

447
00:25:15,000 --> 00:25:17,399
Speaker 1: sort of ahead right in the future, you could. So

448
00:25:17,440 --> 00:25:19,560
Speaker 1: the model is the model at every given time. It

449
00:25:19,640 --> 00:25:23,800
Speaker 1: is trained on the path data. So we're simulating like

450
00:25:23,840 --> 00:25:26,200
Speaker 1: what if, you know, if you train the model back

451
00:25:26,240 --> 00:25:28,199
Speaker 1: then based on that it was only available up to

452
00:25:28,240 --> 00:25:30,960
Speaker 1: that point. I think more broadly, there's a good question

453
00:25:31,000 --> 00:25:35,160
Speaker 1: there of um, it's hard to say which which patterns

454
00:25:35,160 --> 00:25:38,359
Speaker 1: are just you know, especially I think with short term investing,

455
00:25:38,359 --> 00:25:41,439
Speaker 1: it's very obvious that any any pattern that exists on

456
00:25:41,440 --> 00:25:44,120
Speaker 1: a scale of seconds is something that could be sort

457
00:25:44,160 --> 00:25:47,960
Speaker 1: of traded away. It's not as clear and I believe,

458
00:25:48,040 --> 00:25:50,320
Speaker 1: I mean, John can speak more to it, right, But

459
00:25:50,359 --> 00:25:52,560
Speaker 1: I believe part of the ethos of long term investing

460
00:25:52,640 --> 00:25:55,960
Speaker 1: is very much that rather than interacting in a place

461
00:25:56,000 --> 00:25:59,080
Speaker 1: where the price most price movements are due to the

462
00:25:59,119 --> 00:26:03,159
Speaker 1: behavior of the high frequency traders, when you're in the

463
00:26:03,200 --> 00:26:06,199
Speaker 1: long term space, the price movement is more tied to

464
00:26:06,400 --> 00:26:09,680
Speaker 1: the actual fiscal performance of the company, and that's maybe

465
00:26:09,680 --> 00:26:13,480
Speaker 1: a more durable pattern. So Joe and I were talking

466
00:26:13,640 --> 00:26:18,480
Speaker 1: about financial players and how quickly they adapt to new

467
00:26:18,560 --> 00:26:21,800
Speaker 1: markets and new situations. At the beginning of this episode,

468
00:26:22,320 --> 00:26:26,320
Speaker 1: from your respective viewpoints, how fast are these sorts of

469
00:26:26,359 --> 00:26:30,680
Speaker 1: technologies and models and applications being developed, And for how

470
00:26:30,760 --> 00:26:35,720
Speaker 1: long would something like, you know, a clairvoyant factor predicting

471
00:26:35,800 --> 00:26:39,080
Speaker 1: model actually give you an edge four until someone else,

472
00:26:39,520 --> 00:26:41,560
Speaker 1: maybe an E t F came along and copied it.

473
00:26:42,480 --> 00:26:46,280
Speaker 1: I think that's a hard question to answer, because again

474
00:26:46,320 --> 00:26:49,280
Speaker 1: it gets back to how you would use this model. Right,

475
00:26:49,400 --> 00:26:53,080
Speaker 1: So in the in the paper we give one very

476
00:26:53,119 --> 00:26:56,560
Speaker 1: specific example. We use the deep learning normal network to

477
00:26:56,600 --> 00:27:00,240
Speaker 1: predict fundamentals and then we plug that into one kind

478
00:27:00,280 --> 00:27:05,000
Speaker 1: of factor model, right uh, in particular operating income predicted

479
00:27:05,080 --> 00:27:08,760
Speaker 1: operating income over enterprise value. But as I suggested, you

480
00:27:08,800 --> 00:27:11,840
Speaker 1: could use it to you know, figure you know, figure

481
00:27:11,880 --> 00:27:15,480
Speaker 1: out whether consensus forecasts are good or bad. Um. So,

482
00:27:16,400 --> 00:27:19,320
Speaker 1: you know, I think that just saying in general, deep

483
00:27:19,400 --> 00:27:23,640
Speaker 1: learning applied to you know, investing is going to get

484
00:27:23,720 --> 00:27:25,919
Speaker 1: used and then a year later is going to be

485
00:27:26,000 --> 00:27:30,080
Speaker 1: arbitraged away. Miss is the point that, look, you know,

486
00:27:30,320 --> 00:27:33,119
Speaker 1: you can use deep learning in a myriad of ways

487
00:27:33,320 --> 00:27:36,960
Speaker 1: to attack the problem of long term investing and presumably

488
00:27:37,160 --> 00:27:42,080
Speaker 1: trading as well. To address your question about how quickly

489
00:27:42,400 --> 00:27:45,919
Speaker 1: is this kind of technology getting adopted? UM my sense

490
00:27:46,520 --> 00:27:49,840
Speaker 1: and based a little bit on an outsider's view as

491
00:27:49,840 --> 00:27:53,879
Speaker 1: an academic machine learning person, um talking to collegues who

492
00:27:53,920 --> 00:27:56,600
Speaker 1: have either gone into fintech or who flirted with it

493
00:27:56,840 --> 00:28:00,359
Speaker 1: or tried to recruit me into it. The sense that

494
00:28:00,440 --> 00:28:03,520
Speaker 1: I get is that actually, obviously a lot of people

495
00:28:03,560 --> 00:28:06,760
Speaker 1: aren't talking about what they're doing, right, But my sense

496
00:28:06,840 --> 00:28:09,120
Speaker 1: is that there's a lot of people doing this kind

497
00:28:09,160 --> 00:28:12,600
Speaker 1: of stuff in the high frequency space, not maybe on

498
00:28:12,600 --> 00:28:14,960
Speaker 1: the scale of you know, fractions of seconds, but but

499
00:28:15,080 --> 00:28:17,720
Speaker 1: on on a pretty short time scale. And the reason

500
00:28:17,760 --> 00:28:23,680
Speaker 1: why is because, um, right, it's it's easy to collect

501
00:28:23,720 --> 00:28:27,159
Speaker 1: a lot of data. If the patterns are very different, um,

502
00:28:27,320 --> 00:28:30,480
Speaker 1: a year from now, Well, you just you have enough data.

503
00:28:30,680 --> 00:28:34,359
Speaker 1: Like if you're trading like at the scale of months

504
00:28:34,440 --> 00:28:37,160
Speaker 1: or years, then you have to look back twenty years, right,

505
00:28:37,400 --> 00:28:39,360
Speaker 1: you have to look back thirty If you're trading at

506
00:28:39,360 --> 00:28:42,480
Speaker 1: the scale seconds, then your whole universe could be formed

507
00:28:42,480 --> 00:28:45,400
Speaker 1: by the previous four days. There's a very fast cycle

508
00:28:45,440 --> 00:28:47,520
Speaker 1: of development. So if you're in it and you just

509
00:28:47,600 --> 00:28:50,080
Speaker 1: wanna you don't you don't have any kind of strong

510
00:28:50,120 --> 00:28:54,000
Speaker 1: beliefs about finance. You're just a machine learning person throwing

511
00:28:54,000 --> 00:28:57,600
Speaker 1: your hammer at financing. Then going in the high frequency

512
00:28:57,600 --> 00:29:00,800
Speaker 1: space gives you or are the comparatively high frequency space

513
00:29:01,200 --> 00:29:04,200
Speaker 1: gives you like the sandbacked box to just really quickly

514
00:29:04,360 --> 00:29:07,440
Speaker 1: test stuff validated, see if it works. My feeling and

515
00:29:07,800 --> 00:29:09,640
Speaker 1: when I've talked to friends who are doing this kind

516
00:29:09,640 --> 00:29:12,040
Speaker 1: of stuff, but what we're doing is that I think

517
00:29:12,280 --> 00:29:14,680
Speaker 1: almost no one that I've talked to out of a

518
00:29:14,680 --> 00:29:16,400
Speaker 1: lot of people doing this stuff with finance, is looking

519
00:29:16,400 --> 00:29:19,640
Speaker 1: at the same kinds of time scales. And John might

520
00:29:19,680 --> 00:29:21,920
Speaker 1: be able to to speak to that because he might

521
00:29:22,280 --> 00:29:25,640
Speaker 1: actually be deeper and the I mean, he's definitely differ

522
00:29:25,680 --> 00:29:28,280
Speaker 1: in the finance community than I am. But my sense

523
00:29:28,320 --> 00:29:30,720
Speaker 1: is people doing deep learning for finance, and there are many,

524
00:29:31,200 --> 00:29:33,600
Speaker 1: Um it's on the rise, but they're not necessarily looking

525
00:29:33,640 --> 00:29:35,800
Speaker 1: at it in the same way, and certainly very few

526
00:29:35,960 --> 00:29:38,480
Speaker 1: on as long a time scale. Yeah, I mean, I

527
00:29:38,520 --> 00:29:40,800
Speaker 1: think if you look, you know, the a q r

528
00:29:40,920 --> 00:29:43,120
Speaker 1: s and the d f as of the world, which

529
00:29:43,120 --> 00:29:46,840
Speaker 1: are you know, these huge you know, quantitative shops. They do.

530
00:29:46,960 --> 00:29:51,000
Speaker 1: They certainly do long term investing, but um, there's not

531
00:29:51,040 --> 00:29:53,560
Speaker 1: a lot of evidence that there's a ton of machine

532
00:29:53,640 --> 00:29:56,440
Speaker 1: learning deep learning going on there. But you know, I

533
00:29:56,480 --> 00:29:59,040
Speaker 1: think if if if stuff is a successful, you know

534
00:29:59,080 --> 00:30:02,320
Speaker 1: it's likely to be a opted, so probably won't be

535
00:30:02,360 --> 00:30:06,920
Speaker 1: true forever. John Elberg and Zachary Lipton, that was a

536
00:30:07,000 --> 00:30:11,440
Speaker 1: fascinating conversation, so much to think about and wrap our

537
00:30:11,440 --> 00:30:15,320
Speaker 1: heads around. Really appreciate you both coming on. Thanks for

538
00:30:15,360 --> 00:30:29,280
Speaker 1: having us. Thank you guys. Well, Tracy, we didn't really

539
00:30:29,280 --> 00:30:30,920
Speaker 1: plan it that way, but I really do think that

540
00:30:31,000 --> 00:30:35,160
Speaker 1: was sort of the perfect follow up to Andrew Lot left. No, Joe,

541
00:30:35,200 --> 00:30:37,400
Speaker 1: you're supposed to pretend we did plan it that way

542
00:30:37,440 --> 00:30:40,240
Speaker 1: so everyone will think we're really working so good. I mean,

543
00:30:40,320 --> 00:30:45,280
Speaker 1: like we should continue this series on quantitative strategies and

544
00:30:45,440 --> 00:30:48,280
Speaker 1: new ways to evolve to beat the market. Let's continue

545
00:30:48,320 --> 00:30:52,160
Speaker 1: this continue. Yes, absolutely, Okay, in all seriousness, Yes, it

546
00:30:52,240 --> 00:30:55,240
Speaker 1: was fascinating. I really like the idea of well, who

547
00:30:55,280 --> 00:30:58,520
Speaker 1: doesn't like the idea of a clairvoyant robot who can

548
00:30:58,560 --> 00:31:00,760
Speaker 1: predict how well a company is going to do in

549
00:31:00,800 --> 00:31:04,480
Speaker 1: the future and then apply that to a factor based

550
00:31:04,720 --> 00:31:08,720
Speaker 1: investment model. If someone comes back in time and they're

551
00:31:08,760 --> 00:31:10,800
Speaker 1: like giving me hints on what the stock market is

552
00:31:10,800 --> 00:31:12,800
Speaker 1: gonna do, it's like just give me the winning stocks.

553
00:31:12,840 --> 00:31:15,200
Speaker 1: You know what I'm saying, Like, if you're time traveling,

554
00:31:15,320 --> 00:31:17,880
Speaker 1: don't like be a tease, just give me the winning stocks. No,

555
00:31:18,160 --> 00:31:21,960
Speaker 1: But in all seriousness, hey, I felt like several times

556
00:31:21,960 --> 00:31:24,840
Speaker 1: in that conversation, it's just like the level that they're

557
00:31:24,880 --> 00:31:28,160
Speaker 1: operating and thinking about the market on is so like

558
00:31:28,520 --> 00:31:31,280
Speaker 1: high above anything that you and I like typically talk

559
00:31:31,360 --> 00:31:33,640
Speaker 1: about it on a day. Like several times I felt

560
00:31:33,640 --> 00:31:37,200
Speaker 1: like I had to catch my breath speak for yourself, Joe,

561
00:31:38,840 --> 00:31:41,600
Speaker 1: because it's just like absorbing all of that, and you know,

562
00:31:41,600 --> 00:31:44,280
Speaker 1: obviously there's probably lots that I didn't get. But then

563
00:31:44,320 --> 00:31:47,680
Speaker 1: the other thing I really thought that last point was

564
00:31:47,760 --> 00:31:51,280
Speaker 1: very interesting about time. So obviously going back to the

565
00:31:51,320 --> 00:31:54,400
Speaker 1: adaptive framework for thinking about markets. You know, if there

566
00:31:54,520 --> 00:31:58,480
Speaker 1: is a strategy that works over a day and you

567
00:31:58,520 --> 00:32:00,440
Speaker 1: can get it, and you can you just have to

568
00:32:00,480 --> 00:32:03,560
Speaker 1: back test four days or whatever, it's very easy to see, Okay,

569
00:32:03,560 --> 00:32:05,760
Speaker 1: this works, this doesn't. Let's go with the thing that

570
00:32:05,800 --> 00:32:07,520
Speaker 1: works and then everyone can sort of figure out the

571
00:32:07,520 --> 00:32:09,880
Speaker 1: things that work and then it doesn't work anymore. But

572
00:32:09,960 --> 00:32:13,320
Speaker 1: this idea that maybe a quantitative approach to long term

573
00:32:13,360 --> 00:32:17,120
Speaker 1: fundamental investing, you don't get that sort of instant feedback

574
00:32:17,120 --> 00:32:19,719
Speaker 1: on whether it's working as fast, and so maybe winning

575
00:32:19,760 --> 00:32:22,120
Speaker 1: strategies might prove to be a bit more do Yeah,

576
00:32:22,120 --> 00:32:25,400
Speaker 1: and presumably it's much more difficult to actually develop them

577
00:32:25,440 --> 00:32:28,239
Speaker 1: and see them evolved. It's like, um, I guess it's

578
00:32:28,320 --> 00:32:33,240
Speaker 1: like if you bred successive generations of like rabbits, right,

579
00:32:33,400 --> 00:32:36,440
Speaker 1: Like it takes like a year two, well less than

580
00:32:36,480 --> 00:32:40,400
Speaker 1: a year, you could breathe like a hundred generations in

581
00:32:40,440 --> 00:32:43,200
Speaker 1: a year. Yeah, right, Or I guess, like, you know,

582
00:32:43,320 --> 00:32:46,320
Speaker 1: it's like laboratories use mice and because they do, like

583
00:32:46,360 --> 00:32:49,000
Speaker 1: they can get so many so fast. But if you

584
00:32:49,080 --> 00:32:52,960
Speaker 1: had to sort of, you know, breed hippopotamus is you

585
00:32:52,960 --> 00:32:56,760
Speaker 1: wouldn't know for a much longer period of time whether

586
00:32:56,840 --> 00:32:59,080
Speaker 1: you're down the road you had sort of created the

587
00:32:59,120 --> 00:33:04,480
Speaker 1: master hip hop. Does that make sense? Wow? This past? Yeah, Okay,

588
00:33:04,560 --> 00:33:07,440
Speaker 1: let's leave it at master hippo. Okay, alright. This has

589
00:33:07,480 --> 00:33:11,320
Speaker 1: been another episode of the Odd Thoughts Podcast. I'm Tracy Alloway.

590
00:33:11,440 --> 00:33:13,960
Speaker 1: You can follow me on Twitter at Tracy Alloway and

591
00:33:14,000 --> 00:33:16,800
Speaker 1: I'm Joe Why Isn't All? You can follow me on

592
00:33:16,840 --> 00:33:19,880
Speaker 1: Twitter at the Stalwart, and you can follow our guests

593
00:33:19,960 --> 00:33:24,280
Speaker 1: on Twitter John Elberg is at at John Elbert. There's

594
00:33:24,320 --> 00:33:27,600
Speaker 1: just one L in that. Zachary Lipton is on Twitter

595
00:33:27,800 --> 00:33:31,800
Speaker 1: at Zachary Lipton, and you can follow our producer Sarah

596
00:33:31,800 --> 00:33:35,640
Speaker 1: Patterson on Twitter at Sarah pat with two teams. Thanks

597
00:33:35,640 --> 00:33:36,160
Speaker 1: for listening,