WEBVTT - How Hudson River Trading Actually Uses AI 0:00:02.920 --> 0:00:16.560 Bloomberg Audio Studios, Podcasts, Radio News. 0:00:18.040 --> 0:00:21.360 Hello and welcome to another episode of the Odd Lots Podcast. 0:00:21.400 --> 0:00:23.599 I'm Jill Wisenthal and I'm Tracy Alloway. 0:00:23.920 --> 0:00:26.920 Tracy, I've always had this idea for the podcast, or 0:00:26.960 --> 0:00:30.319 a thing that I've wanted to do. Okay, conceptually with 0:00:30.440 --> 0:00:34.400 podcasts is schedule every guest for two interviewers. So you 0:00:34.479 --> 0:00:36.280 have the opening interview and you ask a bunch of 0:00:36.360 --> 0:00:39.159 questions and then it's, oh God, I really wish I 0:00:39.200 --> 0:00:41.440 had followed up on that. I had more. I was 0:00:41.600 --> 0:00:43.519 just starting to sort of get my head around this thing. 0:00:43.600 --> 0:00:43.680 Now. 0:00:43.720 --> 0:00:45.640 I could have asked the good questions and then like, 0:00:45.720 --> 0:00:48.640 have the person come back next week. Also, the audience complains, 0:00:48.640 --> 0:00:50.480 I wish it as that and then fill in all 0:00:50.479 --> 0:00:54.080 those gaps that had been inspired by the previous conversation. 0:00:54.320 --> 0:00:55.920 I don't think it's a bad idea. I think it 0:00:55.920 --> 0:00:58.840 would double the number of episodes that we put out. 0:00:59.320 --> 0:01:03.320 But sure there are topics that come up, usually things 0:01:03.360 --> 0:01:06.080 that were just kind of new to and we're trying 0:01:06.120 --> 0:01:08.800 to learn about specifically technical things, and one of those 0:01:09.000 --> 0:01:09.319 has to. 0:01:09.319 --> 0:01:12.600 Be AI, right, Ai? And also, you know, I really 0:01:12.680 --> 0:01:14.880 had a great time. I guess last month we were 0:01:14.880 --> 0:01:17.039 in Chicago. Yeah, we talked to a bunch of different 0:01:17.080 --> 0:01:21.880 it was like got trading related trip. We interviewed Don Wilson, 0:01:22.040 --> 0:01:23.960 we interviewed the head of the CMME. We had some 0:01:23.959 --> 0:01:25.880 other chats. So they're all about the world of trading. 0:01:26.200 --> 0:01:27.920 When it comes to trading, it's like, you know, we 0:01:27.959 --> 0:01:31.319 talked to long term investors, portfolio managers and daomas. We 0:01:31.400 --> 0:01:33.280 talked to some people in the hedge fund space who 0:01:33.319 --> 0:01:36.279 like maybe have a holding period of several weeks or whatever. 0:01:36.840 --> 0:01:39.080 I actually really want to learn more about the trading 0:01:39.160 --> 0:01:41.080 like these people who have like a holding time of 0:01:41.360 --> 0:01:43.399 one second or something like that, because that's where a 0:01:43.400 --> 0:01:45.360 lot of the tech and a lot of the actual 0:01:45.440 --> 0:01:47.880 like action is and how that world makes money and 0:01:47.920 --> 0:01:51.200 how they actually deployed technology is very interesting, but still 0:01:51.200 --> 0:01:52.600 something I don't have my handle. 0:01:52.360 --> 0:01:55.560 On, well, the practical application, right, and also the culture 0:01:55.640 --> 0:01:57.960 of AI on Wall Street. I find that really interesting 0:01:58.000 --> 0:02:00.200 because I remember, I guess it was like more than 0:02:00.200 --> 0:02:02.520 a decade ago, but remember Lloyd Blank find saying that 0:02:02.560 --> 0:02:06.680 Golden Sachs is a technology. Yeah, and all these bank 0:02:06.760 --> 0:02:09.720 CEOs saying we're going to install pingpong tables to get 0:02:09.760 --> 0:02:12.800 all the coders, and now I see ads at trading 0:02:12.840 --> 0:02:15.160 firms and it's like, we have a data center full 0:02:15.200 --> 0:02:17.280 of B two hundreds, or we have a data center 0:02:17.320 --> 0:02:19.359 full of G three hundreds. Come work for us. 0:02:19.440 --> 0:02:21.440 The only thing besides all their tech that I know 0:02:21.680 --> 0:02:23.880 is like every time you read a profile of any 0:02:23.960 --> 0:02:27.320 trading company, like and they love to play backcam and 0:02:27.320 --> 0:02:30.959 they love to play all the article the chessboards are out, 0:02:31.280 --> 0:02:33.760 they could be seen playing chess over launch, et cetera. 0:02:33.960 --> 0:02:36.120 I get it, Okay, they like us, they like games, 0:02:36.160 --> 0:02:38.799 they like whatever, let's move the ball for Well. 0:02:38.720 --> 0:02:42.880 There's also the underlying theme of is this all hype, right? 0:02:43.160 --> 0:02:45.640 Because you two get the sense sometimes that companies are 0:02:45.680 --> 0:02:48.520 putting out press releases where they just mention AI to 0:02:48.560 --> 0:02:50.800 tick a box, to be seen to be doing something 0:02:51.240 --> 0:02:54.080 and hope that their stock actually goes up. And because 0:02:54.120 --> 0:02:57.200 so much of this is proprietary and people kind of 0:02:57.240 --> 0:03:00.360 have an excuse not to go into detail about it, 0:03:01.040 --> 0:03:04.320 sometimes you do get the feeling that people are just 0:03:04.520 --> 0:03:06.480 talking about it and not actually using it. 0:03:06.960 --> 0:03:09.360 Cynics and I'm not saying. 0:03:09.160 --> 0:03:11.640 This myself, I know you're not a cynic. 0:03:11.720 --> 0:03:17.639 Speaking of trading and technology. Cinics would say that comes 0:03:17.720 --> 0:03:20.000 deal with Google to both of clouds, to you put 0:03:20.040 --> 0:03:21.920 trading on the cloud with hype, that that was a 0:03:21.919 --> 0:03:24.200 press release. People have said that people have made that 0:03:24.280 --> 0:03:26.560 charge and they don't understand why. You don't have to comment. 0:03:26.600 --> 0:03:27.840 You don't have to say anything further on that. 0:03:27.960 --> 0:03:29.720 I do have a comment, but I'll hold it for 0:03:29.800 --> 0:03:30.119 our guess. 0:03:30.120 --> 0:03:32.640 I'm just there is this world where people do press 0:03:32.680 --> 0:03:34.960 releases and cynics go. I don't really understand the point anyway. 0:03:35.000 --> 0:03:37.560 There's a very long lind up. Let's learn more about 0:03:37.600 --> 0:03:40.240 the world of trading. Let's learn more about AI and 0:03:40.280 --> 0:03:43.160 tech specifically, what does it even mean to apply AI 0:03:43.200 --> 0:03:45.320 within the realm of trading. We're going to be speaking 0:03:45.360 --> 0:03:47.920 with Ian Dunning. He is the head of AI at 0:03:48.000 --> 0:03:51.840 Hudson Rivert Trading. He's previously at deep Mind, so his 0:03:52.440 --> 0:03:55.960 trading and AI bonafides are about as good. 0:03:55.840 --> 0:03:56.360 As it gets. 0:03:56.440 --> 0:03:58.520 With me, you've established them, we've established that. 0:03:58.920 --> 0:04:00.840 Really the perfect guest answer all our questions. So I 0:04:01.080 --> 0:04:03.200 thank you so much for coming on the podcast. 0:04:03.280 --> 0:04:04.480 Yeah, I'm really happy to be here. 0:04:04.560 --> 0:04:08.800 I agree you as the mystique factor is kind of overblown, 0:04:08.920 --> 0:04:11.000 even if it's understandable white people embrace it. 0:04:11.040 --> 0:04:13.720 Sometimes we're gonna blow past the mystiq. Let's start with 0:04:13.840 --> 0:04:16.919 some like really just like rhudimentary questions, Like just the 0:04:16.920 --> 0:04:19.960 first one is like Huns the River trading as a company, 0:04:20.360 --> 0:04:21.400 how does it make money? 0:04:21.720 --> 0:04:27.440 Yeah, so we are a sort of quantitative automated proprietary 0:04:27.480 --> 0:04:30.240 trading firm. Which is a lot of words, but I 0:04:30.279 --> 0:04:31.760 guess the way I see it is we are a 0:04:31.800 --> 0:04:35.520 service provider to markets. Okay, the most clear example is 0:04:35.720 --> 0:04:38.120 market making. There is like a sort of utility to 0:04:38.160 --> 0:04:41.760 the world of being really just buy yourself any product, anytime, anywhere, 0:04:41.839 --> 0:04:46.600 and for us that means stocks, futures, options, crypto bonds. 0:04:47.240 --> 0:04:50.279 And if you could say, build a magical machine to 0:04:50.800 --> 0:04:52.960 quote a price to buy it or sell at any instrument, 0:04:53.040 --> 0:04:55.400 and you would want to be like the best possible price, 0:04:55.440 --> 0:04:56.359 like the tightest price. 0:04:56.680 --> 0:04:57.640 People would trade with you. 0:04:58.160 --> 0:05:00.560 They would be happy because there's a count for their 0:05:00.560 --> 0:05:02.320 trade and they get a kind of good price, like 0:05:02.360 --> 0:05:03.080 a low spread. 0:05:03.320 --> 0:05:04.400 And we're happy because. 0:05:04.160 --> 0:05:06.400 We essentially pick up a penny in front of a steamroller, 0:05:06.440 --> 0:05:08.920 like we are making sort of money from that spread, 0:05:09.040 --> 0:05:10.920 and we can pick up the pennies in front of 0:05:10.920 --> 0:05:14.600 a steamroller if we have a really magical device which 0:05:14.600 --> 0:05:16.200 tells us how what everything should be. 0:05:16.120 --> 0:05:17.120 When the steamroller is coming. 0:05:17.240 --> 0:05:17.840 Yeah, it tells us. 0:05:17.760 --> 0:05:19.680 When the steam roll is coming. And so I think 0:05:19.720 --> 0:05:22.720 that's kind of the very very sophisticated sort of middleman 0:05:22.880 --> 0:05:24.520 in some sense. And the same we've had Amazon is 0:05:24.560 --> 0:05:27.480 Amazon doesn't make stuff, but it's a very valuable, profitable 0:05:27.480 --> 0:05:30.200 company provides a service. People get value of same thing. 0:05:30.240 --> 0:05:34.560 We're moving stock spons through time and space between different counterpartties. 0:05:34.600 --> 0:05:36.120 And yeah, we will. 0:05:36.000 --> 0:05:38.760 Ask you about the steamroller in a few minutes. But 0:05:39.320 --> 0:05:41.960 before we do that, how does AI or the way 0:05:41.960 --> 0:05:45.679 you're using AI actually differ from the algorithmic or quant 0:05:45.680 --> 0:05:48.640 trading of olds, Because I guess that one of the 0:05:48.720 --> 0:05:52.600 questions is, is this, you know, a sort of evolutionary change, 0:05:52.720 --> 0:05:55.919 you know, maybe a marginal improvement on what already exists, 0:05:56.120 --> 0:05:59.279 or is this something seismic and a step change the 0:05:59.480 --> 0:06:01.560 big ship in the way trading actually works. 0:06:01.920 --> 0:06:05.120 Yeah, I mean I don't want to overstate ourselves in 0:06:05.160 --> 0:06:07.840 some sense, because into space, as you mentioned before, it's 0:06:07.920 --> 0:06:10.400 very like opaque what sort of different firms of this 0:06:10.520 --> 0:06:13.640 class are doing. I can siddenly speak to our own experience, 0:06:13.800 --> 0:06:15.760 which is we've been doing this type of trading for 0:06:15.880 --> 0:06:19.240 twenty plus years, and much like everyone who was doing this, 0:06:19.440 --> 0:06:22.960 the way it kind of worked was you handcraft features. 0:06:22.800 --> 0:06:24.359 It's sort of based on human intuition. 0:06:24.480 --> 0:06:25.560 Oh, I don't know. 0:06:25.600 --> 0:06:27.479 If the order book looks imbalanced, there is more people 0:06:27.480 --> 0:06:29.200 wanting to buy than sell, the price is going to 0:06:29.279 --> 0:06:31.880 go up soon, or something like that. And maybe you 0:06:32.360 --> 0:06:34.000 get a bunch of very smart people and they think 0:06:34.080 --> 0:06:36.080 very hard, like it's almost like making a very fancy 0:06:36.160 --> 0:06:39.240 watch kind of artistanally craft all these pieces, and then 0:06:39.240 --> 0:06:43.080 maybe you use relatively simple mathematical techniques like linear regression 0:06:43.080 --> 0:06:46.680 to combine those predictors. And I've been going to conferences 0:06:46.720 --> 0:06:48.640 and things and recruiting for a long time, and even 0:06:48.680 --> 0:06:51.520 if today's going the internet, you'll people say things like, oh, 0:06:51.560 --> 0:06:54.479 that's all you can do in finance. For some reason, 0:06:54.480 --> 0:06:57.240 they'll say this. They'll say something like, oh it's too noisy, 0:06:57.320 --> 0:07:00.080 or markets are too nonstationary or things like this, and 0:07:00.120 --> 0:07:03.080 so that's all you can do. And I guess that 0:07:03.320 --> 0:07:06.320 belief isn't really backed up by anything in my opinion, 0:07:06.440 --> 0:07:09.760 and like lived experience, I guess, And so we sort 0:07:09.800 --> 0:07:12.000 of viewed it more for a long time as well, 0:07:12.760 --> 0:07:15.360 everything that's happening in the world, and ideally you would 0:07:15.360 --> 0:07:17.040 put this into kind of like a machine that does 0:07:17.120 --> 0:07:18.920 not have any human biases. 0:07:19.240 --> 0:07:20.040 I don't know how to. 0:07:19.960 --> 0:07:23.800 Trade stocks myself, like I buy broad market ets, what 0:07:23.880 --> 0:07:25.640 do I know? And so we but if you could 0:07:25.640 --> 0:07:27.840 put all the data into a box and it kind 0:07:27.840 --> 0:07:30.360 of could jurn all about data, it would find things 0:07:30.400 --> 0:07:33.280 that you would never be able to do, this handcrafted thing. 0:07:33.960 --> 0:07:36.880 And we started doing that very early, relatively in se 0:07:37.000 --> 0:07:40.960 twenty fourteen twenty thirteen period, and over time, over less 0:07:41.000 --> 0:07:43.400 a decade or so, much like in other contexts that 0:07:43.440 --> 0:07:46.000 are not finance, there has been sort of a hockey 0:07:46.000 --> 0:07:48.080 stick and you can measure by the sizes of the 0:07:48.120 --> 0:07:52.400 models the compute deployed and over time that way of 0:07:52.560 --> 0:07:56.840 modeling the markets initially was not like a hybrid with 0:07:56.880 --> 0:07:57.560 the traditional way. 0:07:57.680 --> 0:07:59.320 Bially kind of just like overtook it. 0:07:59.360 --> 0:08:03.280 Entirely and so now our trading is entirely driven by 0:08:03.680 --> 0:08:07.640 this magical machine consumes of a data. I kind of 0:08:07.720 --> 0:08:09.560 keep saying this magical machine. It consumes of a data 0:08:09.560 --> 0:08:11.640 for a reason, which is that this is how chat 0:08:11.680 --> 0:08:15.360 GPT is trained, it consumes all the data all the Internet. 0:08:15.520 --> 0:08:18.400 It's kind of scraped and connected into one place. When 0:08:18.400 --> 0:08:20.600 you train a model, that kind of takes it all 0:08:20.760 --> 0:08:23.280 and something emergent comes from it. And that's why I'm 0:08:23.360 --> 0:08:25.160 kind of a bit of leading. But that's why I'm 0:08:25.160 --> 0:08:26.520 talking about in the sense, and I think that is 0:08:26.560 --> 0:08:28.800 materially different from the like I'm using my intuition of 0:08:28.800 --> 0:08:31.320 the markets to kind of construct a predictive model. 0:08:31.640 --> 0:08:34.839 So just to be clear, how much of usefulness of 0:08:34.880 --> 0:08:38.600 AI here is about execution and the fact that you 0:08:38.640 --> 0:08:41.760 can crunch a lot of data really quickly with hundreds 0:08:41.880 --> 0:08:47.400 or thousands of GPUs versus spotting sophisticated patterns or discrepancies 0:08:47.440 --> 0:08:48.480 that you can exploit. 0:08:48.800 --> 0:08:49.760 I think it's both. 0:08:49.840 --> 0:08:52.400 I think one of the things that people sort of 0:08:52.480 --> 0:08:54.560 missed with a whole like do a linear regression type 0:08:54.559 --> 0:08:57.160 thing is when you really think about how much data 0:08:57.200 --> 0:08:59.640 there is in financial markets generated. And when I say data, 0:08:59.679 --> 0:09:01.880 I think it's important to think of it as every 0:09:01.960 --> 0:09:03.960 event that happens in market. It's not the sort of 0:09:04.000 --> 0:09:06.480 time serious of prices, but like the actual low level 0:09:06.679 --> 0:09:11.360 substrate people are quoting trading retracting quotes that like low 0:09:11.440 --> 0:09:15.440 level stuff is internet scale data set sizes, and one 0:09:15.440 --> 0:09:18.240 of us sort of bitter lesson ye type things of 0:09:18.320 --> 0:09:20.800 AI was like, you know, you shouldn't think too hard 0:09:20.800 --> 0:09:22.959 about how to feature engineer this in pre process, that 0:09:22.960 --> 0:09:25.960 you should kind of throttle in to something a form 0:09:26.040 --> 0:09:28.280 of computation that can kind of make use of internet 0:09:28.320 --> 0:09:31.200 scale data. In the twenty tens, it was like computer 0:09:31.280 --> 0:09:34.520 vision people used to make detectors for edges of images 0:09:34.520 --> 0:09:36.320 and things and they would combine them and same thing. 0:09:36.360 --> 0:09:38.320 It's like that was a good approach, but you know, 0:09:38.520 --> 0:09:40.839 it's completely dominated by the idea of getting very large 0:09:40.880 --> 0:09:42.920 umber of GPUs and a kind of a pretty generic 0:09:43.160 --> 0:09:46.520 neural network form empowering through it. As for like the 0:09:46.840 --> 0:09:49.240 how is it finding things that other methods could not. 0:09:50.000 --> 0:09:53.679 It's very hard to say our models are not very interpretable. 0:09:53.760 --> 0:09:57.400 And I think that's fine because, as Joe mentioned, our 0:09:57.480 --> 0:09:59.920 sort of trading style and holding times, a bit of 0:10:00.200 --> 0:10:03.040 it is like minutes, hours, maybe like a low single 0:10:03.040 --> 0:10:06.320 digit days for the most part, And I guess in 0:10:06.360 --> 0:10:09.840 my mind it's unreasonable to expect them to be interpretable 0:10:10.400 --> 0:10:13.000 because I don't know if I looked at the autobook 0:10:13.080 --> 0:10:16.200 data for Tesla or something. Am I really going to 0:10:16.200 --> 0:10:17.680 be able to tell you better than random with the 0:10:17.679 --> 0:10:20.080 price of Tesla will be in a minute's time? And 0:10:20.120 --> 0:10:21.440 so I kind of think it like that, if you 0:10:21.440 --> 0:10:25.239 have something that's clearly superhuman already, what level of interpretability 0:10:25.280 --> 0:10:28.000 because you expect like this is very different right to 0:10:28.040 --> 0:10:28.520 normal AI. 0:10:28.679 --> 0:10:32.120 Right, this is gets into some areas that I'm very interested. 0:10:32.280 --> 0:10:35.280 But just to like establish what we're talking about, you're 0:10:35.320 --> 0:10:38.240 trading a stock like a Tesla in video, et cetera 0:10:38.960 --> 0:10:42.839 with your magic machine machine. We had another episode where 0:10:43.400 --> 0:10:44.200 that was the money box. 0:10:44.240 --> 0:10:46.800 That's a magic bo different, that's a different one. 0:10:46.920 --> 0:10:50.960 With this AI machine, it is sort of arguably grown, right, 0:10:50.960 --> 0:10:52.800 and it's sort of grown in a lab more than 0:10:52.840 --> 0:10:55.600 it is programmed, much like a chatbot. I know, it's 0:10:55.679 --> 0:10:59.520 very different technology, like what is the price of in 0:10:59.640 --> 0:11:01.720 video going to be tomorrow? Or what is the price 0:11:01.760 --> 0:11:04.600 of ennvideo going to be this afternoon? What you're saying 0:11:04.840 --> 0:11:09.280 is with your technology, you have a better chance of 0:11:09.320 --> 0:11:12.000 getting that right, that you actually might be able to 0:11:12.080 --> 0:11:14.840 make an informed prediction about the future in a way 0:11:14.880 --> 0:11:17.680 that you couldn't have done, say ten years ago. Yes, 0:11:17.840 --> 0:11:20.120 and that people who talked about this they would come 0:11:20.200 --> 0:11:22.160 up with reasons. Oh, the stock market. It's not like 0:11:22.240 --> 0:11:24.920 chess or go, and therefore you can't really do predictions 0:11:24.960 --> 0:11:27.760 the same way. But what you're saying is that with 0:11:27.920 --> 0:11:31.120 these models, which are different than lllums, there is some, 0:11:31.280 --> 0:11:33.640 at least on a short time scale, predictive capacity. 0:11:33.800 --> 0:11:36.680 Yes, I think I find this still to mistake a 0:11:36.679 --> 0:11:38.319 little bit hard to believe. I think you get this 0:11:38.400 --> 0:11:41.240 kind of efficient market hypothesis stuff jumped into your head. 0:11:41.400 --> 0:11:43.320 It seems if someone is saying they can predict like 0:11:43.360 --> 0:11:46.280 the price of a stock in an hour, your instinctual 0:11:46.320 --> 0:11:48.840 reaction is incredulity, Like just sounds like you're kind of 0:11:48.880 --> 0:11:52.240 bluffing or making it up. But no, these models can 0:11:52.320 --> 0:11:53.960 predict this, And I think it's the way to kind 0:11:53.960 --> 0:11:56.640 of reconcile the like really man like kind of distinction 0:11:56.760 --> 0:11:59.640 is that the predictions are very bad in some sense. 0:11:59.679 --> 0:12:02.199 We don't no way to talk about like accuracy. But I 0:12:02.240 --> 0:12:03.520 think the way to think about it is like the 0:12:03.559 --> 0:12:06.800 accuracy is like fifty point one percent type thing, like 0:12:06.840 --> 0:12:09.880 they're only a little bit better than random. 0:12:10.240 --> 0:12:12.800 But I suppose an extra one percent like blows up 0:12:12.840 --> 0:12:14.040 your profits if you're doing it. 0:12:14.720 --> 0:12:16.880 Doing it scale doing it enough times and over time 0:12:16.880 --> 0:12:19.920 you kind of realize the biased coin flip. And as 0:12:19.960 --> 0:12:22.000 for why it might be possible to do this without 0:12:22.080 --> 0:12:26.400 kind of invoking magic, It's like markets are very beautiful 0:12:26.440 --> 0:12:29.439 interaction of like many different potties, all the different kind 0:12:29.440 --> 0:12:32.560 of utilities and risk preferences and things, and the only 0:12:32.559 --> 0:12:34.559 way you really see what people are doing is by 0:12:34.600 --> 0:12:36.520 like the actions they take in markets, and you kind 0:12:36.520 --> 0:12:40.040 of it's sucking up all that like signal, that micro 0:12:40.120 --> 0:12:42.160 signal and extrapolating. 0:12:41.600 --> 0:12:45.600 The synicism or the skepticism about the possibility of machines 0:12:45.640 --> 0:12:48.480 that could predict the price of stocks is a little strange, right, 0:12:48.520 --> 0:12:51.840 because machines ingest data then whatever, maybe they see a 0:12:51.880 --> 0:12:54.760 pattern more likely than not, this consolation of data means 0:12:54.800 --> 0:12:57.200 tomorrow will be green. Humans do this all the time. 0:12:58.040 --> 0:12:59.560 What else do we have besides data? 0:12:59.679 --> 0:12:59.880 Right? 0:13:00.120 --> 0:13:02.800 You have an analyst and they put out of Tesla 0:13:02.920 --> 0:13:04.679 or whatever in video is going to go to five 0:13:04.720 --> 0:13:05.520 hundred dollars a show? 0:13:05.640 --> 0:13:08.600 How dare you insinuate I'm not smarter than a computer show? 0:13:08.800 --> 0:13:11.800 We were like all humans have this data and much 0:13:11.840 --> 0:13:15.160 less data, and yet humans are making predictions all the time. 0:13:15.240 --> 0:13:17.520 They have a whole industry of it, so the idea 0:13:17.559 --> 0:13:20.560 that therefore was for some reason a computer couldn't do 0:13:20.600 --> 0:13:25.000 this with much more data analysts ever, have I understand 0:13:25.000 --> 0:13:27.760 why the cynicism comes off as a little strength. 0:13:28.200 --> 0:13:30.880 I think some of the doubt stems from this idea 0:13:31.080 --> 0:13:34.079 that a lot of these models tend to be backward looking, right, 0:13:34.240 --> 0:13:38.439 and some of them occasionally are pretty bad at spotting 0:13:38.559 --> 0:13:41.800 or reacting to big regime breaks. And I guess the 0:13:41.880 --> 0:13:45.440 thinking again sometimes is that maybe humans are more flexible, 0:13:45.480 --> 0:13:47.959 maybe more adaptive in their thinking, and they can kind 0:13:47.960 --> 0:13:51.560 of spot these big cultural shifts. How do you actually, 0:13:51.840 --> 0:13:54.720 I guess, prepare for those big pattern changes. 0:13:55.000 --> 0:13:58.720 Yeah, I was at HIT for COVID, and I thought 0:13:58.760 --> 0:14:00.840 that was kind of like the most that was a pattern, 0:14:00.880 --> 0:14:04.880 that was a big pattern break, and things went totally fine. Actually, 0:14:04.880 --> 0:14:06.920 it was more of an engineering crisis in some ways. 0:14:07.000 --> 0:14:09.720 Stock market volumes exploded and every system was just like 0:14:09.760 --> 0:14:11.840 screaming trying to keep up with the volume of activity. 0:14:11.880 --> 0:14:15.800 But in terms of the predictions they stayed quite good. 0:14:16.080 --> 0:14:19.400 And I had like Rickoncilus in my head as well. 0:14:19.560 --> 0:14:21.840 I guess it is a matter of like horizon and 0:14:21.880 --> 0:14:25.000 like how far in the future are we talking intra day. 0:14:25.360 --> 0:14:27.640 I think a lot of the price movement is driven 0:14:27.800 --> 0:14:30.840 by just observing, like the flows. It's hard for us 0:14:30.880 --> 0:14:33.720 as humans to observe, but it's like the relative patterns 0:14:33.720 --> 0:14:35.840 of buyers and cells in the markets. And it's like, yes, 0:14:35.920 --> 0:14:38.720 during COVID, volatility was massive and prices were moving up 0:14:38.760 --> 0:14:40.440 and down a lot, but they were going up and 0:14:40.600 --> 0:14:44.000 down during say March twenty twenty, and so these models 0:14:44.040 --> 0:14:45.520 it was sort of out of domain for a human. 0:14:45.560 --> 0:14:47.560 But I don't think out of domain in some sense 0:14:47.560 --> 0:14:48.200 for the models. 0:14:49.200 --> 0:14:51.560 But I guess I also don't know how you would 0:14:51.560 --> 0:14:54.000 apply this thinking if you were trying to make it a. 0:14:53.680 --> 0:14:55.240 Month ahead predictions. 0:14:55.400 --> 0:14:58.280 I often get like people being like, oh, everyone knows 0:14:58.320 --> 0:15:00.800 hedge funds, which were not a hedge fund is like 0:15:00.000 --> 0:15:03.280 like flipping coins, and it's some survivor bias thing. And 0:15:03.720 --> 0:15:06.600 you know, I genuinely don't know about months out prediction 0:15:06.720 --> 0:15:09.400 stuff that is not a data rich environment. 0:15:10.080 --> 0:15:13.240 Just by definition, there have been more days than months, 0:15:13.280 --> 0:15:16.480 so therefore prediction on a day basis you're offered a 0:15:16.520 --> 0:15:17.240 lot more data that. 0:15:17.920 --> 0:15:21.520 The rule of thumb is basically very useful and it 0:15:21.560 --> 0:15:24.280 extends all the way down to seconds, and we see 0:15:24.280 --> 0:15:27.120 that empirically all the time. And so yeah, I guess 0:15:27.240 --> 0:15:28.920 all the things I'm saying, do you have this heavy 0:15:28.960 --> 0:15:30.560 out that it does rely on sudden of it being 0:15:30.600 --> 0:15:33.520 a certain level of signals and noise. I definitely cannot 0:15:33.680 --> 0:15:36.560 make reasonable claims about the price of things in like 0:15:36.600 --> 0:15:39.080 a month using the same kind of like AI hammer. 0:15:39.680 --> 0:15:42.920 I guess also to be specific, I'm talking a lot 0:15:42.960 --> 0:15:45.560 about using market data to make these predictions. And that's 0:15:45.600 --> 0:15:48.280 because on the sort of intra day timescale that is 0:15:48.320 --> 0:15:50.440 the most important thing. It's all about flows and things 0:15:50.480 --> 0:15:52.720 have been back and forth. If you're thinking about things 0:15:52.760 --> 0:15:55.240 in a month's timescale, I think that's fundamentals. 0:15:55.840 --> 0:15:57.880 And can AI be used for that? 0:15:58.200 --> 0:16:01.040 I don't know, to be honest, and it's definitely outside 0:16:01.280 --> 0:16:04.640 my wheelhouse. And I guess people have various opinions about that, 0:16:05.320 --> 0:16:07.120 and maybe some people very much would like to claim 0:16:07.160 --> 0:16:09.840 that they can, and you know, others maybe don't. But 0:16:09.840 --> 0:16:14.120 it's definitely outside of my Arab expertise. And I don't know, Wait, talk. 0:16:13.960 --> 0:16:16.360 To us about the data that you're using, or talk more, 0:16:16.400 --> 0:16:19.440 because this is another area where people tend to talk 0:16:19.560 --> 0:16:23.880 in PR speak, sometimes we have access to all this data, 0:16:24.240 --> 0:16:27.600 unusual data, alternative data, and that's going to enable us 0:16:27.600 --> 0:16:30.200 to use AI better. What are you actually looking at 0:16:30.240 --> 0:16:32.360 and what have you found? I guess most useful? 0:16:32.600 --> 0:16:35.200 Well, I think the thing that I found most counterintuitive 0:16:35.200 --> 0:16:37.560 when I started was that when you're thinking about predicting 0:16:37.600 --> 0:16:40.640 the prices of anything a minute, an hour out, I 0:16:40.760 --> 0:16:43.520 far the most useful thing is just market data. This 0:16:43.640 --> 0:16:46.160 is the market data feeds you can buy from the 0:16:46.240 --> 0:16:49.080 exchanges for a pretty reasonable price. People often think of 0:16:49.160 --> 0:16:51.880 some sort of like competitive moat. The data feeds for 0:16:51.920 --> 0:16:54.480 these exchanges are not particularly high. I mean crypto, you 0:16:54.480 --> 0:16:56.480 know where it's just like a wild West. But everyone 0:16:56.520 --> 0:16:58.880 can collect these feeds, and so that is the most 0:16:58.960 --> 0:17:01.920 useful raw ingredient. That is the most true expression of 0:17:01.920 --> 0:17:04.399 everyone's intense right they're going to the market, that quoting, 0:17:04.400 --> 0:17:07.320 the buying, selling, That is the primary ingredient. People get 0:17:07.359 --> 0:17:08.960 kind of caught up on the whole, like do you 0:17:09.000 --> 0:17:11.720 have a Twitter feed type of thing? And Bloomberg sells 0:17:11.720 --> 0:17:15.760 a Twitter feed through a state of products, and it's 0:17:15.840 --> 0:17:18.920 every now and then obviously something happens, news happens during 0:17:19.000 --> 0:17:21.800 market hours that moves, the price justicates the price. But 0:17:21.840 --> 0:17:24.520 if you really coldly rationalize that that is a relatively 0:17:24.640 --> 0:17:28.560 infrequent thing compared to the overall massive markets. So I 0:17:28.800 --> 0:17:31.800 thinking entry day, I think these market data feeds, it's 0:17:31.840 --> 0:17:34.439 literally like a little events someone quoted that this price 0:17:34.720 --> 0:17:35.600 and this size. 0:17:35.760 --> 0:17:36.760 It's all anonymous. 0:17:36.920 --> 0:17:39.080 Market data feeds are anonymous, and so that is the 0:17:39.200 --> 0:17:41.880 raw stuff, and it is vast. There are just millions 0:17:41.880 --> 0:17:44.160 and millions of events per day, per stock, per future. 0:17:44.960 --> 0:17:47.880 When you get to the day days timescale, that's where 0:17:47.920 --> 0:17:50.399 the alternative data quote unquote kind of really comes in 0:17:50.400 --> 0:17:53.840 as an alternative to market data, the SEC filings, the 0:17:53.920 --> 0:17:57.960 news feeds, balance sheets, brokers reports. 0:17:57.640 --> 0:17:58.159 Things like this. 0:17:58.320 --> 0:18:01.359 That's where that comes in, a vast sea of data 0:18:01.400 --> 0:18:05.239 offerings that people try and sell that I think in 0:18:05.240 --> 0:18:07.120 that kind of situation, you know, it's a very low 0:18:07.200 --> 0:18:10.399 shop environment you start getting into, and it can be 0:18:10.440 --> 0:18:12.840 hard to attribute the extra shop needs of these things. 0:18:13.280 --> 0:18:16.480 But in some sense it's also very democratized where maybe 0:18:16.480 --> 0:18:19.919 people collecting very secret data sets, but my inbox, and 0:18:19.960 --> 0:18:21.439 I'm not even the person in charge of buying these 0:18:21.440 --> 0:18:24.240