- 3 years ago
Kaiju ETF Advisors builds, trains, and employs robust artificial intelligence (AI) and machine learning technologies designed to improve fund management decision-making. By empowering these innovative technologies to curate and provide direct management of our ETF, we’re striving to go places no Registered Investment Advisor has gone before.
Category
🗞
NewsTranscript
00:00 (upbeat music)
00:02 - What is up Zinger Nation?
00:04 Welcome back to another episode of Benzinga Live.
00:06 Happy Tuesday, everybody.
00:08 (upbeat music)
00:10 All right, Mr. Pinnell, how are you doing this morning?
00:16 - Good, how are you doing, Aaron?
00:18 Thanks for having me back.
00:19 - I'm good, we're here in sunny Detroit.
00:22 I understand you're probably a few hours behind.
00:24 It's probably real early for you, huh?
00:26 - I wouldn't say real early.
00:27 It was real early, but yeah, it's a little earlier
00:30 here on the West Coast of Canada right now.
00:32 - Beautiful, I hope you guys,
00:34 are you guys being impacted by the fires over there at all?
00:39 - We've usually got a number of fires burning in the summer.
00:42 This year has been pretty good
00:43 about not dumping the smoke this way.
00:46 We've not had a lot of rain though,
00:48 so it'd be good to see that roll in.
00:51 Not something I would normally say.
00:53 - Okay, well, yeah, we've been following the news here
00:55 'cause it's been a few times
00:56 from the more like mid regions of Canada.
00:58 We've gotten some smoke down here in Michigan from there.
01:01 But Ryan, let's hop into it with Kaiju.
01:04 I know we have the link and everything in the description,
01:06 but can you just give us a brief overview
01:09 of what Kaiju is, what you guys do?
01:11 - Sure, I mean, we originally started
01:15 as a private fund manager,
01:16 which is something that Kaiju Capital Management still does.
01:21 And we were an early adopter of AI.
01:24 So I'm gonna get boring by constantly saying
01:28 we've been doing this long enough
01:29 to have been made fun of for it.
01:30 I think this is the fourth time I've said that,
01:32 but it's true.
01:33 You know, I mean, it's not that long ago,
01:35 like 2019, you tell somebody
01:37 that you make end point investment management decisions
01:41 with AI and it was like, you know,
01:43 you practice the dark arts.
01:44 You got lobsters crawling out of your ears.
01:46 They would just look at you like that was insane.
01:48 And now today that doesn't seem so insane.
01:52 And we moved into the ETF space last year
01:56 after we noticed that, you know,
01:58 the technology just wasn't really showing up
02:00 on the public fund side.
02:01 You know, it was becoming more available.
02:04 I wouldn't say common on the private fund side,
02:06 but on the public fund side,
02:08 there were some early offerings with thematic exposure,
02:13 but nothing that was really harnessing the power of AI
02:15 to make the end point investment management decisions.
02:18 We weren't sure why that was.
02:21 The more we looked into it,
02:22 the more we realized that, you know,
02:24 derivatives manufacturers that focus on ETFs
02:26 just don't have that capability.
02:28 And AI is an extremely heavy lift.
02:31 I mean, it's taken us years and years
02:34 to build the infrastructure that our systems run on.
02:37 That's become easier more recently
02:41 because obviously the technology is more powerful,
02:44 but it takes a long time to set up that kind of system.
02:47 Talent acquisition is another real challenge.
02:50 There's no such thing currently.
02:52 I mean, I'm sure there will be,
02:53 but there's no such thing as like an AI degree, right?
02:55 I mean, you don't see that.
02:57 How do you find experts?
02:58 How do you find experts in a field
02:59 that's only a couple of years old?
03:01 Exactly, right?
03:02 And I mean, you've got, you know,
03:04 our director of AI systems, Nicholas Sabryan,
03:08 like he's got a master's in machine learning from 1994.
03:11 So, you know, like the technology goes back a long way,
03:16 but in terms of its current level of application,
03:20 I know it's relatively recent.
03:21 So what you're finding is, you know,
03:23 derivatives mathematicians, physicists, cryptographers
03:28 who have used AI in their discipline
03:32 and then are coming to, you know,
03:35 financial market management, et cetera,
03:37 bringing that technology with them.
03:39 So a lot of the talent that we pull out of Imperial College
03:43 or University of Zurich,
03:44 like has never worked in financial markets before,
03:47 but arguably they got expertise in AI.
03:50 - Interesting.
03:51 Yeah, that's something you probably don't think about a lot
03:54 for a lot of these AI companies is the recruiting
03:56 and talent acquisition process,
03:58 but I'm sure it's not easy or at least not,
04:01 I don't know if it's easy, you know,
04:03 in a lot of fields,
04:04 but it's probably a little bit more difficult here.
04:07 Ryan, how does AI improve?
04:09 Like how do you optimize,
04:11 use AI to optimize fund management?
04:13 And you guys were doing this before, you know,
04:16 on the, before the public side,
04:17 when you guys were still just being a private fund,
04:19 you got like, how are you guys using AI
04:20 and how was it used to optimize that fund management?
04:23 - That's a good question.
04:25 I think the easiest way to start to answer that
04:30 is to acknowledge that, you know,
04:32 AI is not good at everything in fund management.
04:35 So, you know, answer number one,
04:37 depends on what type of fund you're running.
04:39 We ran quantitative funds from the very beginning.
04:43 So we were never global macro.
04:46 We were not sort of fundamental investment managers.
04:50 So for us, the transition early on
04:53 just was a natural evolution.
04:55 You know, we believed in the profitability
04:58 of identifying specific patterns
05:00 and being able to predate on them.
05:02 And AI is fantastic at that.
05:04 If we were running something
05:07 that was more global macro, fixed income,
05:11 something where a deep knowledge of the, you know,
05:16 financial market response to big socioeconomic changes
05:21 or, you know, interest rate fluctuations
05:24 and quantitative easing or sheet tightening,
05:29 you know, like guys that run those kinds of funds
05:32 have decades of experience
05:34 and a deep, deep well of knowledge
05:37 that they can draw from in terms of how the world
05:39 might react to an event or a new scenario.
05:43 Whereas on our side, we are just picking up the patterns
05:48 that come from those events.
05:50 You know, ultimately any event,
05:52 any change in interest and an underlying equity
05:56 in a market, et cetera,
05:59 is gonna show up in price, time and quantity.
06:01 And that's what we're looking for.
06:02 And AI does that well.
06:03 Like it does the number crunching, the filtering,
06:06 the culling really, really well.
06:09 It's awful at context or nuance,
06:12 which, you know, you see from GPT
06:14 and the echoing challenges, it's a pretty cool mimic.
06:19 And it's extremely powerful in that implementation.
06:22 But there are moments where you're going, what?
06:24 You know, because it's just, it's guessing, right?
06:26 It's guessing what you expect to see next.
06:29 - Yeah, and I think it's, I mean, one thing is when people,
06:32 I think are almost like too hard on the AI
06:34 when it does make it, even on chat GPT,
06:36 if it gives a wrong answer, it makes a mistake.
06:38 'Cause I think people like need to realize
06:40 how these things learn and get better
06:43 is through making mistakes.
06:44 And then like, we're so such in the early stages
06:46 that right now it's clear as day
06:49 that these things are just going to continue
06:51 to get more powerful, more advanced.
06:54 So bringing the AI to like a tangible trading example,
06:58 dip, your ETF, DIP is AI powered
07:03 and it's kind of centered around this idea of buying the dip.
07:06 So how is AI implemented in doing this?
07:09 Is it as simple as like, okay,
07:10 if the stock goes down 20% in the week,
07:13 that's a dip, we're going to buy it
07:14 or what exactly is the AI doing here?
07:16 - Yeah, so I think the important thing
07:22 is to start by identifying what's a dip, right?
07:26 Because you've got, everybody has a different definition
07:29 of what a dip actually is.
07:31 And it's not-- - I get scared
07:32 when I see my portfolio down 1%.
07:34 - Yeah, exactly.
07:36 And you know, there's some people
07:37 who are looking for like any retracement
07:39 of like eight or 10%, that's a dip, I'm going to buy that.
07:42 It's like, yeah, but what if that's
07:44 the appropriate retracement
07:46 from a massively overvalued underlying, right?
07:48 Like, you're buying back into fair market value,
07:51 you're not buying a dip,
07:52 not every retracement is a buying opportunity.
07:55 So, when we talk about dips,
07:57 we're talking about an artificially oversold condition.
08:00 And I continuously forget
08:03 that that is not an obvious definition.
08:05 So, we came up with this analogy
08:07 that made that sort of simpler to understand.
08:09 And it's, so imagine that there's like a basketball
08:12 floating in a swimming pool, right?
08:13 And that's your, at any given time,
08:15 fair market value for an underlying.
08:17 Now, you can add more water to the pool
08:19 and the basketball will go up,
08:21 you can remove water from the pool
08:22 and the basketball will go down.
08:24 Water represents the general consensus
08:27 of the collective of market participants
08:29 that are in that underlying.
08:31 So, that's where it wants to live.
08:33 You know, it's not gonna stay there,
08:35 but it requires the water in the pool up or down
08:38 to sort of maintain an even level.
08:39 Now, if I come and I put my hand on the ball,
08:41 I stuff it underwater,
08:44 it's gonna be under the surface of the water,
08:46 but only so long as I continually apply that pressure to it,
08:49 which I can't do forever.
08:50 At some point, I'm gonna let go,
08:52 it's gonna pop back to the surface of the water.
08:55 The same works the other way around.
08:56 I can lift it out of the water,
08:57 but I'm not gonna hold it there forever.
08:58 Eventually, I'm gonna drop it,
09:00 it's gonna fall back onto the surface of the water.
09:02 So, what we're looking for
09:04 is the hand on the basketball under the water.
09:06 This artificially oversold condition
09:08 that can occur with HFT downside overflow,
09:12 liquidity pockets and voids being predated on,
09:15 it's a temporary collapse in share price
09:18 that's not really related to the consensus
09:21 on fair market value.
09:23 So, we are looking for that pattern,
09:25 which is fairly reliable.
09:27 We identify it, we buy it.
09:29 When the ball pops back up to the surface,
09:32 that's when we're gone.
09:33 So, this isn't timing the market.
09:35 AI is awful at long-term investing.
09:38 It has no idea what's coming in three months,
09:41 but it's pretty good in a short-term timeframe
09:43 because those patterns,
09:44 the integrity of the patterns hold up.
09:46 So, that's exactly what it's looking for.
09:48 It's taking largely S&P components and huge index ETFs.
09:53 It's identifying this temporary oversold condition,
09:59 buying ball pops back up, exit the position,
10:02 wash, rinse, repeat, rebalance daily.
10:04 - Got it.
10:06 So, how much in this fund is actually being managed by AI?
10:10 Like, is the AI the one executing the trades?
10:13 Is it giving signals?
10:14 And then humans are the ones executing the trades.
10:16 Walk me through kind of exactly what's happening there.
10:19 - So, the AI system, I mean, and when we say the AI,
10:22 we refer to it as the AI system or the machine or whatever.
10:25 You're talking about like thousands of independent neural
10:29 nets with discrete jobs, right?
10:31 I mean, we just say the machine or the AI to make it easy,
10:35 but you've got regime classification engines
10:39 that are trying to determine what regime
10:42 this underlying, the sector,
10:44 the broad market is generally in.
10:46 There's several of those.
10:47 We run a regime change detection engine,
10:49 which is constantly trying to determine based on what it
10:52 sees right now, what does it think is coming?
10:55 And then obviously on every single underlying,
10:58 we run discrete engines on multiple levels,
11:01 both looking for that artificially oversold condition,
11:04 and then sort of the checks on, you know,
11:07 is this a regime that's favorable to this pattern?
11:10 Should I pick this?
11:11 So, the AI is doing all of the curation
11:15 and the trade direction.
11:16 It does not do the actual trading.
11:18 That goes off to a portfolio management team,
11:22 and that's executed manually.
11:23 Now, that said, you know,
11:25 there are layers of quality assurance there
11:28 where we're making sure that there's no obvious error,
11:32 but, you know, you're talking about a system
11:34 that every day takes 220 terabytes of data
11:37 and performs 2 billion discrete examinations.
11:41 In terms of a human's ability to say,
11:43 gee, I don't know, maybe the waiting there
11:46 just seems a little off to me.
11:48 There's no chance you're going to outperform the machine
11:50 when it comes to those determinations.
11:53 So, yeah, the AI does the curation and direction.
11:57 We make sure that it doesn't contain
11:59 some sort of horrible critical fault,
12:02 you know, some sort of technological breakdown,
12:04 and a PM team executes.
12:06 And when we say, you know,
12:06 it does this at a nanosecond scale,
12:08 it's the analysis that's performed at that scale,
12:11 not the trading.
12:12 We're not a high-frequency trading firm.
12:14 - Understood.
12:16 So, right now, I mean, I mentioned this
12:19 on the top of the show, but on Wall Street this year,
12:22 in investing, AI has been kind of the story,
12:25 the biggest thing.
12:26 It's been driving, you know,
12:27 a lot of the stock market moves
12:29 that you've seen in stocks like Nvidia.
12:32 So, you have a lot of people on CNBC and stuff now
12:35 that are financial experts that are talking about AI,
12:38 and I don't know, you know,
12:40 how many of them really understand exactly.
12:42 Myself, I don't understand a lot when it comes to AI.
12:44 What are the biggest misconceptions out there right now
12:48 when it comes to AI?
12:49 - Well, I think there's a couple.
12:52 One, the doom and gloom scenario, right?
12:54 You know, there's a lot of disquiet around it.
12:57 Is it gonna end all these jobs?
12:59 Is it gonna take over?
13:00 Is it Skynet?
13:01 Are we all getting nuked at the end of this?
13:03 And, you know, it's understandable
13:06 because it can be a bit of a creepy technology.
13:09 You know, we once fed an output mechanism
13:13 with three years of our trader chat
13:14 so that the AI could talk to us more effectively.
13:17 This was sort of an earlier generative version of chat GPT,
13:20 and it was eerie, you know,
13:22 and we eventually turned that off
13:23 'cause it was a little odd,
13:25 and it turned out we were pretty rude.
13:27 So, so was the AI that was talking to us.
13:30 But, you know, when you work in this field,
13:34 you realize, or you obviously know,
13:36 that it can't do anything
13:38 that you don't tell it specifically to do.
13:41 So, if you had a late-stage version of GPT, GPT-5,
13:46 and it could actually talk to you
13:48 instead of just outputting text,
13:50 there's no capacity for it to suddenly,
13:52 in the middle of that exercise,
13:54 start speaking a foreign language,
13:56 because it felt like it.
13:57 I mean, you might, in this interview,
13:58 you might absolutely switch to some other language
14:02 just for the hell of it.
14:04 But, you know, you're a human,
14:05 and you have complete autonomy,
14:07 and you could do whatever you want,
14:08 but the AI would need to be told,
14:11 first of all, it wouldn't know how to speak the other language,
14:13 it would need to be told that it can speak another language,
14:16 and it would need to receive some sort of rule set
14:19 on when it might want to mimic a random condition
14:23 and speak the other language.
14:25 Like, unless you program that all in, that can't happen.
14:28 So, it can't blow up the planet
14:30 unless you give it permission to blow up the planet.
14:33 In the investing space, you know,
14:35 I think the fear is that these AI systems are all black box.
14:39 Black box meaning that there are no rules and guardrails,
14:42 that the AI system is simply allocated a huge pile of money,
14:45 and it trades whatever it wants
14:47 at whatever frequency it wants.
14:50 And those firms, like Renaissance Technologies,
14:52 been doing that for 30 years,
14:54 and that's why they have a 67% average net of all fees
14:58 for all three decades,
14:59 because black box systems are hugely powerful.
15:02 But they can also, like, have 70% drawdowns.
15:05 So, you know, if you're a responsible manager using AI,
15:10 you're not gonna run a black box system.
15:11 You can put rules, guidelines, controls in.
15:14 DIP has dozens of them.
15:16 Like, it can't just do whatever it wants.
15:18 It has specific sector weighting guidelines.
15:21 It can't unbalance risk to one side.
15:23 So, one side of the AI question is, you know,
15:26 the terror around it just running out of control.
15:30 And I don't see that really happening
15:32 unless we're irresponsible
15:34 and give it that kind of latitude.
15:37 You know, we just take off all of the guardrails,
15:40 remove the handcuffs and say, yeah, no, complete trust.
15:43 Go for it.
15:44 But I don't see people doing that right now.
15:46 I mean, if it's really profitable,
15:48 history will tell us that they will do that,
15:50 'cause I think you've seen before, we saw that in crypto.
15:53 But that's, you know, the biggest misconception,
15:57 I think, is that.
15:59 The other is that it can do everything, which we've covered.
16:01 It just, it's not great at everything.
16:04 - Right.
16:05 And I think, you know, it's interesting seeing Sam Altman,
16:07 of course, the CEO of OpenAI,
16:11 basically like begging Congress for more regulation
16:13 in certain regards.
16:15 And I think, you know, seeing that, okay,
16:17 if used properly, it's gonna be, you know, great.
16:20 And like you said, if as long as certain people
16:22 aren't just trying to profit, you know,
16:24 do things that aren't great with it,
16:27 then I think that's why Altman was kind of like,
16:29 yeah, maybe some guardrails on here would be good.
16:32 So, Ryan, what right now, I mean, how do you think,
16:36 we're seeing how you guys are using AI right now
16:39 in terms of fund management,
16:40 but let's look down the road 10 years,
16:42 assuming AI, you know, the AI networks improve
16:46 in a lot of these manners in financial markets,
16:49 how do you think it will be implemented?
16:51 I mean, you know, like you said,
16:53 we've seen quant trading now for 20, you know, plus years.
16:56 What do you think about 10 years down the road?
16:58 - I think 10 years is like, you know,
17:01 an ice age in AI time, right?
17:03 I think you're like, you could speculate
17:06 wildly in the three to five year timeframe, you know,
17:09 10 years, we could all be just driving around
17:11 in AI directed cars, I suppose.
17:13 But in the shorter timeframe, I mean,
17:17 the exciting thing for us is that the infrastructure
17:20 is becoming more available to the average person, right?
17:25 It's still horribly expensive to build one of these things
17:28 and cumbersome to run an AI shop,
17:31 but I'm hopeful that you'll start to see some tools
17:34 trickle down to the retail level,
17:36 where, you know, you can backtest your own strategies.
17:41 Like you have an idea, you've had success
17:44 trading in a certain way or managing a portfolio
17:47 in a certain way, and you wanna see
17:49 what that would have looked like, you know,
17:51 10, 15 years, you know, that kind of a timeframe.
17:54 And I mean, there are tools to quantitatively do that,
17:57 but quantitative backtests are garbage, you know,
18:00 they're just fallible on every level
18:02 because the testing mechanism has the advantage
18:07 of the future knowledge.
18:08 And I think what people don't know is with AI systems,
18:11 you know, when you're testing these models,
18:13 you know, you train the model,
18:15 you rewind the stock market 15 years, 10 years,
18:17 whatever you're gonna do.
18:19 You basically give it amnesia,
18:21 and then it's only left with the decision-making capability.
18:24 So you play the market forward,
18:27 I mean, we play it every day at the tick level.
18:29 So it's looking at every print of every stock
18:32 that's ever traded and making determinations
18:36 as if it was in real time.
18:38 That's a vastly more robust way of testing something.
18:41 I'd love to see that come to the retail space.
18:44 My fear, of course, is that you're probably first gonna see
18:47 like AI robo advisors, I'm so thankful when they died out,
18:51 and now I think you're gonna see them come back
18:54 and be like, "Hey, just, you know,
18:56 "throw a hundred grand into this
18:58 "and let our AI manage your portfolio."
19:01 And, you know, again, it's not good at that.
19:03 It doesn't know you,
19:04 it doesn't know what your risk appetites are,
19:07 diversification is your job.
19:09 I think it's great if you can, you know,
19:12 invest in discrete funds
19:14 that employ a strategy successfully that you like,
19:17 fills a niche part of your portfolio,
19:19 but handing over the keys completely and then walking away,
19:24 I don't think it's ever going to do that well
19:26 in the short term.
19:28 I mean, I'm curious what you see on the news side
19:32 from retail, like what would you say is the general outlook
19:37 or fear or interest right now when it comes to AI?
19:40 - Honestly, I think from a lot of what I've seen,
19:43 and this is more on like social media
19:45 compared to the actual news,
19:48 I've seen some, you know, people basically being like,
19:50 "Oh, you're not hearing about chat GPT as much anymore."
19:53 So I think like a lot of these things
19:55 are kind of like behind the scenes,
19:56 companies are starting to figure out,
19:58 "Okay, how are we going to use this?
20:00 This is super powerful."
20:01 But I think just on the consumer facing side right now,
20:04 actually I think a little bit of buzz has died down,
20:07 at least in regards to like chat GPT specifically,
20:09 'cause that's the one that everyone got taken by storm by,
20:12 I don't know when this was,
20:13 back in like January, February, five months ago or so.
20:16 And so, you know, I think people are really starting to,
20:22 you know, from the investing side,
20:25 it's still pretty much all anyone can talk about.
20:27 You have Nvidia, you have C3 AI,
20:29 you have all these different publicly traded companies
20:31 that people have been buying up.
20:34 Upstart has been on a tear,
20:36 which is using AI to try to basically qualify people
20:40 for personal loans and whatnot.
20:41 But I mean, it's gonna be fascinating
20:44 because I think like chat GPT,
20:46 every time we see one of these new iterations come out
20:49 of something AI related that shows its strength
20:53 and shows how powerful it can be,
20:56 people will just continue to be amazed by it
20:58 until it's ubiquitous, like you said.
21:01 - Yeah, I think, you know, the fade with GPT
21:04 is the same that we've seen
21:05 with the mathematical side of things.
21:08 It's again, it's really good at some things
21:11 and GPT is sort of shockingly good at mimicry, right?
21:16 And, you know, for something like a university paper
21:20 on Napoleon's failure at Waterloo,
21:23 I think it does a very credible job
21:25 with something like that.
21:26 But for something that's really nuanced,
21:28 a heartfelt letter to a loved one or something,
21:31 you know, if you've played with it,
21:33 it's not great at that.
21:34 Like the prompt education is really important.
21:37 It needs you to provide more and more guidance.
21:42 But I think when the flash of, you know,
21:44 what it could do, you know, sort of faded
21:47 and it's like, well, when am I gonna use this
21:49 in day to day life?
21:50 Some people we know use it extensively,
21:53 you know, in day to day life.
21:54 Our dev team uses it constantly.
21:57 It's pretty fantastic at writing code
21:59 and doing coding lifting. - Yeah.
22:01 - That they can then look over after the fact.
22:04 So you keep the experts-- - Tweak it from there.
22:06 - Right, right, exactly.
22:08 So, you know, same thing that we see,
22:10 you know, you sort of push it, push it, push it, push it,
22:12 and then it blows something up and you think,
22:14 oh, too far right now.
22:15 So I think that's-- - Yeah, I think people got,
22:17 I think people got swept away
22:19 because it was better at Google than a lot of things.
22:22 But like in my day to day, I'll use ChatGPT
22:25 instead of Google for maybe like one or two things,
22:27 but still 98% of the time, I'll just Google something,
22:32 you know, so it wasn't like this overnight takeover.
22:34 People on Twitter and stuff are, oh my God,
22:36 Google's in so much trouble.
22:37 And it's like, I mean, yeah, like you're gonna use
22:40 ChatGPT for a few of these tasks
22:42 that you can't just use Google for,
22:44 but for just searching like, you know,
22:46 how late is this restaurant open?
22:48 Google still does a pretty good job of that.
22:50 So, well, Ryan, I know we've gone over,
22:52 but I think, you know, for now,
22:54 this has been a great introduction.
22:56 I know we're gonna keep doing this
22:57 and probably get a little bit more
22:59 into the weeds of certain things,
23:00 but this has been a great introduction to Kaiju.
23:02 Thank you so much for joining Benzinga Live today.
23:04 I'm looking forward to our future conversations.
23:07 - Absolutely, thanks for having me on again.
23:09 It's been fun.