Welcome to NVIDIA's AI podcast. If ever there was an industry ripe for artificial intelligence to take hold, it's finance and investing.
There's almost infinite data to consume to make sense of and then make a decision based on all that intelligence.
Of course, it needs to be the right decision.
In recent years, hedge funds have taken the lead on algorithmic investing or robo-trading as it's sometimes called.
But there's no reason the hedge fund world should have all the good stuff.
And a number of startups are starting to bring that same machine learning investing approach to the rest of us.
Our guest, Gaurav Chakravorty, is the co-founder and head of trading at one such company, Qplum.
Gaurav, welcome. Thanks for having me. I've said in my intro that AI is perfectly suited to the investing world and finance, but is it?
What's your point of view on that? I think it's perfectly suited.
And let me explain why. My background was in machine learning.
Finance was something that I always thought is, let me say, it's not a clean world.
I thought it was not scientific. Right.
There's lots of like, at least in the movies and things we see like people, you know, in rooms making decisions based on intuition or gut or something.
Exactly. And edge often was what I knew that others didn't know.
You can't model that, you just have to wine and dine.
So that was the Wolf of Wall Street sort of impression that I had, but in 2005, in Wharton, I took a few courses.
They were around financial applications of machine learning.
You know, I quit PhD to join a company to use those ideas.
Frankly, it worked very well for me. I was given $10,000 and told, do what you can with it.
And since then, my strategies have made more than $1.4 billion from it.
Wow. Well, let me back up to that first, take this $10,000 and make more money from it.
How do you start in terms of your thinking and applying what you knew as a machine learning PhD candidate?
Where did you start? I approached it from a textbook way that, what do I have to do?
I have to predict what's going to happen to prices.
At the time, the dominant methods in machine learning were linear regression and ridge regression, things of that sort.
So I used that, made a model, I predicted the next 30 seconds, 600 seconds, something around that range.
I did what the model told me. It said buy, then I bought.
And it worked. Yeah, so it worked so well that I joined in May, and by late August, early September, I was trading the government bonds.
I was told it was a mandate. US government bond.
And there I was trading more than any bank, including the Goldman Sachs of the world.
Well, so when you say you were trading more and that your window was 30 seconds to 600 seconds, um, And let's get back to kind of what you built.
What data were you consuming and what were you trying to make sense of to then activate these trades?
Like what was the information that came in and then what was the action that went out?
So we had access to all the data and people before me used some part of it, not all of it.
I took literally every little piece update from the market.
So for instance, every trade, every bid that was posted in the market.
So nowadays it's called high frequency trading.
Back then it didn't have a name. It was just called trading.
But yeah, I used the methods that, same linear regression, ridge regression, I used that on all the data I added what is now called a feature engineering approach, which is that you use your intuition to make features or indicators out of that and feed that into the model.
And did the model, did you train it in advance of like putting this $10,000 to work or was that really the training?
Like, okay, here's $10,000. Let's figure out how to train a model that gets better and better.
No, I was a little bit old school in the sense that I took around Four years of data.
I trained it and put some data aside for cross-validation.
I saw that the results held up. and then the testing period as well.
And then I started trading with it. Pretty much from day one, it started working and that helps a lot.
You know, if you start something and if you start it on the right foot, then it helps a lot.
Well, so you were trading bonds and from the outside, we always like to hear like why bond prices went one direction or another.
What did this model, what did it become really good at?
And I think one of the things about... high frequency trading, or you can call it robo-trading, or we can talk about whether that's a fair term, It's sort of a black box, or at least it seems like it from the outside.
Like, we don't know why... certain decisions are made.
So how did your model make decisions and how do you describe the ups and downs of a market as you see it through this machine learning perspective? system that you built?
I got better at interpreting the model over time.
Initially, it was a black box to me. And initially, it was a model that I used to only based on the results I got in backtest.
I started trading, started with bonds, but within a year or two I was trading every single bond, future, foreign exchange, equities, everything in the world.
That's what led to the starting growth. Slowly, we understood what was the model getting good at and how to turn it in a different direction if I needed to.
And so what was it getting good at? It was getting good at finding very short-term inefficiencies.
So For instance, if Apple stock went up half a percent, And in that time, let's say Google stock didn't do anything, then the model would sort of say this is a good time to buy Google, just because in the last 30 seconds or so, Apple has moved up a bit.
Wow. And that's I mean, I know we're sort of used to it now, but again.
The sort of movie version that we have of trades and Wall Street, I mean, that's not something any human could really ever do unless you just sort of landed on it and it was on your screen.
And you saw it in that 30 second period.
Exactly. The scale of it was what was new in the sense, if I give you the same trade over the last five years and I say, you know, as a human put it on and hold it for the next month or a year or something that's possible.
But we were trading in notional value hundreds of billions of dollars every day in a very flawless manner.
That's why my computer science background honestly helped a lot.
Explain that for us. Did it help a lot because there was...
No emotion tied to it or no kind of preconceived, like this is what investing is.
And like you say, it's 90% whining and dining and kind of gut.
And information that somebody else doesn't have.
So in our case, we were trading on information that everybody else had.
We were just... Somewhat like Google, we were faster at extracting information from data and more accurate.
The models had to be better and faster. The reason I'm referring to my computer science background is now I'm placing millions of trades every day.
I mean, not manually, the computer is placing millions of trades every day.
Any single mistake, I could lose more money than I make in all the other trades.
So I have to do all of this without making any mistakes.
Well, how is that possible? Our track record of applying models over the long term hasn't been good.
I mean, there was the infamous case of the hedge fund up in Boston.
What was it? Long-term capital? Is that what it was called?
Yeah. LTCM. Yes. Long-term capital. Yeah.
And like their model was at one point completely wrong.
So how can you get it right all the time?
And millions of trades a day, like you say.
Yes, you're right. I mean, by the way, when I said no mistakes, I meant systematic mistakes.
I meant like, you know, where you, instead of a price of 100, you give a price of 1.
And you suddenly lose a lot of money. I see.
But not like you could, the model could, you know, sort of pick the wrong direction on something every now, you know.
However often. Yeah. Okay. Yes. Yes. The benefit in trading high frequency is that since you're placing Many trades, the law of large numbers is on your side.
LTCM, for instance, they were hardly wrong, but they were hardly wrong in one trade.
Right. I could make 300,000 mistakes and I could have 400,000 correct trades and end the day with a positive 1 million.
Right. I see. And again, your model gets better and better.
So when you think about high frequency trading today, How does it approach the market?
And you know, there's value investors. Can you dial things differently?
There's, you know, there's hedge funds, obviously.
Can it be a sort of certain type of investor or is it like just the best investor across all kinds of categories?
So one of the things I have talked about recently is that some of these trades that we know be it high frequency trading, be it statistical arbitrage, which was a huge trade earlier, or trend following, which is famous Paul Tudor Jones, 1987, that trade, all of these are today not that good.
They're not making enough money. And there's a reason for that.
Roughly, if you see a pattern, when a trade comes up and starts getting limelight, In about 10 years or so, we see it slowly dying.
1987 to 1997, we see trend following. 97 to 2007, In fact, literally until August 2007 is when we see StatArb, where every quant fund was using StatArb. crashed horribly in August 2007.
High frequency trading again. came into the limelight roughly around 2006, 2007.
Today, honestly, it's not a good trade. Very, very few firms make money from high-frequency trading.
And we see failures almost every day, most recently being Arxis Capital.
Today, the trade that I think is gonna change The world, every fund that wants to make money is adopting is deep learning.
And the cycle, it roughly takes 10 years because it takes a lot of time for the knowledge to spread, for people to invest in it. technology to catch up.
So I think today the competition in value investing or other trades that we refer to is so high that they're not the best today.
Pardon the interruption, but we have a favor to ask.
Leave us a review on iTunes, Google Play Music or whatever your favorite podcast platform of choice is.
It helps more people find us. As always, thanks for listening.
Now back to the good stuff. We started this discussion by talking about edges and that's what you need to have in investing.
Otherwise, you know, everybody has the same information and I guess the performance is the same and you want to outperform everybody. others.
So is the edge now deep learning? And if it is the edge, who's got it?
So I believe that we are finally getting to a point that With deep learning, we can make investing, robo-investing, a science, a utility such that It no longer needs to be a competition about edge or a game or who has more resources.
For instance, what Google did for information retrieval.
Google search isn't perfect. There are people who know better about a subject, but it's good enough.
We can... achieve that using deep learning.
And then who benefits? Because I think I would agree with you.
Google search and kind of access for information has been this incredible global thing.
It's also been a very leveling thing in a good way.
But level playing fields is not how Wall Street works, right?
Good point. Let me tell you who benefits from what we have now is more than $164 trillion is managed by the financial services industry And we pay, as investors, we pay more than a trillion dollars in fees every year.
Trillion dollars. That's a huge number. And yet we are no closer to solving the problem.
In health care and in other areas, We do research to solve a problem.
Granted, it's tough, but we try to solve the problem.
So who benefits from the current status quo is obviously the financial services industry.
Right. Who get their 2% in management or whatever it is every year.
Yeah. Yeah. it's not so hard i mean think and think about what we're doing if we can use deep learning to make a level playing field where investing is a utility, then who benefits is among us.
Many people are just not connected to this process of investing because they don't know if they're going to do it right.
They don't know if they know enough. Right.
Is that the basis for Qplum then? This idea that you can make investing a science?
Yeah, yeah. So I'm carrying on exactly what Dr. Steele told me in FN952 in Wharton, that if we use these methods, then we can make investing a science.
In 1952, It's a big inspiration for me.
And once I've read more, I've learned how others have worked towards it.
Have you heard of Benjamin Graham? Yes, I have.
In 1952, he published a paper and he was not an academic.
He was a portfolio manager. So publishing a paper is rare for him. but he published a paper towards the science of investing where he talks about in future, investing will be by a trustworthy tool, not experts.
And he says that this tool needs to be thorough.
It needs to look at every aspect of the balance sheet.
He was big on balance sheets. And then he says that tool needs to be affordable.
Like it should be fairly priced. And last thing he says is it should know everything that I know.
It shouldn't feel like I can do better, somewhat like Google makes us feel.
Only additional thing that I think we have learned now is that the tool needs to keep learning And that's why I believe deep learning is the right tool.
If it is a level playing field, if deep learning can kind of bring this science to everyone, what happens to markets?
Secondary markets are going to be still functioning exactly the same as today.
Secondary markets are a place where investors with different risk appetites and different things going on They come together and the shares change hands.
I don't think markets necessarily have to be a game. or a competition, they're just a way to transfer ownership from buyer to a seller.
On the other hand, the process of investing will change The process of investing is us seeking out those who know how to invest and investing with them.
And that will change. to us, be it a pension fund or be it a normal investor, we will say that, yeah, there's this tool that's super awesome.
There might be a few buffets out there who can do better, but this is as good as it gets me.
I'm going to use this. Honestly, I can't wait for that day.
My own experience in the investing world is such that you kind of, you know, There's always a background of worry.
Like, oh my God, is this going to really work?
Yeah. And no matter what you do, like at tough times, you'll feel that, you know, I messed up because I don't know enough.
And the fact is you don't have any investor for that matter, doesn't have the time, resources, The reason we are talking about this machine learning and trading and over the last five to six years, we have seen, the industry has seen machine learning based trading doing better than the best portfolio managers.
BlackRock fired its top portfolio managers about four weeks ago.
And you see almost every day, you see articles in the journal about quants and people shifting to quants.
This is gaining acceptance sometimes. Do you think deep learning systems are going to be the portfolio managers of the future?
And is it systems plural or has it become just AC?
Yeah, very good point. I believe it should be a system, but that system has to be collaborative.
I think that, again, Google being a great example as to how we all create information without our web pages, Google cannot give anything.
I believe that it has to be a system, but how we reach there, needs more collaboration from the financial services industry.
Is that a matter of sharing data and having the same data to work from?
When you say cooperation, what do you mean?
Yeah, it's tough. It's not from industry, which makes a trillion dollars every year.
It's going to be tough. So I think the way it would work is somewhat like in other areas where people collaborate on an experiment. is that the deep learning methods that we are working on, if we are more transparent about it, if we sort of have tie-ups and people sort of work together, to improve the state of the art.
I'll give you an example. Google and Facebook, being as competitive as they are, they would collaborate on big data technologies, but on the other hand, in financial service industry.
In my previous job, I was told Hey, the guy next to you, if he asks you what you do, don't say anything.
If he asks you what book you're reading, you don't read it, say anything.
Wow. Yeah. So it's got to change. Yeah. You described this kind of single system and you mentioned Warren Buffett.
Do you have a sense of, you know, what kind of investor this single system will be.
And, you know, is it a long-term investor?
Is it, you know, in the market to, I don't know, make slow, steady gains or, you know, Does it have a personality or is that kind of beside the point?
I think that it needs to do everything. If there's financial markets are very unique in their nature, like you mentioned earlier, it's very competitive.
If the system, if there's any obvious thing the system is not doing, then people will lose trust.
One example of that is, imagine you have a basketball player and he's great, except that If you defend him one way, he's going to miss every time. then this player cannot, like, it's very easy to defend this player.
Right. So when we're talking on the system, it's got to do everything.
It's got to be able to, like, No matter who is it from the bench who comes up and defends him, he's gotta play.
And so how do you, as a person who develops and designs these systems, How do you achieve that and how do you know kind of what you don't know?
You know, I don't, but one thing that is going in my favor is that I have an experience of many different sort of trades.
I've traded high frequency, low frequency, statistical average, medium frequency.
I have a computer science approach to most of the trades that we have done in the past.
On the other hand, I also have seen the way traditional quant funds work, which is very people intensive and they hire PhDs and make indicators, so I know what not to do.
I know that's not what we should be doing.
We should be following the same Facebook, Google sort of approach where we are learning from data and letting the system learn from data.
Finance, at some point, we need to visualize.
And I think that's the sanity check that we try to do is that no matter what the trade comes out, So that it's not a black box.
We try to visualize when it would have worked and why it would have worked.
So that's our approach. You know, Gaurav, you paint this picture of the future, but it's happening now.
And we know that Machine learning and AI and the concepts have been around for decades, since the 50s.
So why is Wall Street and why are you guys... finally digging in and putting your money where your deep learning is these days.
Yeah, so these methods have been around for a while and research labs and stuff have been using them, but innovation, especially in industry, innovation happens in small companies.
Machine learning today is at a stage where through the work of NVIDIA, GPUs, and then you have software like TensorFlow, and even big data technologies, small budgets still achieve a lot.
We can try these things at an affordable cost.
And that was not there before. Right, you had to be a hedge fund with hundreds of millions or billions under management and the people and the resources to try and figure it out.
It's this access to data and technology.
I mean, I think we see that across the board in technology, but especially in the machine learning realm where If the data is available, the tools are available to put that to work and to make it do things and learn things that...
It really would have taken, like you say, either a government or a very large corporation to do it.
Exactly, exactly. Five years ago, I think to do what we're doing today would have taken upwards of, $100, $200 million a year.
Wow. Well, it's fascinating work and we'll...
Keep an eye on how it's working out for you guys at QPlum and more generally, how it's changing the nature of markets and how we all invest.
So Gaurav Chakravorty. co-founder of Qplum.
Thanks for joining us. Thank you for having me.
