051 - Samir Varma - Classify Risk Don't Chase Alpha

People viewing this are going to want to consider what they can do to make their own trading better, otherwise what's the point of watching the podcast. There are several things that they really should be thinking about very carefully. So the first is the concept of robust statistics. So the question is, the thing that you're looking at, the thing that you're targeting in terms of your performance, is that something that with a small change in the inputs will still produce the same output or not? So one of the things you need to be doing is thinking about what's robust in your system and what's not robust in your system. So you see people often optimizing their systems and saying, okay, this is the optimal set of parameters. That's really the wrong question. The right question is, is there a bucket of parameters such that the result is more or less indifferent to the change in the parameters? And the second bit of robustness. Welcome to the algorithmic advantage. We're here to expand the toolkit of the Quant Trading Community and introduce investors to the many advantages of systematic trading. Our goal is to educate and inspire as we embark on a captivating journey into the vast knowledge and experience of leading portfolio managers and other experts in the field. Is supposing that you've got three or four inputs or more, it doesn't matter into your system. If you add noise to the inputs, does the rate of return of your system change or remain the same? What you want is to see that as you add noise, the results gradually degrade. Yeah. What you don't want to see is that as you add noise that they vary all over the place, that's terrible. And you certainly don't want to see, you know, your returns start to go up as you add noise. Because that strongly suggests that you fit in noise to begin with. Those are sort of the two major principles that I'll keep in mind all the time. And then the third and biggest principle is that all too often, people worry a lot about their percentage of correct market calls. You know, I got 80% of my calls right. I got 90% of my calls right. But the truth is that your overall compounded rate of return over any significant period depends only on outliers. So negative outliers would lose you money and positive outliers that make you money. And so you always want to set up your system not worrying so much about, you know, what your percentage of positive trades is. You want to set it up so that when you lose money, you lose smaller amounts than when you make it. And you want when you can to make significant amounts of money and when you lose it to make lose insignificant amounts of money. Now obviously that's easier said than done. I do suggest that you use trend following and so on. But it doesn't mean that you have to follow trends. But if you're going to be a counter-trend trader, for example, be aware of what the trend is and also be aware of the fact that it might hand you your head. And you need to be very careful about that. I love it. And this just made me think of an article that you had on Sub-Sac about efficient markets. And given that they are perhaps frightfully efficient, but can never really be completely efficient or aren't completely efficient. Those outliers, the little wins, the little wins is the wrong word, but the less common outliers slash opportunities need to be capitalized on. Yes, that's right. I think this George Soros who said that when you're right, be a pig. And that's sort of what he was getting at is that, you know, the biggest and most important thing in the long run in what your rate of return is going to be is your risk control. And again, all too often I see people that want to create trading systems that are in it because they want to be excited or they want to be interested or whatever. I want to be bored. I want to be as bored as I can be so that I can be fiddling and looking away on my screen and not even bothering with what's going on because I'm just bored. The more boring things are the happier I am. The last thing I want in my trading life is excitement. And I think, you know, there's a there's a great quote in those market wizard's books, which I'm sure you've read by Jack Schwager. I forget who it is, but it's one of the great traders. And he said, everybody gets from the market what they want. Yeah, I remember that. And that's true. If you're looking for excitement, you'll get it, but it doesn't mean you'll make any money. All right, well, I just wanted to jump back then. Let's go high level for a second, so, you know, just the fact that you manage funds, you've written books, you've co edited a quantum gravity textbook, you've written your own book on free will, you know, physics papers, and the articles and the lot, is there a unifying question driving all of it? Yes, good question. The unifying question is why? And, you know, I, my undergraduate degree isn't even in physics. It's actually an engineering and electrical engineering. And that was a very frustrating experience for me because in engineering, the goal is on how that's the question, how do you do X? Well, I mean, look, that's interesting, but that's not my mindset. My mindset is why? And so in every one of these things, the why question, you know, what is free will? Why do we think we have it in a deterministic universe? That's what the book is about. Why do economists, they've started to do it less, but they still pretty much keep saying markets are efficient. Nobody can beat them. Why do they say that? I'd like to know. You know, the same question is why was what I call the world's most capitalist culture, which I think is pretty true? I mean, it has a goddess of wealth, for God's sake, that everyone prays to. Why was it a basket case from 1947 to 1991? That's what my sub-stack is about. You know, India was an absolute basket case, economic disgrace, actually. And I grew up in it. So I again, why? So the India series on the sub-stack is to try to answer why this disaster occurred. And so on. So yes, the unifying thing always for me is why? I like to know why. Curious. So that curiosity mindset is essential even as a systematic trader. Completely. Well, tell me a little bit about your background, your journey into systematic trading, because you've also written that your trading edge must be 100% congruent with your personality. Otherwise, you'll capitulate it exactly the wrong moments. And everything will go terribly wrong for you. So how do you diagnose that fit before you start losing money if you've got a proposal on that? Well, so the short answer is basically you can't. You go to college and pay lots of tuition to become a doctor. You trade the market to pay lots of tuition to become a trader. Having said that, there are things that you can do to minimize the amount of tuition you're going to pay. And so the way you make something congruent with your personality is that I think I even said this in my piece. You test it and you test it again and you keep testing it and then you throw up and you tested some more and you throw up again and you keep testing it until there's just simply no more tests that you can think of. And so that's where I came up with all of these tests that I was telling you before, which is add some noise to the system. See if the parameter combinations are robust. See if the model that you're trying to work with works across asset classes, for example. You know, try to make sure that you can, excuse me, you can say what your edges in some sensible way. Ideally, if you can, this is not always possible. Ideally, if you can say, look, academic finance says ABCDEF. But in fact, C and D are wrong and that's my edge. And that, now you're in fantastic shape. So there are things that you can do to try to reduce the amount of tuition you have to pay. But it is true that you have to pay the tuition because it's the tuition that teaches you what is not congruent with your personality. And you don't necessarily know that hip-hop. I like this comment of yours. Your edge must be yours, not borrowed, not copied, not theoretically optimal for some abstract investor, but yours. And, yes, matching all of that, marrying all of that together. I guess all that testing, working out what works for you. But also what you're comfortable with understanding. I want to get into the understanding a bit later because there's some questions around logic versus data mining and so on that I want to ask about later. Just because there is such a limited edge in the market, you want to be very comfortable with what you're doing because you've tested the heck out of it and you've tested till you've thrown up so to speak, but that gives you that confidence that you know what you're doing. Yep. And also, want to be sure, you know, there's that old joke that if you're sitting at a poker table and you don't know who the sucker is, you're it. Right. So you have the same issue in trading, you know, whose target are you? You should know that because that helps you structure your trading in such a way that you don't get hammered too badly. And then you also have to understand if you get discouraged, you know, for example, in 2022, which is one of my crappiest years, actually didn't do anything wrong. It's just just worked as I say, by the way, that we got a website nine times in a row. Now, was I getting my hair out? Of course I was. Was I annoyed and upset? Yes. But did I for a second thing that I wasn't going to follow my own signals? No. That's what I'm getting at. So I knew from the beginning, having tested it, as I said, agnasium that this could happen. Was I happy about it, of course not. Did I lose my sleep? Absolutely. But that's the life of a trader, which you have to accept it and move on as it were. But not tested it, you can't accept it. That's the issue. Yeah. No, 100%. So just in terms of the adding noise, are there any favorites, any things that you lean on in particular, do you add artificial data to your bootstrapped data, skip trades, add, alter open, open prices and closing prices? Are there any things that come to mind when you say add noise? Yes. Yes. So the what you don't want to do, generally speaking, is change the order of the prices. So for example, people is shuffle days around. Except for a limited thing, which we can talk about at the end, having to do with resampling and bootstrapping, which we can talk about at the end. That's actually not a good idea. Because in fact, more often than not, your edge lies in the fact that stuff is seriously correlated. So you don't want to do that. But what you do want to do is you want to add noise to the prices or whatever your inputs are. So let's say that you're using closing prices, for example. So then take some quantity of standard deviations of the price. I don't know, point one point one standard deviation, a point two standard deviation, something like that. And add that to the price. Just draw it from a random number generator. And then see if the results that you get are still roughly the same. And then start running up that amount point two sigma, point three sigma, point four sigma, point five sigma. See what happens? If it degrades gracefully, you've got a signal. It's a good chance you've got a signal. You never quite know, but this is a good chance you've got a signal. If it doesn't and it bumps around all over the place, yeah, you got a problem. So that's what really good way of checking. And the other is you want to try essentially all the parameter combinations you can to try to find the plateau. And not just look for whatever the maximum is, because quite often the maximum will be, you know, a maximum here, but everything around it is not good. Whereas something else that doesn't quite have the same maximum, but everything around it is really good. That's the one to perform. So and then, you know, related to that is the concept of process. So you always want, if you had to choose, you want a good process with a bad outcome versus a bad process with a good outcome. Because the latter is the way you end up losing money in the long run. Because it reinforces the wrong message. Let's get into that a bit later. Just my mind is racing, though, already on the on the concept of that plateau of parameters. Do you have a thought, perhaps a philosophical one, around the relative value of essentially trying to find strategies that are very robust over time that you expect to endure through time. Versus a sort of walk forward constant optimization to current regimes approach where perhaps whether it's quarterly, monthly, annually, there's a need to review that space, review that plateau, reoptimize, see what's working in a current market. Yes. So my preference is the former. I like strategies that don't have to be changed or reoptimized. The issues with the latter are twofold. The first is you need a much larger stuff, need more resources and so on. Because that's something that you have to keep doing, you have to keep reoptimizing. Yes, you can write computer programs reoptimized, but, you know, if you've got multiple strategies, it's very difficult for one person to keep a track of them. Now, in the last two months, yes, you could probably teach Claude to do it for you. If you're willing to trust Claude, but you know, that's a different issue. I don't much like reoptimization algorithms. And my skepticism comes from the fact that most of the philosophy of those algorithms comes from academic finance. And the problems with academic finance are legion. The biggest problem with it is that they like to assume that things like, for example, stock betas are constant. They're not. They like to think that you can optimize portfolios. You can't. They like to think that, for example, if I've got, you know, 25 stocks and I've got some, you know, correlation matrix or covariance matrix, then I can now, you know, produce an optimal portfolio. You can't, etc. Now, if you make an optimal portfolio and hold it 15 minutes and then do it again, fine, you know, good enough. But as a general rule, in every of those, every one of those cases, if you want to do something robust, you're better off finding a rule that you don't change. So, for example, this is shocking to me, but it's true. So imagine that you've got a portfolio consisting of five stocks and two assets, right? So let's say that you've got five individual stocks, pick whatever your favorites are and two ETFs, one that covers the bond market and one that covers the currency market. Let's say I'm just, I'm making this up. You would say then, okay, well, you know, the stocks in my portfolio should be 50%. And so each stock gets, I don't know, one over one tenth and then the other two are 25% each. Okay, so that's the first set of robustness. You've actually done something relatively intelligent. You haven't tried to over optimize the portfolio by fiddling with its various weights. But here's the most peculiar thing. Even in that stylized scenario going forward, you're almost certainly better off holding one seventh in each position. In other words, you treat a stock and a bond market ETF as the same thing, which is, you know, something that somebody in academic finance would tear their hair out and say you're insane. But that turns out to be the most robust one to do. It's one over n robustness. And essentially, if you've got n things to invest in, put one over n in each one. And don't worry about it. There's nothing else I, in my opinion, that you can do that is robust, that will actually protect you from risk. And you won't get hammered by something that you didn't know anything about. We've got a lot of similar approaches. I want to jump back just quickly. We might have covered some of this in the first show about your background. So because you described your transition into trading from particle physics as something that happened when the super collider was cancelled in 1993. And I just wanted to quickly go back to that. And especially just ask again, if there was what were the things from physics that specifically transferred, what had to be unlearned? And you know, what was that, what was that introduction into trading for you? Well, it was, I actually ended up in the deep end because what happened is that I defended my dissertation in October of 1993. And then applied for jobs, which wouldn't start until the summer of 1994. And so I had a few months to basically do anything I felt like. And I happened to read an article in the economist called the mathematics of markets. And it said that there were people that were now using mathematics to make investments in the stock market. So I said, oh, this is interesting. Let me look at it. And so I started fiddling with it. And I found a particular mathematical technique which we can get into if you like, that seemed to work very nicely to make next day predictions for the S&P 500. And then, you know, well, once the SSC got cancelled, it became clear that I wasn't going to get a physics job. And at the same time, I found that this stuff worked or seemed to anyway. And I said, okay, I'm a trader. And so then I spent the next year and a half trying to find my first client, which I did. And I started trading in I think October. Is it of - Yeah, I think October 1985. I've been a trader ever since. So that's what 31 years now. - So yeah, I'm curious about these initial algorithms as it were. So I believe you started by trying to apply chaos theory to the markets, is that correct? Can you talk to us a bit about that? - Yes, so in chaos theory, you have a concept of something called an attractor. And so what happens is that you have, supposing you've got, let's say three variables that describe your system. So then you can plot the motion of those variables in a three dimensional space, you know, x, y, z. And what you find in chaos theory is that frequently the system will settle into some region of those three axes and not everywhere. Well, that's an attractor. And attractors have different shapes and whatever it is. But one of the things is that because, because there is an attractor, it means that you can make short term predictions. Because what happens in chaos theory is that as you follow two infinitesimally close trajectories forward, they'll start to diverge. But it takes some time to diverge, before they diverge and eventually they end up in completely different places. But while you, if you do the prediction just, you know, one step forward, you have a good chance of making prediction that makes some sense. And so what I discovered, this is no longer true, so I can talk about it. What I discovered is that if you took the prices, the actual prices, the index values of the S&P 500, and you embedded them, this is a way of doing mathematics, but you basically embedded them in a space of end dimensions. You could make one day ahead predictions of the price, accurately enough, such that you could trade it and make money from it. And so the best of my knowledge, I could be wrong. I believe I was the first guy to trade the S&P 500 futures algorithmically. - Well, I guess Jim Simon's was, when you read his book, that you get an impression that I would also doing something similar in those short term predictions. - Yes, although they were looking for signals, which is not what I was doing back then. They were looking, they were trying to do use to find lots and lots and lots of predictive signals, and then put them into a portfolio. And then if you know each signal has a 50.1% chance of succeeding, that doesn't matter, 'cause I've got thousands of these signals, and then I use them. And that's a fantastic approach, of course, as we've seen. There's, I have no complaints about it whatsoever. I wasn't doing that. I was trying to find a market, S&P 500, which by the way, then I expanded to the food see in the DAX and the NICA, I think, yes, that's right. And using this technique of embedding the prices and then making next-day head predictions. - Mm. If, even to this day, perhaps you were using algorithms that are inspired by a quantum field theory, like what does that mean in practice? Is there, is it about the mathematical structure? Is it something more literal that you can do? - No, it's a good question. The, it's not that the algorithm is derived from quantum field theory. It is that the rigor of physics teaches you that whatever it is that you're writing down needs to have some relationship to the real world. So you don't get to do, as frequently economists do, to make up a very idealized model of the world and say, well, in this idealized model, ABCDEF holds. That's fine if you're, if you're trying to sort of put together any economic theory, but if you try to use that in the market, you're gonna get killed because at some point it won't work. As gold when it's covered in 2008, which is a funny story. But anyway, so what you really need to do there is you need to understand what the market is saying as it were so that you can then model that. That's the question. What is the, what is the market saying overall that you're trying to model? As opposed to saying, hey, I have this beautiful mathematical theory based on the following assumptions and it says ABCDEFGH. That's how economists tend to do it. That doesn't really work for trading purposes. Well, that leads me into the article of yours that I just skimmed over this morning with the Emperor Hasno Alpha, which was a really great read. So maybe this is a nice transition into, let's just talk about the strategies that you trade, the basics, the markets, the time frames and so on. What's of interest to you now and why? But just to kick that off, in the Emperor Hasno Alpha, the post sort of essentially shows that data mining matches peer reviewed academic papers as predictors and if that's true, then what is it that say a 30-year quant could have that a GPU cluster doesn't, that is going to give them some real edge in the markets. And you can go back and explain that paper a little bit if you like as well. Yes, so the paper was very amusing. So what they did is they took a whole bunch of academic finance theories and they said, okay, over the last, I forget what a lot of years, more than 30, how have these strategies held up and they looked at values and factors and growth factors and so on and so forth. And then they said, well, let's just say that back then I didn't know anything at all. And I just looked at the data and I said, fine, I'm going to mine this data for whatever patterns exist in it and I'm just going to trade those patterns. End of discussion without asking anything about what they are or what they're not. The worst case was they ended up matching the academic finance, more often than not they beat it. Which tells you that what's going on in the market because it's a very complex system cannot be reduced to just a few factors and despite the fact that Eugene Farmer has a Nobel Prize, he's wrong. And I think the one exception to some extent was good old momentum. - Momentum, yes, that's right. Momentum is always the exception. And it's the exception in every market, everywhere in the world and that's why trend following works. That's the other thing, let's see it on neo-fight trader and you want to get something that will work that is unlikely to hand you your head and you can get your feet wet. - Well, become a trend follower and don't use too much leverage. - Yeah, exactly. I think it was that paper as well, actually no, it was a different paper about market efficiency and article about market efficiency that I also enjoyed reading. It relates, I guess, like in so many ways the markets are terribly efficient, but it's, I guess evident that they are not by the amount of work that gets put into attempting to beat them and that there are, of course, pockets of success which drive further work. And so there is that work that's being done. So in light of that, Emperor has no alpha article. Yeah, are there a few, I guess, insights into then where we're looking for alpha? So the one place you can always find alpha and probably always will be able to find alpha is in momentum. And you can come up with any kind of momentum you like and it does a good chance at the work. So one form of momentum is how the moving average. If something is above a moving average is different than when something's below a moving average. But there's other things you can do. So for example, you can do what's called factor momentum. So one of the things that people like to do, which is just insane in my opinion, you never mind, is they like to break the stock market up into a collection of factors. And so you basically say this stock is a value stock with a medium market cap and some medium trading volume, something like that. But this one is a large cap stock and it's in the tech industry, whatever. So you call all of these things factors and then you say, well, the return of the market is just a collection of factors. Now, that's a technology in some sense. You can always break anything up into a set of factors. That isn't the question. The question is whether trading with those factors is ever going to produce for you a good rate of return. And so what ends up happening is that this turns into, particularly for stock investors, like for example, AQR, which is a large hedge fund, this turns into religion. So for example, they'll say, the value factor, I know it hasn't worked in 12 years, doesn't matter. Now it's got to have its best period coming out because it hasn't worked for so long. I mean, look, if a factor hasn't worked for 12 years, then I'm sorry, but the scientist is probably not working. It's probably not the right way of doing things. So what you do is you say, well, I don't really know anything. But let's say that I do break the market up into a set of factors then what can I do with these factors? Well, the answer is look at the momentum of the factor. So which factor is working now invested in that? That's another thing that you can do. do. That's another kind of momentum. And then you can do what's also called cross-sectional momentum. I've given you two kinds of momentum, giving you time series momentum, which is just how is the thing doing in its time series? The factor momentum, look at the factors that are being expressed by the market right now and invest in those the market seems to like, even though the whole approach is silly. But there's a third thing you could do, you could do cross-sectional momentum. And then you take an entire portfolio of similar stuff. So let's say you take all stocks, say, or all stocks in the S&P 500, I don't know, whatever. And then you break them up into, you know, the top, pick a number, 5%, 5% top and top 15%, 20%, 25%, 30. And then you ask, if you break everything up by in 5% buckets, do these buckets change their forward rate of return in any sensible way? And if you find that, yeah, you know, on average, the top bucket tends to outperform all the other buckets, now you found something else. So that's a cross-sectional momentum. And so you can always do that whenever you've got a collection of assets. So that's three easy ways of using momentum. And they're all pretty valid actually. They're always that have a good chance of working in the future. Just while I think about it, there was something mentioned in your article, the emperor has no alpha, about the papers, the academic work and the data mining work, that the real challenge was what was working out of sample. So it just made me think of out of sample testing and the what role that played for you. And I do want to get more into your process later, but the out of sample evidence is the ultimate evidence. So how important is it for you to reserve data out of sample during your testing? Is that part of it or you're just as likely to look at all of the data because there's not enough of it? Bolt. So one advantage of being a greater than 30 a quant is that as you're building a system of thinking when idea or trying to do something, you already know with a great deal of precision actually. If this is likely to work or not likely to work, if it's just pure fitting noise or if it's something that you're doing something, you know, interesting with. And when that happens, you can more or less use all the data. And then in that case, you're basically looking for consistency. You're asking, did it work in all six month periods and all one year periods and all three year periods and overlapping periods and so on. So you can ask questions like that. If you don't know what you're doing, which is where I was 30 years ago, then you absolutely should have a large data set and of it, you should take a significant portion and hold it out of sample. And by holding it out of sample, it means truly until you are finished because one of the problems is you then run your in sample piece and yard sample piece and it doesn't work and you're like, ah, I need to go back to the drawing board. No, that doesn't work. If that didn't work in that case, it's telling you something that you need to take on board, which is this probably isn't going to work at all turn to it. So you have to keep that in mind, but as you gain more experience, you can stop worrying about that quite so much because what you're then looking for is other stuff. You're looking for, as I said, robustness. You're trying to use the robust statistic and so on. See, one of the problems that I see is when people write, for example, forecasting tools. So they write a forecasting tool that's meant for trading and then the forecast will be judged by root mean square error. So the idea there is, you know, ah, I forecast 22, it came out to be 23. Okay, that's a difference of one. Then I forecast 27 and came out to be 26. That's another difference of one. Then I forecast 28 and it came out to be 24. That's a difference of four and so on. And then you take each difference, you square those differences and you take the square root and you call that root mean square error. That only works when the underlying distribution of data is what's called Gaussian or normal. But the entire point of being a trader, your entire reason of that is that the underlying distribution is not normal because if it is, then what are you doing trading? You have to buy all. You cannot use root mean square error. And by the way, that's also why you mustn't ever use a sharp ratio. That's another question. You mustn't use root mean square error to judge your forecasts. What you need to judge, judge them by is your PNL. And the reason is that if your forecasts are directionally correct, but have big errors when huge moves take place, then you're actually better off than the situation where you're directionally wrong when huge moves take place. And so that's the question. So that's why I was saying earlier, your rate of return depends entirely on getting the large moves correctly. Both up and down, both against you and for you. This is the same reason why you mustn't use a sharp ratio. A sharp ratio is absolutely contradictory. Going to that way, we used to talk about that a little more earlier on in the podcast. The please, yeah, remind us of sharp. It's contradictory by definition. You are saying I have a strategy to trade in the market. Then you are saying by definition, there is something inefficient in the market. And inefficient market is overwhelmingly likely not to have a normal distribution. So then why are you using a standard deviation, which is a definition that comes from a normal distribution to talk about your risk in a non-normal situation. It makes no sense to me at all. It's crazy. It's madness. It's a short, it's shorthand, but people have taken that shorthand to mean reality, which is just it's silly. It's like the pillar of the industry to quite a sharp. It's everywhere. Yep. And it's wrong, but whatever. It's good that there are things that are wrong. Yes, that's the point. When I get too irritated, I have to tell myself, if there was nothing wrong, I wouldn't have anything to treat. That's right. All right, so let's just, I guess, cover where you're at now. So where did all of this land for you in the fund? What's your approach and what are your trading models look like today? So I started off very much in the daily prediction game. From that, I became a higher and higher frequency trader. Until in 2005 to 2008, I had a joint venture with a very famous trader, Chicago, or George E. And Joe and I did a lot of high frequency trading together, basically as his prop capital, his capital we were trading. And then there was a certain point when I realized that we were talking about light speed limits. And as a physicist, I said to myself, if we're talking about light speed limits, I don't think I want to be in this game anymore. So then I said, no, I want to be in the game of lower and lower and lower frequency trading. In fact, I want my frequency of trading to be so low that people would not even call me a quant. And that's really where I've ended up. My frequency of trading is, if you average over all the years, is maybe about one trade a year, one and a half trades a year, that's about it. And that's because of this system that I created called risk timing. And that came from the answer to another question you asked earlier, which is do I prefer systems that I don't have to change or do I want to keep re-optimizing my systems? And what I realized is that one of the problems with trying to find alpha directly by picking stocks is that there's only a limited amount of it. Everyone in his uncle is looking for it. It's moving into higher, higher frequency realms. So why wouldn't I just stop looking at alpha? Why don't I try to find something that's permanent? And ask myself, what's permanent? And the answer is risk. Why is that? Because risk, particularly extreme risk, cannot be arbitraged away. So imagine that something happens, there's some piece of bad news, and there's some heavily levered fund that it has to sell because of the piece of bad news. It doesn't want to, but it's levered so it has to. Then the next guy owns very similar stocks. So he says, oh crap, my fund's going down, I need to sell. Then the third guy sees his stocks going down and he has to sell. So even though they may all realize upfront, this is a negative expected value trade, they have no choice, they have to sell. You can't arbitrage that. It can't be made to go away. And so my idea was, why don't I create a fund where when risk is low, and by risk, I don't mean standard deviation and I don't mean volatility, I mean the chance of a large catastrophic drawdown. When risk is low, I want to be in the market, long and levered. And when risk is high, I want to be out of the market, or partly out of the market, just depending on how high this system things, risk is, and in cash. And I don't want to be dealing with complicated lever, hedges. I don't want to deal with, you know, heavily leveraged positions that I hedged with some other heavily heavily leveraged positions and so on and so forth. Because often those don't have to be Texas hedges. I just want to do the simplest possible thing, which is be literally in cash. So nothing bad can happen to you. And that's what the system does is, and you know, it's been running 10 years now. And doing extraordinary well, extraordinarily well. The principle, I guess, to that you talked about in your paper was that alpha is not as predictable as risk. Yes, that's correct. It's not. And the reason it's not as predictable is because imagine that this is the old example, right? Imagine that I think that the stock is going to go up, well, they'll buy it today and it's already gone up, right? So as people start to exploit alpha, it will disappear from the exploitation. Whereas that's not true for risk. Because you can't exploit it. Yeah. So the risk events tend to cluster and the the adage that you know, one should always be in the market or else they'll miss their best days is false. Tell us about that. You mentioned that in the paper. That's a great question. So the old adage is, oh, if you missed the 10 best days of the market, you'd have your return over a long term period or missed the best 20 days. I don't know, something like that. And the idea is to show you that you should always remain invested in the market. The error is that the very best days in the market occur alongside the very worst days in the market. They are not randomly distributed. If you are out of the market during its very best days, you are also out of the market during its very worst days. And since the left hand tail, I eat bad stuff is always bigger in trading than the good stuff, right? Because the purpose of the market is to make as many people as poor as possible. You actually are better off being out of the market when really bad stuff is happening, even when the really good stuff is happening. As long as you have some systematic way of getting back into the market again. So the problem is not so often the fact that people sell it the right time and they'll get out of the market is that they don't have any method by which they'll get back into the market. If you don't have a method, your door. So ultimately your approach, you're really classifying these regimes of risk and then depending on your outlook for risk, you're effectively deploying ETFs potentially levied to capitalize on market come and when risk comes in, you step aside. That's exactly right. That's the idea. And it took me a long time to build this system. Traders crypto is not a wild, unruly market. It's an asset class with 500% more time for price discovery and opportunity. My new course crypto creator's edge is now live through Algo Advantage, one of the top quad pods and academies out there. A full systematic framework, back testing, strategy development, validation, real edges with full code, not hype, links in the description or head to my sub-stack at Recteligence. Now back to the show. It's a very interesting question as to how I got to where I what I am today. I had this idea that you should be trading risk and not alpha in 2003 actually believe it or not. And I kept trying to make it work and it kept not working and I couldn't understand why it wouldn't work because it would work for three, four years or so and then it would blow up and it worked for three, four years or so and then it would blow up. And each time I would go ask people like in Wall Street, hey dude, you know, I'm using all the same risk models that you guys are, you know, with all the fancy names that are one Nobel prizes, you know, Arima and Garch and EGarch and whatever. And it just handed me my head. What's up? And they would always say, always. Don't worry about it. This is just a once in a lifetime event. After the third or fourth time, somebody said to me, it said that to me, I said, dude, what are you talking about? Life them of what? A lab mouse? I mean, come on. So no, they wouldn't work. I tried all the famous models that you can think of. I'm a physicist. This is fairly trivial mathematics. I just programmed them all. I tried them all and they all love. And so the first thing I realized is the problem, the risk models as they are currently used is that when the shit hits the fan, they don't work. Which is literally the one time you need them to work. Yeah. Like VAR, right? Value at risk. It's completely worthless because it tells you what your value at risk is 95% of the time. I don't care. I only care about the 5% of the time, but it doesn't work then. So what I realized after banging my head against the wall for many years, 12 years actually, is that the problem is that people are trying to predict risk. They're trying to say the annualized risk over the next month is, I don't know, 32% or the annualized risk over the next month is 7%. In one case, high, the other case, low. But that's irrelevant. What's relevant is it's high. That's it. And the moment you realize that, you realize the what you need to do is not predict risk. You need to classify risk. And once I realized that, I stopped bothering with all these models. I trashed all of them and as developed my own classification system. And the idea is to take some set of variables that tells you something about risk in the system and then use them to classify the current state of the market. And that state is actually a classifiable, be pretty robust and see has some legs in it because as I said, risk is something you can't arbitrage away. If the, for example, particularly in the old days, if the federal reserve is tightening reserve requirements, well, I mean, I can scream as much as I want from the rooftops that risk is in all that high. If something happens and banks not have very many reserves and the market starts to fall, there's no cushion. It's going to keep falling. Can't arbitrage that away. So once I realized that, I realized I could start a risk classification system. I created the risk classification system. I coded it up. I started trading it and that's what I've been doing ever since. And it's basically been running the same philosophy identically. We've made small changes to it. Has we've gone along as we've learned things and we made it a little bit better. It's actually becoming increasingly difficult to improve it, which is really good. It's fun. Which means we've done something worthwhile. And you know, it's a sort of permanent model as it was. It's not something I have to change. It's not going to change it. Just talking about market efficiency, you're understanding of the markets in general then. We've covered a bit of ground already and I'm just thinking back over some of your articles as well. There's some articles about sort of agnostic theory free papers outperforming the theoretically motivated ones. To what extent should we trust or distrust signals that we can explain? I think you mentioned this, but Renaissance has Peter Brown said that some of the signals that couldn't be explained were sometimes the strongest. So especially in the age of machine learning and AI, what's your theory around signals we can explain versus can't explain? The key there is something that he didn't explain naturally because that's his secret source. But if you think about it logically, you can sort of figure it out. They have thousands of signals at this point. We know this already. Some stronger than others. The problem that you have is that at some point, a signal is not working. Is that because the market has changed so the signal no longer functions? Or is it the case that there's nothing wrong with it is just that the signal is going through a fellow patch right now? You don't know. So the key, the real key of what Renaissance has done, and I don't know how they did it, is they figured out a way of allowing the signals to reweight themselves automatically, presumably by which signals are working now or maybe which are forecasted to work. I don't know. But there has to be some reweighting of those signals as you go along. If you can do that and you can do it successfully, then you can be completely agnostic as to what the source of those signals is. You just basically say, I found a signal, I don't know where it comes from. I'm going to trade. And of course, you need a lot of them, say, a portfolio of signals and so on, but you also have to make sure they don't clash. So you don't want to end up with just the market return because what was the point? That's the secret source. I don't know how they do it, but that is really the key. And so if you want to be an explanation agnostic, then you need to have some way of selecting from the signals that you are getting, particularly when they're not performing. And this, by the way, is the same problem with an autonomous AI trader. which people keep talking about. You have the classic problem for economics, the principal agent problem. The issue is that the AI, it's not its money. The money is owned by somebody else. And so the tendency is going to be, since the AI can't explain itself to you, the tendency is going to be the AI is going through losing patch, OK, you're out, buddy. We're shutting you down. And if the AI is a good trader, presumably, then you probably are going to hit-- you're going to shut that trade it on exactly the wrong point. So in the long run, you'll end up making less money than your back desk would show, because you'll keep shutting you down as time goes on. And that's something that I haven't seen people address. How are they going to handle this? Because it's a complete black box. It's the same problem. What do you do with signals? Is the same as what do you do with an AI that's losing money? I want to come back to your research and development pipeline. I know we covered some really good stuff about really just test and test and test till you can test no more, or test till you throw up. I like that. I do because I feel like sometimes my approach is difficult to really explain in detail, step one, step two, step three. And I realize, as you say that, it's because that's my approach. Just test until you throw up. And you feel like you know that thing intimately. And by then any possible way of disproving it, you've hopefully come across. You've just got to be honest with yourself, like with the early trader who tests data out of sample, as you said before. Once you've done that out of sample test, you've got to treat that as you went live for that year or whatever. There's that, to me, that art of asking the right questions and thinking the right way about it. But you've just got to put in an enormous amount of work. But is there anything I've missed in terms of your process? So like making sure that some things robust, tackling stock market data of which it is highly efficient. But then also we've lent these things about it, volatility clustering. And various other inefficiencies that we're trying to capitalize on, is there a bit of a sort of defined process for you? It's like here's an idea. Is it robust? Take it through these steps one, two, and three. How scientific is that, so to speak? And do you do all of this in Python? Do you use some standard retail software? What is the sort of data analysis that I look at? I started off doing it in C 30 years ago. Then I switched to retail software, which is trade station. Then trade station became a brokerage. And I said, why do I refuse to be beholden to any software provider ever again? And I'm a physicist by training. So I actually rewrote my entire trading package in Mathematica, the high end mathematical modeling package sold by Wolfram Research. And at this point, all my quotas in Mathematica, I have my own completely made. My partner used our own completely custom software, all written in Mathematica to do all of our testing. And it's written to do exactly what I wanted to do. And part of that, if you like, part of that process is that I also read a lot, particularly in economics and in finance. And I was counting the other day, it's been 30 years now. I've been probably reading three or four papers a day now for 30 years. That's one of about 30,000 papers. Give a take. It doesn't mean every word of them, but I read enough of them to understand what they're about looking for ideas. And you end up then finding commonalities amongst those ideas. And because you find commonalities amongst those ideas, like momentum, now you've got a head start in having some confidence that this is going to go. So that's part of the process. And then the last part of the process is actually the way I measure my risk having to do resumpling and bootstrapping. But we can get into that in a separate segment, if you like. That I would say is sort of the key aspect for me is using that because I'm just worried about catastrophic risk. That's all I think about it. I don't think about really anything else. You've got a really contrary and nature and everything that you do. And that's 40 to go polar opposite. What can I do that nobody else is doing, which is-- Yes, fantastic approach. And it's incredible. The returns you've been able to derive out of that. Given that you've got your own software and you can really poke it in product any way you like, one of the advantages of that is that obviously you're not beholden to the constraints that many packages will throw at us. And then when you don't know what those constraints are, you're potentially not even asking the questions of how do I build a different model. Because the option isn't even given to you. So you've got that tool there to what extent are you using AI, if at all, AI and/or machine learning techniques that you can close the loop there with your engine and your ideas. Can you feed data and questions into AI and have it do some analysis for you and throw it back into the system? Are you using it in these practical ways in your trading? Yes, but not in the way you might think. I use it all the time, literally. I use Cloud Code. I use Codex all the time. But what I did is I taught CodeClood and Codex by writing skills for them what my code base looks like. And I also taught them what my trading philosophy is. And I taught them how I want them to think about my ideas in trading and so on. And once I did that-- and of course, it took some time to make it work, but not that long-- what I can do now is I can use Cloud Code or Codex. It doesn't matter as my front end to my own trading system. And so it knows how my trading system works. It knows the functions. I've written it, knows how I've written those functions and how I want them called and all that stuff. So I can say to it, hey, I have this idea. And it will take my code and rewire it so that I can test that idea. So it speeds up my iteration cycle, very substantial. But what I don't use it for is what everybody else is using it for, which is, hey, give me a trade idea. Yeah, so I don't do that. The other thing I don't do is I don't use machine learning. And this is despite the fact that I actually know machine learning reasonably well for other projects, I don't use it deliberately. And the reason I deliberately don't use it is that I don't know how to know when it's not working. So we're right back to that black box problem again. So let's say that there was not that it can happen because everything's backed up in 10 places at this point. But let's say that for some reason, my computer system wasn't working. And I needed to do a trade tomorrow. Well, firstly, how would I know that I need to do a trade tomorrow? Well, because I know the rules I programmed. So I can, if I need to take out a pencil and paper and just work out what the trade tomorrow is supposed to be, or if there is one because most of the time there is one. We couldn't do that with machine learning. You can't do that with a black box. And that's the difference between something that's intelligible and something that's unintelligent. It's fine to use unintelligible things, but then you need some criterion by which you're going to judge when it's going to be on and when it's going to be off. And I've never found one for me personally that works. People will use all kinds of things. They'll generally, everybody falls back on some variant of Kalman filtering. That's the name of the technique. That's the buzzword. You can look it up if you like. I don't find that convincing, so I don't do it. And again, this is the same issue that we started with. It's not congruent to my personality because I know under pressure, if it starts losing money, if I don't know why it's losing money, I'm going to turn it off. Yeah, I think that's the ultimate point is that in those methodologies, if you don't have a real understanding of when it shouldn't be working, you won't know really when to turn it off at that comfort level. Yep. Yeah, so with the AI, I mean, and I love that. I've got a colleague who is doing very similar things that you've just described with the AI working with your backtesting engine. Given that you do talk so much about AI, do you think the time is coming when this whole process could be run by the AI so that it is, in a sense, constantly refining, building, finding, and deploying better and better systems? What does this mean for the market? Market efficiency? Like, where are we going? Big picture with AI and trade. are we in trouble? What is very likely to happen, but again, I'm talking my own books, so you have to discount it. But nevertheless, what is very likely to happen is that as you make these AI agents more and more sophisticated, they're going to discover more and more sources of alpha. I think that it's hard to argue that that's not true. And if that's the case, the most obvious pockets of alpha when it mind away, which means that what you'll be left with is the less and less and less obvious sources of alpha. And the problem with less and less and less sources of alpha is again, back to legibility. I guess they can find it in a, you know, billion dimensional space. But I don't know how you would trust it. I always fall right back to that problem again. Now, one thing that might happen is that AI's may in some not-so-distant future have their own money. Because already you can imagine a scenario in which there's an AI that lives on X and it makes its money by selling subscriptions to its tweets. And let's say it's very good at telling jokes. People find it very funny. So they pay it and they pay it whatever currency you get. Usually I think it's Bitcoin, whatever it doesn't matter. Now that AI can pay for itself because it's paying for its own way on X. It's paying for its servers and everything else. And it may have some money left over and not once to invest it, which it decided it wants to trade in crypto, something like that on its own. That's possible. In that case, the AI is at least legible to itself, even if it's not legible to anybody else. So that's a possibility that when will that happen? I don't know what it's going to happen. But yeah, I think the obvious source of alpha is the first place where the advent of AI is going to make things an issue. And this is probably going to lead to the following. In the back of my book in the Kindle Edition, I actually have presentations on this. But in my opinion, you're going to get a volatility barbell or a risk barbell to be more accurate. Let's call it volatility, if you like. And I suspect what's going to happen is you'll have longer, co-ecent periods and then shorter, bursty periods, but the shorter bursty periods will be even more violent. And the reason for that, in my opinion, is that a lot of the trading that the AI is going to do is going to be very similar to other AI's because they're all combining the same data verse. And if you have a large number of them, they're all going to be making the same trades at the same time in the same instruments. And so that means that when liquidation happens, it's going to be very violent. That's my guess. But the rest of the time, since again, you're mining inefficiencies, you're actually going to compress the volatility barbell. That's my guess. Fascinating. All right, Samir. Well, look, I think we've covered a lot. I've thoroughly enjoyed this. I think it was really great. There's a lot of excellent stuff to chew on there for the listener. So maybe we'll wrap that up and we'll go into the members' only section and talk a bit about your risk-free sampling and using that as a risk measurement tool. Sounds good to me. Happy to do it.