Okay, it's my pleasure to introduce a special guest in our studio today, Rossen Roussev, whom I know for 10 years or more probably. He's a quantitative analyst at JPMorgan Chase. A machine learning enthusiast working on applications of machine learning in commodities derivatives, and finance in general. Thank you very much, Rossen, for joining us in the studio. Well, thank you, Igor. I'm really honored and excited frankly to be here. It's a very interesting experience for me. I want to just give a very quick disclaimer that the views that I will present in our discussion are purely mine, just my personal views, and they don't necessarily reflect the views of my employer. Yeah. What I wanted to ask you, first, let's start maybe with your background. You are a physicists like myself, who did not have much of formal training in finance, and then, basically you were involved more in self-education, self-training. How can you share your experience? How did it work? What you can advice on those who want to follow similar way? Sure. I did take one course on financial math in my last year of grad school, also strictly speaking at that. But, in that course, that's where I heard about Black-Scholes stochastic processes, we did those things. Really, I would say what gave me the depth to be a quant, was a good deal of self-study both before I started work and after. Most of my learning, frankly, was done on the job, and that turned out well, at least for me. The real prerequisites for being a quant in my mind are a scientific training, or at least a scientific mindset or an engineering mindset, and I will put an emphasis on this, it is programming. Yeah. Back when I was starting as a quants, my academic work had exposed me to some C++ on UNIX-like platforms. I was also curious about systems development, and that had little to do with my research. I'll give you an example. I like to tinker with the Bash shell, various scripts, and so on. Then, I was doing that way more than was remotely necessary for my thesis. Then, I was pleasantly surprised that these things were actually in demand for a quant position. So, I would say, programming is what opened the door for me. Then, asking questions and formulating problems in math terms, was what kept me inside, and it still does. The quant industry has been declared mostly by quants to be past its peak. But I see it as reinventing itself. Machine learning is certainly playing a big part in this transformation. Yeah. Machine learning could certainly, it's a kind of language, a new set of tools, right? Yeah. So, how exactly you see this transformative role of machine learning? Yeah. Machine learning is of course old. It didn't just come about overnight, but it's certainly not old news. I think it may well be one of those ideas that become powerful when their time comes, which will be now. We know that many tech companies were transformed before our eyes. I'm not even talking about the stuff that grabs the headlines. Yes, we're waiting for the robots to come. We're waiting for the cars to drive themselves. But looking around us, we already see a lot of automation in small ways, and big. This is all driven by data. It's also driven by computational power and new algorithms. Above all, it's driven by passionate people, but at the bottom of it is data. Now, finance has always been recognized as a data-heavy industry. Does that mean that we can expect quants to be doing no more than fine-tune robots? Will that happen before the next class of students derives their third way of how to prove the Black-Scholes formula, or how to derive the Black-Scholes formula? We don't know. For one, financial data is not as signal-rich as the stuff that taught a machine how to recognize the image of a cat. In theory, when markets are efficient, then prices follow a random walk. Also in theory, there's no difference between practice and theory. Yeah. All I'm saying is that the problems we can, and will have to solve with machine learning and finance, are both harder and more rewarding. I tend to view machine learning as a tool not so different from traditional math. A tool that empowers us to ask questions, to answer those questions. So, is it transformative? Yes. Will we have to retool our arsenal of how we think about an applied math in the field of pricing, hedging, finding relative value, automating, executing strategies, et cetera? Yes, I think we will have to do all that. Will machine learning do our thinking for us? I hope not. Yeah. So, we won't be outsources anytime soon hopefully, right? Yeah, what you said reminded me of what Vladimir Vapnik, the father of support vector machine, once said in his book that, "There is nothing more practical than a good theory." Which I totally subscribe to the view. Coming back to your experience, what would be your advice for students who think about either moving to finance from other fields, or maybe, while working in finance, moving towards more of machine learning, can what they do? My advice is to be curious definitely while in school, but especially afterwards. In terms of opportunities, look around. The times that we live in are simply great. The sources to learn from are excellent. Learn new results that people publishing papers all the time. They first going in the public domain that the utilities and the ways we have to search those results to filter them, that the're unprecedented, it's not a new phenomenon but the amount of available results we have today and the ability to read feed back on on those papers it really is something that we didn't see even a few years ago. In addition, there are all these powerful computational libraries for machine learning that they're easy to get, easy to learn, easy to run. I won't mention all the examples of those, I'm sure you're doing your class. Also that there are all these high-quality online lectures and then full courses to learn from. I frankly tend to view the MOOC initiative, that Massively Open Online Course initiative is one of these luxuries of civilization, that the comparisons people make 150 years ago, you would have to feed a horse first to travel 20 miles from your house. Nowadays, you don't have to feed a horse to attend a lecture. Just one click away. Exactly. Yeah. We've come a long way from even 15 years ago. At the same time, the MOOCs, they are all available in their volume and their focus, I think they're somewhat commoditised, and that's one reason I viewed your course as filling a very important gap in the two fields machine learning and finance. With all these available resources, to go back to my advice for students, with so many available resources, it can be hard to stay on topic and long enough to master it. The signal to noise ratio that we're observing, pretty much everything is not on my list of luxuries of civilization. It's hard to deal with. So, I would say it's important to figure out already on what you really want and to stay focused on it. So, I realize I might be giving conflicting feedback here, so on the one hand. Life is conflicting. That that is true, but it may not, my advice is stay focused on a number of important things, at the same time be curious that it is it is conflicting, but it also reflects how machines learn and I'm sure you'll probably mentioned that in your course. The trade-off between exploration and exploitation. I also believe that's how humans learn as well. It's also something that's important for life to be able to make that trade off, and to realize where to put the emphasis. Yeah, I totally agree with. All what we do is continuous trade-off. You always balanced something versus something. Let me ask you, coming back as soon as we started to talk about physics, what's your today's relation with physics? Do you do miss it? Do you see any interesting development coming from physics? Yeah. what can you say about that. I actually removed myself from almost entirely from developments in physics after I left academia and this wasn't an intentional choice, but it was more a reflection of my full absorption into the quantum world. I didn't have a vague desire to find a practical application of re-normalization group methods and finance, but other than a few brief literature reviews, I hadn't done much. Also while being out of touch with physics, I felt for one of those psychological biases. It may be called unavailability bias. I imagine that the whole field was frozen in a state where I had left it. Quite recently, I got in touch with a friend of mine from grad school who stayed in academia and is now quite active in one of those interdisciplinary branches off of theoretical physics. I could label him as a biophysicist, but he's interested in in a truly broad range of things. I wouldn't exaggerate if I said that that he opened my eyes to what has been happening in physics while I thought it was frozen in time. I learned about DMRG, density matrix re-normalization group sensor networks, and most importantly I learned that there's an active effort to put the efficiency of deep learning on a solid theoretical footing. Deep learning is already borrowing from physics, for example, the ideas around tensor train and I figured there's so much for me to learn and borrow as well. Simply by paying more attention to the field I grew up in. After years of reading and studying on the job, at times greedily doing this, reading on statistics, optimization, operations Research, classical, numerical, algorithms, and more recently on machine learning, frankly, it gives me satisfaction to things like quantum information and entanglement, they could be coming to join the vocabulary of machine learning as a distinct contribution from physics. I think it's really exciting to be working on machine learning, applications on finance, and to reconnect with physics because it has a very rich set of ideas to contribute. I totally understand what you're saying because when I've heard from you about Tensor Networks, thank you, and I had the same feeling when they start to look it's like you feel like you coming back home, right? After all these travels in statistics, t-tests, F tests, you know, support vector machines, all of a sudden you see yourself in a familiar terrain, right? And by the way, more to what you said about how much physics contributed to deep learning. Most of deep learning models are Ice Inc models, inspired by Ice and great organization group, Ice Inc model, it's all like common. I think you put it nicely. It's almost like coming home, it feels like a natural place to be using the machinery and the ideas and just a way of thinking. But outside the physics, yeah. But outside physics, yeah. It's a very broad field anyway. Yes. Okay and also it was a pleasure to have you here. Thank you so much. Pleasure is mine. Thank you.
