AI & Machine Learning in Trading | Michael Himmel | Quant News


Michael Himmel is a Founding Partner, Portfolio Manager and Director of AI Research for Essex Asset Management.   Michael began his career in 1985 with Shearson Lehman Bros., in the pits of the New York Mercantile exchange [NYMEX].  He spent the following decade in the institutional commodity futures, options and derivatives business, including three years in London trading Futures and Physical Crude Oil. After moving back to the U.S., he became head of the Cinergy region trading team at PG&E Energy Trading and in mid-1998 he went on to co-found Lambeth Master Fund, a Global-Macro hedge fund and CTA.  The Lambeth Fund was the No.1 Performing Global Macro Hedge Fund in 1999.  Mr. Himmel then joined forces with the current Essex Asset Management team in 2005, to establish the Vero Beach, Florida office for Bank of New York-Mellon and, later, Essex Asset Management.

Mr. Himmel considers his professional mandate to be:  consistently allocating capital by scientifically sound methods; managing a long-term portfolio of robust strategies; and being a trusted advisor to successful investors and families. You can contact him and learn more at

Kurt Molina: Take us back to how you first got into trading and what that experience was like.

Michael Himmel: I started on the New York Mercantile Exchange in 1985. I was just as a runner so my first experience in markets, or trading, or futures was down there. At that point the crude oil contract didn’t exist so it was mainly heating oil, gasoline, platinum and palladium…that was my first introduction to trading.

KM: And at what point did you first realize some of the capabilities of artificial intelligence and machine learning in trading?

MH: In the early 1990’s I started reading about Complexity Theory and Chaos Theory. It was an emerging science and very few people were applying any of those concepts to trading at that point in time. However, that was during the period, the “Turtles” came to the forefront; as well as a whole bunch of other systematic traders. It was obvious that people were finding that “pattern recognition” and data crunching using computers was giving them an edge, however I don’t really think that there was much machine learning or many neural networks actually running back then.

“People were finding that pattern recognition and data crunching using computers was giving them an edge”

KM: Can you discuss with us some of the some of the biggest moments in your hedge fund experience?

MH: My very good friend and business partner at that time, Jeff Fisher, and I started our fund mid-1998 and we did become the number one small global macro hedge fund in the world in 1999— I think our net performance that year was around 260%.  We were not using machine learning or any AI at that time but we were definitely using a fair amount of trend following and turtle-esque money management, such as Kelly criterion derivatives. We were also using a lot of behavioral finance to the degree that we could at that time. We got very lucky because we caught the entire run up in Internet stocks in 1999. Then, by the time 2000 came along and the first tech crash occurred we were pretty much on the sidelines. We had taken all of our profits and were on to other things. When 9/11 happened in 2001 we did not have a lot of U.S. equity exposure in the fund so we were actually up a little bit, and we had a fair amount of survivor guilt. We were up that year and we were up every single year that we managed the fund.

KM: I know you’ve been looking into this for a couple of decades now, as you referenced earlier. What do you think have been some of the biggest advancements in AI and machine learning since you first began?

MH: From my perspective it’s absolutely “the cloud” [both processing and storage]. That is the big breakthrough for me and for hundreds of thousands of other people who have tried to climb the ladder in this game. The biggest breakthroughs in the world in AI and machine learning so far have not been really in finance– AlphaGo and the new Google Mind experiment, for example, are fantastic. To my knowledge the biggest breakthroughs have been in things like synthetic biology, image recognition, natural language processing, robotics, fraud detection, autonomous vehicles, cryptography, and things like that. But there’s definitely a bunch of people spending their life and all of their attention on applying these concepts to finance.

KM: When you first got started with machine learning for your trading systems, what do you think were some of the biggest struggles that you faced and how did you overcome them?

MH: I think my biggest struggle then (and still now) is definitely being seduced by backtests. Data mining, being seduced by backtests, or in general just not recognizing that much of the data analysis that’s taking place (especially in backtesting and feed forward testing) is really biased. Knowing how easy it is to get tricked into that state of belief is the biggest thing anybody will encounter. Also the understanding that actual performance really isn’t that actual. Even if it’s a well done experiment and the feed forward is done properly and all of the variables have been chosen scientifically, we can still get a bunch of bad data in the walk forward.

“My biggest struggle then (and still now) is definitely being seduced by backtests”

KM: Let’s say we have a retail trader who doesn’t really have any computer science background or is interested in using artificial intelligence in trading. How would you recommend them to go about this process?

MH: When I first asked this question to myself I was reminded of what Warren Buffett says to everybody when you ask that question:   ‘know thyself’. I really don’t mean to be pithy here, but it is important that anybody that tries to do this or get into this world that you have to know yourself, you have to know what you wish to accomplish. For example, some people don’t mind being up at 3 AM day trading or night trading currencies for thirty pips or something like that. Some people only want to do trend following, some people really only want to do mean reversion trades or index arbitrage strategies. And so the beginning study is in yourself, as you need to know what kinds of systems you would like to participate in and which ones fit your sensibilities. Once you do that then go out and find people who are doing that successfully right now. The beauty of the cloud is that the small trader or the small investor or small systems designer has a nearly unlimited amount of capabilities in accessing other people out there who are using cloud databases and cloud analytics engines to design and implement their systems. So I would recommend that a person start with the big question of ‘what are you trying to accomplish’? And then go out and find people who are doing it successfully right now — either subscribe to their databases or subscribe to their systems.

“Start with the big question of ‘what are you trying to accomplish’? And then go out and find people who are doing it successfully right now”

KM: In terms of your machine learning system, how do you go about filtering out what would not be a proper input?

MH: That’s the one hundred million dollar question. Ultimately I have always let the machines decide. It’s important to reduce the problem and make it as simple as possible, but no simpler. It’s very easy to fall into the pattern of throwing more variables at a machine learning system and that only drives you even closer to curve fitting. In general I think it is possible for quantum computers to contemplate every single variable that you throw at them and yet find the actual optimal value for each but that’s a different kind of question we’re not really dealing with just yet.   I would say the art of the entire process is to decide which variables you need and which ones you don’t and let the machine make some sense of the ones that you keep.

KM: Where do you think we are in this technical evolution?

MH: I think that we are at the knee of the exponential curve. So we are about at the point where the change is going to happen so fast it will feel like a blur to people. And I think that the answer right now in terms of the evolutionary process of finance is that man is still important, machines are very important, and the combination of man and machine is the winning combination, as it’s the only way to get the amplification that’s necessary to stay in the game. I’m aware of funds for example that comb every single newspaper in the world every single second of every single day in every single language and read everything that there is to read about every single stock and come up with a score on it. It seems like an astronomical undertaking but in fact there’s plenty of people doing that sort of stuff in the world right now and the only way you can possibly do that is with a machine. So that’s what I mean by amplification.

KM: What do you think are the best uses of AI and machine learning in finance?

MH: Long term asset allocation, for me personally. I’m applying those techniques to asset allocation as opposed to applying machine learning to long/short equity or short-term trading. I also use it for position sizing. I think that in the long run AI will be a great adjunct to any risk management program.

KM: We have seen quantum computing picking up steam recently. What effect do you see quantum computing having on financial markets in the future?

MH: I think it’s going to be ineffable. Once the quantum computers come into the financial market I think it’s going to be a complete reshuffling of the deck, I really do. And they are already out there and they’re already working and I know that there are people who are using quantum computers already in finance. I don’t know exactly what they’re doing or what their results are but Google, Microsoft, D Wave and a few other players have already announced open-sourcing their quantum computers to programmers to create. They’re trying to encourage a new ecosystem to be built around new programming languages, for example to communicate with the quantum computers. Once that gets going and those become cloud-based I think it’s going to change the game completely.

KM: What advice do you have for those looking to embark on this journey of quantitative trading?

MH: Firstly, you need creativity because systems that will be successful in the future are not necessarily ones have been successful in the past so there needs to be some level of creativity brought to the game. Secondly, a very deep understanding of machine learning and avoiding the pitfalls of overfitting and curve fitting and backtest seduction as we talked about before. Number three, programming skills and/or database management. Nobody needs to be a very high level Python or C++ or R programmer but it sure does help. I’ve put myself through many Coursera courses over the years and it’s fantastic. The fourth thing I would say is hard work because it really does take a long time to put these three things together and make something out of it that hasn’t been done before.

Risk Warning: The FXCM Group does not guarantee accuracy and will not accept liability for any loss or damage which arise directly or indirectly from use of or reliance on information contained within the webinars. The FXCM Group may provide general commentary which is not intended as investment advice and must not be construed as such. FX/CFD trading carries a risk of losses in excess of your deposited funds and may not be suitable for all investors. Please ensure that you fully understand the risks involved.