By: Simon Burns
Quant Development Intern16.07.2013
Stellar Interview with Spencer Connaughton: 21 Year Old Quant Fund Manager at Archivolt Partners
Spencer Connaughton had a chat with our Growth Hacker Simon Burns on the learning curve in becoming an algorithmic/quant trader using his school’s Bloomberg terminals (and crashing them), his firm’s use of the Residual Income Valuation model in algorithmic trading and the potential of markets with 20-somethings running them. Spencer started and runs a quant, grey box algorithmic trading fund called Archivolt Partners alongside Andrew DeTrempe, a Finance student. Archivolt Partners has built its strategy through close collaboration with Residual Income Valuation model pioneer and Southern Methodist University professor Dr. Greg Sommers. We had a great time talking to Spencer. His insightful, captivating and at times hilarious stories and analysis on the state of algorithmic trading, starting an ATS fund and democratizing model creation was enjoyable. Here’s our interview with Spencer Connaughton:
Simon Burns: Thank you for joining us today Spencer Connaughton. Spencer you are notable and have acquired some fame as a young quantitative fund manager. Your fund is called Archivolt Partners and was started, as many funds have started, from academic grounds. Can you take us through the history of the fund and how it came to be?
Spencer Connaughton: Sure. My former college residential advisor, Andrew DeTrempe, approached me one day and said: “Spencer, my professor Dr. Greg Sommers just told me about this highly accurate equity valuation methodology that he has studied for decades. But it takes an army of grad students to gather all the data for it and crunch all the numbers. It sounds like the kind of problem that you could write a computer program to solve. Want to give it a shot?” So we sat down with his professor and talked through the issues involved. I was thrown off the deep-end into the world of earnings-based valuation methodologies. We decided to set up this program on the Bloomberg platform to look at every stock in the universe, and run our valuation formula on each of them. Once we got into backtesting, we realized there was something there. We spent about 6 months developing and studying the algorithm, then we formed a company and raised money from our friends and family. So that’s how Archivolt Partners started.
Simon Burns: Great. So it’s interesting to me that you are originally from a computer science/electrical engineering background. Was that your first foray into finance?
Spencer Connaughton: Well, I spent my teen years doing one-off web and coding projects and later working for startup companies writing software, and that’s how I discovered my passion for tech entraprenuership. I decided to study electrical engineering and math in college to further broaden my skill set. After getting a good grip early on about hardware, software, and the underlying theory of both, I then dug into the fundamentals of business. A friend in New York was working on some interesting, unconventional financing stuff at the time. I called him up and said, “Hey, can you teach me a few things about finance?” Before I knew it, I was reading through corporate filings on EDGAR, studying warrant forms, looking for arbitrage opportunities, and doing all kinds of crazy research on Bloomberg. At a certain point, the business students at SMU (Southern Methodist University) began asking me for help on their projects, and that became kind of fun.
Simon Burns: Just to clarify, what was the language you were coding your algorithm in? Would you say it was an easy switch from the programming side over to the quant side or did you have to learn some industry specific programming skills to be able to be successful coding a finance algorithmic model?
Spencer Connaughton: I prototyped the system in Mathematica and built my own backtesting system first in C++. I then moved everything over to Bloomberg formulas on my school’s subscription. I found out that Bloomberg had begun beta testing a factor backtesting system and a portfolio backtesting system, so I spent a few hours one afternoon calling people at Bloomberg until they let me into those betas. I think a lot of programmers, and engineers for that matter, know how smart they are and go into finance bullheaded saying “Yeah, if can solve these finite difference time domain problems, I can easily do finance.” That was how it started for me. I thought, “It’s going to be simple, It’s all algebra.” Of course, that wasn’t the case, and I ultimately ended up spending a massive amount of time learning all the ins and outs and all the terminology and the language of finance. It took a lot of hard work. Much more than I thought. But it has definitely been well worth it.
Simon Burns: Awesome. So in terms of your inspirations and influences at this earlier period where you moved into finance, where were you getting inspired?
Spencer Connaughton: My Uncle John has always been pretty inspirational. He’s been incredibly successful in private equity, and sees the world in a really interesting way that allows him to discover and implement. The guys I worked with in New York have a completely orthogonal perspective. They are always looking for the opportunities that have been overlooked by larger institutions or are simply too complicated for most people to take the time to understand. They taught me a contrarian approach to finance. I think learning how to look at markets and deal structures from an unconventional perspective has been really interesting to me in my development.
Simon Burns: Spencer, you mentioned you had a 6-month process of going through, iterating, finding correlations, data mining and backtesting your strategy. Can you take us through your backtesting process and how you were strategically adjusting through the 6 months.
Spencer Connaughton: Sure. We primarily used Bloomberg for our strategic backtesting. I wrote an early version of the algorithm in C++ just to see how fast I could get it, and to see if it could run on a high-frequency, low latency (HFT) basis. We ultimately ended up going with a low-frequency strategy where, for example, we rebalance once a month instead of once a second. My Mathematica model was also really, really valuable because Mathematica lets you play with different variables one at a time and isolate elements, draw graphs and visualize your system.
Simon Burns: Mathematica, by the same people who did Wolfram Alpha is that right?
Spencer Connaughton: Yes, I guess you could call it the more advanced desktop software version of Wolfram Alpha. The methodology we use is a long algebraic equation that involves a lot of different variables. Once, we tried solve for our ‘R’ variable, the risk of the company. So instead of looking for the dollar value of the stock based on a risk assumption, we want to start with the dollar value of the stock and calculate the implied risk. So we plugged the equation into Mathematica and tried to solve for our risk variable. We ended up with a 6-page long equation. It’s amazing what you can do with these mathematics packages, because there is no way that anyone is going to spend hours and hours in front of a whiteboard solving these problems by hand. Instead, I just hit enter on my keyboard and I had the answer I was looking for. So that was pretty cool! Ultimately, we ended up back testing using Bloomberg’s EQS back testing feature. This is sort of a funny story (at least to people like us!). I was building this very complicated methodology into the EQS formula system. One day I was testing my formulas, adding terms to them one by one. I had terms nested in terms. I was doing a lot of debugging. And I was running simulations, multiple simulations all at once, so I could compare different manipulations of different metrics supplied by Bloomberg. Suddenly, the EQS system stopped working and I found myself locked out of the Bloomberg terminal. I called SMU’s Bloomberg sales rep and told him that I couldn’t get into the system. He asked “what were you doing right before you got kicked out?” and I said that I had just been playing with the Bloomberg factor backtesting package. And he said, “Oh, you’re the guy! We’ve had hedge fund managers calling us all morning complaining that EQS was slow. Now none of our clients have been able to run EQS backtests for the last hour because of what you were doing. We can’t let you back in.” I, of course, was very displeased with that. I negotiated my way into talking with the programming team that was actually working on the features that I has been beta testing to try and get the problem resolved. They were obviously a bit peeved at me!
Simon Burns: Alright, so take me through how exactly you crashed the industry standard, hegemonic financial market software in Bloomberg!
Spencer Connaughton: Absolutely! What was happening, and I’m sure that you guys at QuantConnect have these kind of problems all the time, was as I was calculating the risk for a particular stock, one of the terms I was using was beta. And when you calculate beta for a stock, you need to go all the way back in time and look at the stock’s correlation with the market on a month-by-month basis, right? So I had written this formula for risk which was used multiple times throughout my system of equations. And these equations depended on other equations that used the same variable. And those equations depended on other equations that used it. When you expanded it all out, I was probably using this term maybe 30 or 40 times. The flaw in Bloomberg’s system was that every time I used it, the system would manually go back and re-calculate everything each time.
Simon Burns: So it would never cache that value…
Spencer Connaughton: Exactly, It would never cache the value. So for every stock, the formula would calculate beta 40 times per iteration, each involving about 60 correlation calculations, for about 260 iterations per stock. So I was sucking up all of the CPU time for all of the computers in their cluster that had been set up to work on this, and eventually crashed the whole cluster. With a bit of badgering from me, they implemented a variable caching system. I think they already had it on their schedule to do, but they decided to prioritize it after their experience with me. Then the sales rep bought me pizza on his next trip to Dallas for helping them improve their system.
Simon Burns: So if you were talking to a new quant what advice would you give them?
Spencer Connaughton: It kind of goes without saying: do your research. There are more quants out there than you think, and most of them are smarter than you and have been making money at this since the only tools they had were written in Visual Basic running on Excel Spreadsheets.
Simon Burns: So in describing the Archivolt Partners model, it’s clear that the model is a very expansive and complicated valuation model. In terms of iterating or improving upon what you’re seeing in the backtesting, are you continuously iterating and coming towards a better algorithm or is your system intended to be static for the long term? And along those lines, the model was originally intended for academic purposes to be an improvement on the DCF and other standard valuation models. Do you expect to see greater adoption for the Residual Income Valuation model that Archivolt is built on?
Spencer Connaughton: Seeing broader adoption of our model would be awesome. Our model’s error bars are significantly smaller than those you see on DCFs or DvDs, especially with Bloomberg’s data. The fundamental theory behind our methodology is still largely the same as when we started, although we figured out how to incorporate data that hasn’t previously been availible to folks studying this type of model. We might consider using different data sources or metrics for some of our variables, but we believe the underlying theory is sound, and will remain so for a good long time.
Simon Burns: So we were talking about Bloomberg, and your use of EQS. In your period of integrating data, testing and looking at your historical data simulations, a big problem quants face is curve fitting. How do you avoid fitting your models to historical data as opposed to looking at how it performed in real time (paper trading)?
Spencer Connaughton: That’s an excellent question. We had a lot of discussions about that. Our model is based on over 20 years of research, so to avoid curve fitting, we had to be very careful to keep to the foundations of our strategy. When we first started, Andrew and I would try to change that one term, or add that extra variable, or discount that value back by the square of the error, and long term returns looked better. But then we really had to ask ourselves, what new data are those modifications actually communicating? How does our change actually speak to the valuation? It was great having Dr. Sommers on the team to guide us because of his focus on the academic validity of what we were doing. So he would look at our work and say, “Why did you change that? Now it doesn’t make any sense.” And we would reply, “Well, it works better!” And he would come back with, “Well, now we don’t know why it works.” The voice of reason. That’s another thing I would definitely recommend that new quants be sure to consult with. You want something that works to also make sense; because if it works and doesn’t make sense, it’s probably not going to work for very long. New quants are always coming in with great programming and analytical skills but don’t necessarily have a finance background. It’s easy to make this curve fitting mistake when you are simply data-mining and finding correlations. You always need to have that qualitative reasoning about what you are doing.
Simon Burns: In terms of non-market data, you are using analyst estimates in your system. Where do you see the future of that non-market data for use in quantitative trading with the emergence or Twitter based Hedge Funds and other non-market data strategies?
Spencer Connaughton: A big advantage of our implementation of this valuation methodology is that it uses Bloomberg’s non-market data. It uses their analyst estimates. What you guys are doing with Estimize is pretty exciting. I definitely want to try it out to see how it compares to our current data sources. It’s huge because you can get at what is really driving the prices of stocks. What estimates people are making and how that reflects in their trading. In a perfect world, maybe market data would be all you need to make money in the market; but of course everything else going on in the world and in the minds of traders plays into it.
Simon Burns: Interesting stuff. I interviewed, Stefan Nann of StockPulse this week, who dug into how they look at the exact words used on Twitter about a stock to determine their sentiment score. So the word “great” versus “outstanding” is valued differently in sentiment processing, which is definitely on the way to real-time voice processing. Spencer, how old were you when you first started gaining quant space or when you started coding with financial data inputs?
Spencer Connaughton: This was just a year and a half ago, so I guess I was 20.
Simon Burns: So how do you see quantitative trading going forward when it’s possible to be a 20-year-old with basic to intermediate coding skills and be able to build a successful algorithm. With this democratization of financial algorithm creation, do you see the future bringing a greater diversity of algorithms, more complicated algorithms, more innovation?
Spencer Connaughton: All I can say on that front is look at what happened when we made it easy for 20-year-olds to get into the computer programming space itself. In the past, the only people with access to computers were bearded scientists in secret laboratories in obscure universities with mega-budgets. When it became feasible for a 20-year-old to get into a computer lab and write his own programs, you got Bill Gates. So I would certainly say that we’re in for some pretty exciting stuff in the coming years with quant trading. Younger people getting involved means that we can look forward to rapid innovation on a massive scale.
Simon Burns: I agree! Given that quant trading is so secretive, the quant trading version of a 20-year-old Bill Gates may still be bearded and in a basement out of the public eye! So the last question is very open-ended. What’s the next for Archivolt? What’s next for you? Where do you see yourself in the 5 years.
Spencer Connaughton: After the first big 6-month push, Archivolt Partners has turned out to be very easy to operate. So I’ve been working on other projects. Now that Archivolt Partners has a solid track record, I think it’s time we raised some real money and look toward the future of what our system could be. It’s a pretty exciting time for Archivolt. We’ll see where it goes.
Simon Burns: It certainly is exciting to see the tools of algorithmic model creation being used by a 21 year-old to innovate on the dominate DCF model, crash Bloomberg and grow a hedge fund! We will be following your story closely and hope for you the most success at Archivolt Partners. Thank you so much for your time Spencer.
Looking to try out QuantConnect? QuantConnect gives you free access to high resolution data for global financial markets to backtest your algorithm in our simulator. Once you’re ready you build and backtest your algorithm right from QuantConnect. You’ll be presented with your strategy equity curve and key performance indicators! Sign Up for Free