Over the past twenty five years, machine learning had evolved from being a collection of rather primitive, yet clever, set of methods and heuristics to do classification, to a sophisticated science that is rich in theory and applications. In the foreign exchange market, for example, the market impact of computer-generated trading activities has been keenly observed. To better understand the situation, a simple metaphor based on the classic game of rock-paper-scissors illustrates how the same problem can be solved by two very different and contrasting approaches: high-speed technology solutions vs. empirical machine-learned solutions.
In the first approach, a player does not try to over-think what the opponent might do next, but instead quickly respond to the situation using a high-speed information loop. One can think of the high-frequency-traders as penultimate masters of this first approach, where nobody wants to be last. In the second approach, a stylized “win-stay, lose-shift” phenomenon that is observed empirically among the players becomes the statistical basis for generating a series of sound strategies designed to win the game. One can think of the quantitative traders as skilled practitioners of this second approach.
We shall call systems that implement the first approach "Type 1 Bots", while systems that implement the second approach "Type 2 Bots". It is well understood that human players are easily defeated by "Type 1 Bots" because of the inherent speed disparity. However, it is somewhat puzzling how "Type 2 Bots" manage to defeat even the most intelligent of human players. It turns out that there is actually a trick to it: humans are all too predictable.
Human beings crave positive reinforcement and cringe from negative reinforcement. This important behavioral concept was understood by David Hagelbarger as early as 1953 and became the basis for Bell Labs very first Mind-Reading Machine. Claude Shannon, who was Hagelbarger’s contemporary at Bell Labs, took note and designed an improved version of the machine. Shannon’s Outguessing Machine, though simpler than Hagelbarger’s in that it uses only 16 bits of memory for storage, turned out to be a superior predictor. The machine predicted "random" human choices. But since no one chooses randomly, the machine always won the guessing game.
What’s perhaps most interesting is the fact that his machine displays an innocent-looking “scoring bar”, i.e., a row of up to fifty ball bearings, in a prominent place at the front. By subtly inducing in the human mind a default way to frame their choices around “what worked or what didn't work the last time”, the scoring bar "tricked" the human players into a more predictable pattern of play, which of course, is precisely anticipated by Shannon’s ingenious “mind-reading” algorithm.
As Benjamin Graham famously noted, “In the short run, the market is a voting machine but in the long run, it is a weighing machine.” So from this perspective, the “trading game” is really not that much different from a rock-paper-scissors game. Benjamin Graham's voting machine and weighing machine can both be viewed as Type 2 Bots that operate at different time horizons, but within the same market. In the short-run version of the trading game, advantage accrues to one who can quickly see what everyone else is doing before the majority catches on, or who can offer a better guess about what everyone else is about to do. In the long-run, it appears that speed is no longer material and fundamental knowledge is all that matters. One game, but two different types of game play depending on your time horizon. Very interesting, indeed. But what of the intermediate times in between? Are there newer types of game play waiting to be discovered?
Space Machine’s automated trading algorithms attempts to find the right optimization points along the equivalent of a science fictional “space-time continuum” so as to have a proper balance of speed and time given fixed geographic distances between trading venues. It is perhaps easier to visualize this concept by studying a special case that relates to the stock markets, where Alex Wissner-Gross and Cameron Freer had recently computed the optimal trading points (blue dots) for Type 1 Bots between different stock markets (red dots) as shown in following map:
Here is how we believe the whole thing should work. On rare occasions when we chance upon a momentary speed advantage in the market, we can simply look back at what just happened to briefly enjoy a near 99% probability of success (e.g., as in the case of the superfast rock-paper-scissors robot). But since we are “speed-challenged” compared to high-frequency traders with proprietary infrastructure, more often than not we have to look instead further ahead into the future to locate our trading niche in the market ecology to find an acceptable probability of success. Overall, we can expect to generate optimized trading performances by judiciously betting on the right mix of models and strategies for all the different market situations that our data analytics algorithms can identify. Don't think this will work? Tell us why. We like to find out sooner than later.
As a start-up trading firm with no particular advantage in high-frequency trading, Space Machine's view on HFT is this: if you can’t beat them, you don’t have to join them. Nor do you have to give up and concede. You can try a different approach instead. You have to believe there always is another novel trading niche that remains hidden deep within the lush canopy of a flourishing market ecology waiting to be discovered. Seek, and you will find. When you are no longer fixated on a singular point ablaze at the distant horizon (i.e., to be the absolute fastest), a whole new frontier of trading opportunities magically opens up before your very eyes. Benjamin Graham's wisdom rings true even today, but with an interesting twist: to best capture value from today's high-speed electronic trading venues now requires machine-learned guesses on not just one, but several, well-chosen points along the opportunity frontier of attainable trading performances.
Therefore, we are cautiously optimistic about using machine learning techniques to build "Type 2 Bots" for this manner of search and discovery, so we don’t have to constantly do battles with speed daemons that are "Type 1 Bots". We require a reasonably low-latency market access infrastructure for automated trading, but that does not have to come at the full cost of approximating the speed of light. In short, we are actively investigating alternate ways to intelligently participate in the current speed-obsessed market ecology, without getting trapped in a ruinous technological arm race in the quest for top speed.
- Chaboud, Alain and Chiquoine, Ben and Hjalmarsson, Erik and Vega, Clara (2013, July 5). Rise of the Machines: Algorithmic Trading in the Foreign Exchange Market. Journal of Finance, Forthcoming; FRB International Finance Discussion Paper No. 980. Available at SSRN: http://ssrn.com/abstract=1501135 or http://dx.doi.org/10.2139/ssrn.1501135
- Superfast rock-paper-scissors robot wins every game it plays. (2013, November 5). Retrieved from http://www.designboom.com/technology/superfast-rock-paper-scissors-robot-wins-every-game-it-plays-11-05-2013/
- Johnston, Casey (2014, May 1). Scientists find a winning strategy for rock-paper-scissors. Ars Technica. Retrieved from http://arstechnica.com/science/2014/05/win-at-rock-paper-scissors-by-knowing-thy-opponent/
- Wang, Zhijian., Xu, Bin, and Zhou, Hai-Jun (2014). Social cycling and conditional responses in the Rock-Paper-Scissors game. arXiv preprint arXiv:1404.5199. Retrieved from http://arxiv.org/pdf/1404.5199v1.pdf
- Poundstone, William (2014, July 17). How I beat the mind-reading machine. Retrieved from http://wpoundstone.blogspot.com/2014/07/how-i-beat-mind-reading-machine.html
- McCabe, Thomas (2010, November 11). When the Speed of Light is Too Slow: Trading at the Edge. Kurzweil Accelerating Intelligence. Retrieved from http://www.kurzweilai.net/when-the-speed-of-light-is-too-slow
- Wissner-Gross, Alex, and Freer, Cameron (2010). Relativistic Statistical Arbitrage. Physical Review E, 82(056104), 1-7. Retrieved from http://www.alexwg.org/publications/PhysRevE_82-056104.pdf