January 19, 2011
Algorithmic trading is getting another cycle of press scrutiny, thanks mainly to a very well-researched article in Wired by Reuters financial blogger Felix Salmon and Ars Technica writer Jon Stokes. In it, they outline how pervasive these high-tech algorithms have become to the everyday running of financial trading. And the problem is no one knows how this software drives the market behavior -- not the investors, not the traders, and not even the people who run Wall Street.
The motivation for all this high-tech trading is, of course, money. And in this case, he who develops the fastest system usually wins. This often comes down to placing the trading servers in the same room as the stock exchange servers to get that millisecond edge on executing transactions. The code too is designed for maximum speed, being constantly tweaked to squeeze the last ounce of performance from the underlying computer chips. Appro recently launched a server based on overclocked Intel "Westmere" CPUs, to give high frequency traders that extra speed boost. But all that digitally enhanced speed means it's that much harder for humans to control.
That's partly because that computer-generated bids can be executed so quickly (10,000 bids per second for a single stock) and in such a complex manner that humans cannot comprehend the ramifications. The feedback loops become intertwined, such that the entire trading system exhibits emergent behavior, untraceable to any particular piece of code.
In a recent interview on NPR's Fresh Air program, Salmon declared. "The man danger about algorithmic trading is that we simply don't understand it." He says although the individual algorithms are controlled, and presumably understood, by their masters, the interactions between them are not.
In researching his article, Salmon talked with Michael Kearns, a CS Prof at the University of Pennsylvania, who has developed algorithms for various Wall Street firms. Kearns told him that the financial markets have become what he called an "automated adaptive dynamical system with feedback." That may sound very cool, but according to Kearns there is no science he's aware of that is able to understand such a system.
It should come as no surprise that occasionally such a system would run the financial markets into a ditch. That happened last May, with the so-called flash crash, when the Dow Jones Industrial Average plummeted 900 points in a matter of minutes -- before regaining most of its value. The cause was traced to a relatively obscure mutual fund company that decided to make a very large trade in a very short amount of time (about 20 minutes). The algorithms monitoring the market interpreted this as a panic and came to the same decision all at once: sell. The reason the mutual fund company decided to dump the shares in the first place was to hedge against the possibility of a future stock market drop. Talk about self-fulfilling prophesies.
In the wake of the flash crash, the Securities and Exchange Commission (SEC) announced some measures intended to prevent a reoccurrence. These include "circuit breakers" procedures, such as automatically halting trading when a stocks share price fluctuates by more than 10 percent in 5 minutes. The SEC is also considering other measures like limiting the size and speed of trades and requiring a complete audit trail of all transactions.
But Salmon considers those rather crude remedies for such a tightly wound system. The flash crash event was actually a rather simple example of what could go wrong. The interactions between all the analytic software inhabiting Wall Street datacenters is much more complex. For example, unlike that mutual fund company that executed the large trade all at once, algorithmic trading software tries to hide a big buy or sell events with a series of smaller transactions so as not to tip their hand.
Meanwhile, other algorithms are simultaneously monitoring the activity to discern the larger patterns that the other codes are trying to hide. In some cases, even more devious codes will purposely initiate transactions with no intention of executing them in order to confuse their rival software. It's very much algorithmic warfare, with no real thought given to collateral damage.
The quantitative analysts themselves have become somewhat innocent bystanders. The Wired article describes a typical quant shop, in this case Berkeley-based Voleon Capital Management, that specializes in statistical arbitrage. The idea is to process mounds of financial data, looking for patterns that would point to an profitable arbitrage opportunity. But the quants have no knowledge of the underlying fundamentals of the assets; they are simply looking for patterns. To them, it's just a pile of bits unrelated to any larger reality.
The software is becoming more sophisticated too. Salmon documents a recently launched service, called Dow Jones Lexicon, that mines the text in financial news stories and attempts to map keywords to market conditions, the idea being to help predict market trends based on external events. Although such software is in its nascent stage, this could add a whole other layer of complexity to trading models.
The fact that so much trading -- the majority, in fact -- is performed algorithmically suggests that the market is no longer balanced between investors and speculators. And since the speculation component is being propelled by superfast computers, the market has become increasingly volatile and unpredictable. Even before and after the May 2010 flash crash, there have been a number of examples of unexplained price fluctuations.
The University of Pennsylvania's Kearns suggests that we should to build a ginormous stock market simulator in order to provide some much-needed science for our market structures. In a recent Reuters blog by Salmon, Kearns is quoted about this at length. Although, the professor doesn't see a simulator as a magic bullet, in his estimation it's certainly the place to start.
Given the importance of the stock market to the economy, its increasing susceptibility to damaging volatility, and the lack of our understanding of current system, a simulator project seems like a no-brainer. Sounds like a nice little science project for the SEC.
Posted by Michael Feldman - January 19, 2011 @ 6:12 PM, Pacific Standard Time
Michael Feldman is the editor of HPCwire.
No Recent Blog Comments
10/30/2013 | Cray, DDN, Mellanox, NetApp, ScaleMP, Supermicro, Xyratex | Creating data is easy… the challenge is getting it to the right place to make use of it. This paper discusses fresh solutions that can directly increase I/O efficiency, and the applications of these solutions to current, and new technology infrastructures.
10/01/2013 | IBM | A new trend is developing in the HPC space that is also affecting enterprise computing productivity with the arrival of “ultra-dense” hyper-scale servers.
Ken Claffey, SVP and General Manager at Xyratex, presents ClusterStor at the Vendor Showdown at ISC13 in Leipzig, Germany.
Join HPCwire Editor Nicole Hemsoth and Dr. David Bader from Georgia Tech as they take center stage on opening night at Atlanta's first Big Data Kick Off Week, filmed in front of a live audience. Nicole and David look at the evolution of HPC, today's big data challenges, discuss real world solutions, and reveal their predictions. Exactly what does the future holds for HPC?