Wall Street Journal | June 22, 2010
NEW YORK-Program traders looking to reduce mistakes that can cost thousands of dollars in the blink of an eye increasingly are testing algorithms using past market data to simulate trading.
The range, complexity and size of the process, known as backtrading, has become more popular since the May 6 "flash crash," when the stock market dropped nearly 1,000 points in a matter of minutes before rebounding.
"It's no longer an afterthought," said Richard Tibbetts, chief technology officer at trading-technology firm StreamBase Systems. "If you're going to be committing tens of millions of dollars to a trade that's quantitatively driven, you want to be very confident in your model."
Most companies test their trading strategies by running them through simulations created by vast databases of past market data. But programs have to factor in the effect of the algorithm itself.
"The mere presence of your trading activity would change the trading activity of other firms—that's what makes backtesting quite challenging," Mr. Tibbetts said.
Just testing reams of past data isn't always enough to see what algorithms can handle.
StreamBase engineers adapted a Nintendo Wii controller to help generate random numbers fed into a backtesting simulation used at a trade show, though the company tends to use less-playful number-generation tools in normal circumstances.
Other companies are roping in higher-powered computer chips to speed up the barrage of data launched at the trading formulas.
Firms are increasing the number of simulations and the level of detail created for the tests, Mr. Tibbetts said. New regulations such as circuit breakers are creating different trading parameters that have to be accurately reflected in the simulated environment for tests to predict an algorithm's performance accurately.
Since the market's precipitous plunge May 6, trading firms have been hustling to make sure their systems can deal with the rush of market messages that accompanied the huge spike in trading activity during the flash crash.
"Data rates peaked at extremely high levels during that period," said Howard Pein, chief executive of trading-software company CodeStreet.
On May 6, the number of messages sent across all live data feeds reached a high of more than 2.8 million messages a second at 2:43 p.m. EDT, according to the website MarketDataPeaks.com, which tracks messaging. That's sharp increase from the not-distant past, when market data feeds used to peak at tens of thousands of updates a second.
Algorithms and market data platforms were "stressed in ways they hadn't been stressed before," said Mr. Pein. Now firms are backtesting systems to make sure they can handle 500,000 updates a second and more, Pein said. "These are projects that are specifically commissioned to deal with the consequences of May 6."
Shock over the velocity of the flash crash has also had ripple effects. Traditional backtesting used to be largely confined to preparing algorithms before they actually entered the market. Companies were reluctant to run testing programs alongside their algorithms once they went live, as they often add small delays, or latency, to execution times.
But after the sharp drops in stocks during the financial crisis and, more recently, the May 6 plummet, traders are willing to run programs that monitor and tweak algorithms in real time, said John Bates, chief technology officer of Progress Software Corp. and founder of the Apama data-processing platform.
"In the traditional trading world, people used to turn off their risk management because it slowed things down and the potential upside was great, but the world has changed," Mr. Bates said. "Fear of regulations and the fact that we've seen you can lose large amounts of money in a short period of time has really turned that around."
-Kristina Peterson
Read full story >
