The speed at which algorithmic trades are executed, usually milliseconds, is shrinking fast: by a factor of 10 since 2007, says Kevin McPartland at Tabb Group, due to investment in dedicated optical-fiber networks and faster routing devices.

The average execution time for one class of small trade on the New York stock exchange fell from 10 seconds to 0.7 seconds between 2005 and 2009.

“The big problem,” says James Angel, who studies financial markets at Georgetown University in Washington DC, “is that US exchanges have no real-time safeguard against extreme malfunction.” If a mistake in an algorithm – or even deliberate sabotage – were to set off a wave of selling, the people tasked with suspending markets in the event of dangerous trading would not be able to react quick enough.

He proposes that the US Securities and Exchange Commission require that all markets incorporate a “circuit-breaker” that would suspend trading on signs of an algorithm-driven crash.