The Dying Lifecycle of an Algorithm

A while back I wrote a cover story for Waters on Lincoln Capital's Stephen Temes and he told me about how he felt there were firms out there that intentionally tried to figure out his algorithms and execute against them-or, game them.

"You have to really keep ahead of the curve," Temes told me. "When you start to see your orders being gamed or when things change in algorithms, you've got to quickly make changes. There are programs that are out there that look for my order and know to shoot against it."

I was reminded of this quote yesterday at Sifma 2011 after I met with Simon Garland, chief strategist of database provider Kx Systems. Whereas, in years past the typical short-end of a lifecycle for an algorithm ran about seven weeks, that timeframe has been significantly reduced, according to some of Garland's clients.

"We were speaking with a high-frequency trader a couple of days ago and he said that the typical life of an algorithm, these days, is about two weeks, because as soon as the people are on to you, then of course it's over and you have to do something else. Two weeks is the minimum now, because that's someone who's really good. So you can't send your good guys on holiday because they've got to have a new algorithm in two weeks."

When I told him that I was surprised by that-especially considering all the time and effort that goes into building an effective algo-Garland agreed.

"The two weeks surprised me, too; I thought it was longer," he says. I thought if you were really smart you could manage it a little bit longer than that."

So my question to the industry is this: Does two weeks sound too short...or even too long?