Every once in a while
my spider senses pick up a familiar signal...that is, The Herd is moving in
lock-step and picking up speed.
This time around it's
about the obsession of shaving latency.
Shaving latency seems to have become the biggest cottage industry in trading.
Every day I get more
and more spam about latency....and you know that things are out of hand when
the spammers are on to it.
How about this one:
Dear Friend,
I am a Nigerian Prince who has been exiled from my country. My father has
deposited $420,000,000(US) in a trading account in Zurich.
If you give me your latency shaving algos, in return, I will share our
increased profits with you 50-50.
Please e-mail your source code, router IPs, passwords, etc to:
honey-I-found-an-answer-to-our-prayers@lagos.net
I agree that speed is important,
but is it the whole ball of wax?
Could it be that the latency
shaving stuff is just another arms race?
A contest where no one
is the clear winner?
Won't we all continue
to have pretty much the same speed as the technology continues to evolve and leak out?
Wouldn't it be better
to spend your time working on algos that are just plain smarter at analyzing
and digesting all of the true market data that is currently available than
simply playing the same old card over and over again of trying to be faster?
I was chatting with
developer friend at a prop shop in Chicago and he told me that their in-house
programming bench consisted of 6 guys and that each and every one of them was
only working on shaving latency.
If a shop's algos can
only thrive with that extra nth of second in their pocket then
perhaps it's time to focus a little more of energy developing cleverer setups
and entries....Just a thought.
Anyone care to comment? Am I off base here?
You're not off base, but here is my take on it.
Shaving latency is a concrete problem. We know current figures, and we can shave off microseconds here and there and immediately measure ROI.
Creating new algos is a different beast. It s like research... and can end up nowhere. Few firms can afford a budget for research - surely not the small shops who need to pay the bills.
Most people who start hedge funds come with an existing strategy/algorithm they acquired/developer while at another big bank. Then it is time to maximize profit on the algo - and reduce latency.
Same problem as in IT industry. Few firms can afford a big budget in R&D (IBM, Microsoft...) Then those researchers leave those big companies and create start ups.
Same story with finance. DE Shaw and other big hedge funds often see their employees leave with a big idea/algo to set up their own fund and chase those microseconds...
I couldn't agree more. It's like the quote from "The Incredibles" (hate to quote cartoon movies, but if it fits...): "If everybody is special...then NOBODY is."
Regardless, the race for ever-lower latency is likely to exhibit asymptotic characteristics. Since "zero" latency will never be achieved, zero is the limit to which it will probably become exponentially more difficult to approach - to the point where any extra gains will not offset the increase in costs associated with chasing them.
I liked you suggestion - concentrate on algorithms that are "just plain smarter". That's the tack I've taken, and it's been working well for me so far.
Ultra low latency isn't for everyone and the profits can be fleeting.
However, there are many instruments that should be trading in lock step but often are not. For example ETFs are supposed to mimic the indices they track but very often there are decent gaps that can open up trading opportunities. This is particularly true when markets are volatile.
And there are so many alternative markets- order books, exchanges and ECNs that there is plenty room for those seeking to trade on inter market latency.
I think that a lot of firms do carry out the pursuit for low latency in-house but the R&D costs can be very high. And as Irwin says above, if trader/programmers can figure out some magic formula they will sometimes leave and set up their own firm.
One of our customers claims to have made enough in one month on our system to buy the building they are in. They are somewhere in Manhattan. So I guess it is a lot of money.
Even though the pursuit appears to offer asymptotic profitability the race to zero goes on.
Do you think anyone will ever run a faster 100m than Usain Bolt? I bet you there's a few sprinters that are training hard to do just that.
All to often in the algo space, the dreaded hunt for latency is leveraged in place of the research effort to find effective algorithms. I agree with the post above that latency measurement is somewhat quantifiable, however variable and dependent on many, often moving factors. While latency (and reduction thereof) is an important part of algo trading, its importance varies with different markets, asset classes, and types of signals. For high frequency trading in very liquid asset classes, it is a paramount concern. For less liquid asset classes, whose lower frequency execution relies on complex dependencies and less on pure execution risk, the efforts are better directed toward tuning the algorithms.
The costs are also a factor. When considering latency reduction techniques such as colocation, one needs to take into account the cost per execution in terms of infrastructure, points of failure, communications cost, hosting, etc against the potential profit to be earned by marginally faster execution.
I have three words for latency reduction:
"Assembly Level Programming"
:-)
Alternatively, why spend time and money developing smarter algos when your infrastructure is slow enough to lose you any advantages those smart algos might have gained? The key is to finding the right balance, and indeed, latency reduction can't continue endlessly. And the costs of shaving microseconds eventually outweigh the benefits.