Insights on low latency execution practices at The Trading Show Chicago
Alex Dziejma of Dymaxion Capital was joined by a colleague at the same firm two weeks ago at The Trading Show Chicago. Mr. Dziejma cut right to the chase, “Why does HFT / low latency matter?” While the answer may seem obvious, what is not so obvious are the challenges and tools used to develop such systems. According to Mr. Dziejma, previous java / windows APIs do not cut it for model development. Existing APIs and off-the-shelf full service systems are too slow. The drivers for the delivery of a low latency trading system are market data feed handlers / book building, balancing model complexity vs. reaction time, and execution with respect to order generation and risk checking.
Upon having a robust system, what is the standard speed of execution today and tomorrow? Mr. Dziejma stated the speed required is in the eye of the beholder but had some thoughts about where the speed is and where it may be in the near future. “Today’s speed is approximately 20 microseconds from tick to trade in 2012, we expect less than 10 microseconds in 2013, and also expect executions of 1 microsecond in 2014.” Dymaxion Capital is currently at 5 usec now..and building to ~1usec on specialized trades for near future.
It’s important to know that getting a low latency / HF trading system off the ground is not an easy task. “Initial costs are high while the initial returns are low; many trades face the back test / capacity surprise as they enter production. Rarely do you find markets more accommodating than your test environment; and do not use big databases but comb through the needed information to execute your trades.” Certain strategies such as aligning development with a partner may help to build a next-gen trading platform, cut initial costs, and share some of the risks involved. The reality of the session kicked in when a colleague presented what he considers is the “18-month tech cycle”. It is possible that within 18-months, your infrastructure can be out-dated given the speed to which technology is moving. Items that should be considered for updating are your OS (although still using kernel bypass techniques), hardware (processor, NIC, internal bus), your network, the trading engine, and liquidity providers.
Specifically, which languages are used for development? The answer to this question is heavy on C / C++ using gcc compilers v.4.6-4.8, “Java, C# may work now but not later.” A colleague warned against elaborate software engineering, which may incur more costs and open more opportunities for misuse. FPGAs are also referenced for their parallel opportunities and speed, but programming them is a little more difficult than other technologies. A colleague added, "Your trading platform is the truest expression of your algorithm model at the micro-structure level. Measure, measure, measure!”
Rounding out the session were a few questions from the audience, “What components add the most latency in your system?” They respond with networks, switches, and geo-arbitrage execution. Specifically, what part of the internal stack adds the most latency? “The core accessing data, which uses 1.5 – 2.0 microseconds to make decisions.”
-Original content provided by Safraz Rampersaud, an on-site blogger at The Trading Show Chicago
