Innovations in Trading and Technology (“Then there is the man who drowned crossing a stream with an average depth of six inches.” W.I.E. Gates.
Every week it seems that some organization or other is making headline claims on a new latency figure. From two digit microsecond matching to low latency access to new trading facilities, the race seems to be on to determine who can report the lowest latency.
Here’s the question: are the numbers we see in these headlines at all useful? I’ve previously noted that it’s crucial to measure latency in a holistic fashion, across all components in your trading flow. However, like many other latency headlines, numbers such as those above tend to be focused on a single point in that flow: market data acquisition, or message distribution, or trade execution. (I also imagined that I could find an Einstein quote suitable for this post, but I’ve failed miserably – let’s hope the remainder of this post makes up for it!).
There’s another reason for doubting the usefulness of these figures. There are millions of transactions flowing through your systems every hour. Market behavior is constantly changing, and volumes can vary by an order of magnitude from one minute to the next. In an environment like this, nothing stays constant for any significant length of time, so specifying latency with a single number is meaningless. You could send orders to market with an average latency of 1ms, but still have 10% or more taking over 100ms – presumably not a situation you want to be in.
Comments | Post a Comment
4 Comments to "Latency – A Sting in the Tail":
louislovas
11 November 2010
I couldn't agree more with this approach. Coming from the vendor side of this equation, a solution/platform is only as good (performance-wise) as the weakest link - and there are many moving parts in end-to-end trading platform. There have been efforts to create standardized measures for performance benchmarks - but other than those that measure just a few of the lower-level components, nothing has materialized. In the end it's a tall order to create such a model. For prospective consumers interested in determining performance (latency and throughput) characteristics of a platform - I suggest they devise a test structure based on their own requirements - ensure they measure end-to-end as this article suggest and gather the statistical metrics from vendor products.
Comments (29)
dmishoe
11 November 2010
Amen! latency stats are marketing and spin, without wholesale reporting of loads within latency stats, along with time of day and which market. we see latency averages vary wildly across venues throughout the cours of the day, sometimes in relation to volume swings, sometimes not. ultimately the orders follow liquidity, not latency. the good news for those venues with low order flow, you get to report better latency stats! You wouldn't want any of those pesky client orders mucking up your pretty latency stats...
Comments (3)
csparrow
12 November 2010
Good article. What I am struck by is that this article could easily be about TCA and the need for a statistical/distributional approach rather than focusing only on the average - outliers are very important to understand! It is also important to look holistically at all of the components such as delay costs and opportunity costs and there needs to be a recognition that changing conditions can have a large impact on results - t-cost and risk models work well in "normal" markets but can breakdown when historical relationships change.
Comments (37)
ptrduffy
15 November 2010
Looks like I've provoked an outbreak of violent agreement! @louislovas - absolutely agree that there's no substitute for measuring your own systems. The challenge that we see (and help our customers address) is that the volume of data is massive, and much of it is unstructured, so you need the right skillset and tools to implement this approach. @dmishoe - in our experience volume is one of the biggest drivers of latency changes, although definitely not the only one. If only those pesky clients would save their orders for quieter times! @csparrow - yes, the statistical/distribution approach is one that can be (and is) equally well applied in many other areas. in many ways, what we're doing is simply bringing that rigorous scientific approach to quantifying the performance and capacity of trading systems. One of the interesting things (for me) is that one characteristic which makes these systems so complex, the volume of transactions they process, is also what makes them well-suited to this approach, as you can easily get the large number of samples needed for a statistically valid analysis.
Comments (4)