How to make $100 Million in a Millisecond – Fighting latency and data loss

IT Editorial

Thursday, 19 February 2009 14:21

Written by Steve Wong

Trading orders are often placed electronically, based on trading decisions made using a variety of data inputs such as news feeds, bid and ask quotes, and trading volume history. These data inputs are sent by information and content providers and transferred to feed handlers and trading engines at trading firms via various computer networks. Trades are also placed electronically and using networks.

In recent years, news feeds have become quite sophisticated. Much of the financial news is distributed in a compact, electronically readable form (such as an xml file) that can be quickly evaluated by software rather than a human reader - often in a millisecond or less. Examples are Reuters’ Data Feed Direct and Dow Jones’ Elementalized  News Feed.

If a trading firm sets out to optimize the entire trading process, it becomes important to manage the input data faster and more appropriately than the other firms. However, certain questions arise at the outset:

• How can I get the news information faster?
• How can I make decisions faster?
• How can I make better decisions?
• How can I get my transaction orders to the market faster?
• How can I get my order confirmations faster?

If you can figure out the answers to these questions more quickly than your competitors, then you might be able to gain as many as a millisecond advantage in the time it takes to process input data and place a transaction order. This could mean the difference between placing a successful order and reaping the trading profits - or having a competitor preempt you.

By one estimate published in Information Week, effective management of data can make a difference of $100 million per year, or roughly $400,000 per trading day, for a major brokerage firm. Conversely, if one of your competitors has an implementation that is one millisecond faster, that competitor could make an extra $100 million a year - most likely at your expense or than of another competitor.

Obstacles to Automated Trading

A first step toward optimizing an automated trading system is to have a way of measuring possible obstructions that can interfere with the overall process of evaluating input data and placing transaction orders. Two main culprits are latency and message loss.

Two distinctly different types of latency affect the automated trading process – application latency and network latency. Application latency refers to the time it takes a trading engine to assess the data that it has at any given moment, placing a transaction order that is optimum for that moment. Application latency can be minimized by writing algorithms and software that execute quickly, and running that software on the fastest possible computer platforms. But, this may come at a price; some complex trading programs may require the crunching and interpretation of a vast amount of data.

On the other hand, network latency can be minimized using a number of methodologies. One of these is co-location – locating the decision-making trading engine or computer in direct physical proximity to an exchange’s server, or as close as possible to the source of a news feed. Much can be done in this area, but there are practical and external limits to this approach. For example, a news service may be located in New York, but an options exchange may be located in Chicago; alternately, the news may affect stocks that trade only in Sydney.

Sometimes a certain amount of latency is inherent to the way a network is laid out. For example, a news feed coming from another location has fixed minimum latency because of the physical distance, but it can have additional latency that varies as a function of the amount of traffic. News feeds are, by design, uni-directional multicast streams. They have to pass through many network elements, including switches and routers, on route to subscribers. These switches and routers can, on occasion, be overwhelmed by traffic even in a split second, causing undesirable queuing of data and - in extreme cases - loss of parts of the message; this network phenomenon is also known as a microburst.

Configuring a network to minimize latency is a critical goal, but is not always under the complete control of the automated trading system. Where external variables can increase latency, it is important to be able to know with great precision just how much market latency exists, for example, in an incoming data stream. It is also important for a news provider to know how much latency can occur between itself and its customers in order to be able to accurately characterize the service that it offers.

How to Track Latency and Message Loss

The most effective way to keep track of latency and message loss is to install a latency and message loss management system. These systems typically work by deploying suitable monitoring and recording probes at various points along the data path. The capabilities of these systems vary from vendor to vendor. However, there are some key features to keep an eye out for. Look for solutions that can capture and store all network data over relatively long periods of time – days, week, even months. Then, be sure that the probes can be accessed from a central server through an out-of-band network so that data comparisons can be performed among the various recording probes located throughout the network. This is the key to monitoring and measuring latency and message loss.

In addition, be sure that the probes have an accurate time-sensing mechanism so that discrepancies between clocks on the probes are greatly reduced, if not eliminated. Often times NTP (Network Time Protocol) is sufficient, but for sub-millisecond accuracy (down to 50 microseconds) a GPS (Global Positioning System) receiver installed on each probe may be required. In this case, be sure that your latency management system supports that. IEEE1588 is yet another mechanism used to provide clocking information to probes. It is perhaps the most difficult scheme to implement because it requires that the network is entirely IEEE1588-compliant.  This means that all network elements must support IEEE1588, which may be difficult to achieve.

Another way to track latency and message loss is to look for a way to be notified when an unacceptable amount of latency or packet loss occurs. The device should have user-settable thresholds and a versatile alarm system. The system should have a way of logging all instances of unacceptable latency or packet loss, as well as having the capability to run scripts. For example, if it becomes clear that data is arriving several milliseconds later than normal, the script could send an interrupt to the trading program, either to stop placing orders, or to put more weight on transaction data – quotes and volume – rather than relying solely on the news feed, which may have reached a competitors’ automatic trading system first. This would help to avert a potential scenario in which a trading engine is making trading decisions based on stale information.

Finally, deploy a system that can watch for unusual latency or packet loss events that occur only intermittently. Sometimes it just does not work to wait for an alarm trigger and then attempt to start a capture to see what is going on. The ability to capture all of the packets may be the only way to detect these intermittent events. Since problems with latency and packet loss occur more often during periods of high traffic load, be sure that your latency management system can keep up with the full network line speed without dropping any packets itself.

Summary

Automated trading can produce substantial profits, but it depends critically on the ability of your system to assess the status of the dual enemies – latency and packet loss. A right decision based on old data becomes not only the wrong decision, but also a potentially disastrous one.

If you are a news feed provider, investing in equipment that can really tell you how quickly and reliably your news feeds are getting to your subscribers is critically important.  It is perhaps one clear way that you can demonstrate your value over that of a competing news service.  If you are an automated trading firm, investing in equipment that can tell you exactly how old a piece of data is, and promptly communicate with your trading software to make the appropriate adjustments that can mean the difference between making a trading profit or not.

About the Author: Steve Wong, Vice President of Marketing, ClearSight Networks

TrackBack URI for this entry

Comments (0)

Subscribe to this comment's feed