I am a former
I am also a former and current developer of exchange technology. The exchange is the focal point of HFT, where electronic buyers and sellers match in a complex web of systems and networks to set the price for assets around the world. I would argue that the computational challenges of developing and maintaining a competitive advantage in the exchange business are among the most difficult in computer science, and specifically systems programming. To give you a feeling of scale, the current exchange technology is benchmarked in nightly builds to run a series of simulated market data feeds at 1 million messages per second, as a unit test. There is no such thing as premature optimization in exchange development, as every cycle counts.
The goal of this article is to introduce the problems on both sides of the wire. Today a big Wall Street trader is more likely to have a Ph.D from Caltech or MIT than an MBA from Harvard or Yale. The reality is that automated trading is the new marketplace, accounting for an estimated 77 percent of the volume of transactions in the U.K. market and 73 percent in the U.S. market. As a community, it's starting to push the limits of physics. Today it is possible to buy a custom ASIC (application- specific integrated circuit) to parse market data and send executions in 740 nanoseconds (or 0.00074 milliseconds).4 (Human reaction time to a visual stimulus is around 190 million nanoseconds.)
In the first of the other two articles in this special section on HFT, Sasha Stoikov and Rolf Waeber explain the role of
In the other article, Stephen Strowes discusses a method for estimating RTT
When I began in HFT, it was a very different world. In 2003, HFT was still in its infancy outside of U.S. equities, which two years earlier had been regulated into decimalization, requiring stock exchanges to quote stock prices in decimals instead of fractions. This decimalization of the exchanges changed the minimum stock tick size (minimum price increment) from 1/16th of a dollar to $0.01 per share. What this meant was that "overnight the minimum spread a
This led to an explosion in revenue for U.S.
By 2005, most shops were also modifying kernels and/or running realtime kernels. I left HFT in late 2005 and returned in 2009, only to discover that the world was approaching absurdity: by 2009 we were required to operate well below the
1. Receive a packet at the network interface.
2. Process the packet and run through the business logic of trading.
3. Send a trade packet back out on the network interface.
To do this, we used realtime kernels with bypass drivers (either InfiniBand or via Solarflare's
Onload technology). At my shop, we had begun implementing functionality on the switches themselves (the Arista switch was Linux based, and we had root access). We must not have been alone in implementing custom code on the switch, because shortly after, Arista made a
As with all great technology, using it became easier over time, allowing more and more complicated systems to be built. By 2010, the barriers to entry into HFT began to fall as many of the more esoteric technologies developed over the previous few years became commercially available. Strategy development, or the
I remember coming home late one night, and my mother, a math teacher, asked why I was so depressed and exhausted. I said, "Imagine every day you have to figure out a small part of the world. You develop fantastic machines, which can measure everything, and you deploy them to track an object falling. You analyze a million occurrences of this falling event, and along with some of the greatest minds you know, you discover gravity. It's perfect: you can model it, define it, measure it, and predict it. You test it with your colleagues and say, 'I will drop this apple from my hand, and it will hit the ground in 3.2 seconds,' and it does. Then two weeks later, you go to a large conference. You drop the apple in front of the crowd...and it floats up and flies out the window. Gravity is no longer true; it was, but it isn't now. That's HFT. As soon as you discover it, you have only a few weeks to capitalize on it; then you have to start all over."
What follows is a
The first step in HFT is to place the systems where the exchanges are. Light passing through fiber takes 49 microseconds to travel 10,000 meters, and that's all the time available in many cases. In New York, there are at least six data centers you need to collocate in to be competitive in equities. In other assets (foreign exchange, for example), you need only one or two in New York, but you also need one in London and probably one in Chicago. The problem of collocation seems straightforward:
1. Contact data center.
2. Negotiate contract.
The details, however, are where the first systems problem arises. The real estate is extremely expensive, and the cost of power is an
Once the servers are collocated, they need to be connected. Traditionally this is done via two methods:
The Cross Connect and NAT. Inside the data center are multiple exchanges or
The key to the internal NAT problem is the core nature of trading: correlated bursts. These bursts are what make HFT networks so ridiculously difficult. Figure 2 is a screenshot of an internal monitoring application that shows the
As the world becomes more interconnected, and assets are more closely linked electronically, these bursts can come from anywhere. A change in U.K. employment will most certainly affect the USD/ GBP rate (currency rates are like relative credit strength). That in turn affects the electronic U.S. Treasury market, which itself affects the options market (the
WAN Links. Outside the data center the systems need WAN links. Traditionally HFT shops ran two sets of links, as shown in figure 3: a
The feed handler is often the first bit of code to be implemented by an HFT group. As shown in figure 4, the feed handler subscribes to a
The tickerplant is the system component responsible for distributing the
In addition to managing
In my experience, most
To begin, let's review some jargon. Figure 6 shows the first two levels of the order book. The order book is split into two sides: bids (prices at which people are willing to buy); and asks or offers (prices at which people are willing to sell). A queue consists of the individual orders within a price.
Orders are executed in a
Trading rates are somewhat difficult to estimate, but there is a clear relationship between the probability an order will be executed and the ratio of its queue size to the opposite queue size (e.g., the bid queue size versus the ask queue size). This is shown in figure 8 as pUP.
If trading rates are hard, then cancel rates are even harder. The question is, given your place in the queue, what is the probability a cancel will come from in front of you (thereby allowing you to move up)? This is very difficult to estimate, but our own trading and some historical data provides the basis for an engineering estimate.
To start, we know that if we are in the back of the queue, the probability of a cancel coming from in front of us is 100 percent. If we're in the front, it's 0 percent.
Figure 9 is a chart of the empirical percentage of time a cancel comes in front of your order; it is a function of the percentage of orders that are behind you in the queue. For example, if you are at 0 on the
We often say you get to the front of the queue via two methods: promotion, which is a
If your order is executed on a large queue, then you have a free option to collect the spread. If one of your orders on the opposite queue is executed, then you collect the difference between the price at which you bought the asset (bid side of $9) and the price at which you sold the asset (offer side of $10).
In the event the queue you were executed on gets too small, you can aggress the order that was behind you. This means crossing the bid/ask spread and forcing a trade to occur. If you get executed passively, you are aggressed upon by another order sitting on a queue. As long as another order is behind you, you can unwind the trade, meaning you can aggress the order behind.
Aggressively unwinding a trade is called scratching the trade. You didn't make a spread; you didn't lose a spread. It's a
An exchange is the collection of systems that facilitate the electronic execution of assets in a centrally controlled and managed service. Today, exchanges are in a fight to offer faster and faster trading to their clients, facing some of the same latency issues as their newer HFT clients.
For exchanges, collocation can be an invaluable source of revenue. The more incumbent exchanges run their own data centers (such as NYSE and Nasdaq), and customers pay collocation fees to the exchanges directly. Newer exchanges must collocate in
When exchanges operate inside
Exchange networking is as challenging as HFT networking but also has a deeper focus on security. Information arbitrage, or the practice of gaining information about a market that is not directly intended for the recipient, is a major concern. An example of information arbitrage is when an exchange participant "snoops" the wire to read packets from other participants. This practice is easily thwarted with deployment of VLANs for each client.
Most exchanges still deploy 1GbE at the edge. This is both a function of history (change management in an exchange is a long process) and practicality. By putting 1GbE edge networks in place, the exchange can protect itself somewhat from the onslaught of messages by both limiting the inbound bandwidth and adding a subtle transcoding hit. For example, a 1GbE Arista 7048 has a
The gateway is the first exchange subsystem that client flow encounters. Gateways have evolved over the years to take on more responsibility, but at the core they serve as feed handlers and tickerplants. The gateway receives a client request to trade (historically in the FIX format, but as latency became paramount, exchanges have switched to proprietary binary protocols). It then translates the request to a more efficient internal protocol, and routes the request to the appropriate underlying exchange matching engine.
The gateway also serves as the first line of defense for erroneous trades. For example, if a client attempts to buy AAPL for $5,000 (whereas AAPL is offered at $461), the gateway can reject the order.
Traditionally the Order Gateway (which receives client requests to trade) and the Market Data Gateway (which distributes
Gateways are often shared across customers, as a gateway for each and every exchange participant would likely require a massive
A is connected to two distinct gateways. In 11b, Client A induces extreme load on Gateway 1, causing Client B traffic to slow. In 11c, Gateway 1, not under load, slows all attempts for Client B to cancel resting markets. Client A has an advantage with the
The matching engine is the core of the exchange, and like its HFT cousins is fairly straightforward. A matching engine, shown in figure 12, is a simple queue management system, with a queue of bids and a queue of offers. When a customer attempts to execute against a queue, the exchange searches the queue until it reaches the requested size and removes those orders from the queue.
The difficulties arrive in determining who receives notifications first. The aggressing party does not know (for certain) it has traded 10,000 shares until receiving a confirmation. The passive parties do not know they've been executed until they receive a confirmation. Finally, the market as a whole does not know the trade has occurred until the market data is published with the new queue. Problems such as this are becoming increasingly more difficult to solve as we move from milliseconds to microseconds to nanoseconds.
The world of
1. Arista Networks. 7124FX Application Switch; http://www.aristanetworks.com/en/products/7100series/7124fx/7124fx-development.
2. Arista Networks; 7150 Series 1/10 GbE SFP
3. CBS News. 2010. Robot traders of the NYSE. Sixty Minutes Overtime; http://www.cbsnews.com/video/watch/?id=6942497n&tag=related;photovideo.
4. Millar, M. 2011. "Lightning fast" future traders working in nanoseconds. BBC News; http://www.bbc.co.uk/news/business-15722530.
5. Moyer, L., Lambert, E. 2009. Wall Street's new masters. Forbes (Sept. 21):
6. Nasdaq. 2013. Nasdaq
7. Nasdaq. OMX
8. Nasdaq. 2012. OUCH Version 3.1; http://www.nasdaqtrader.com/content/technicalsupport/specifications/TradingProducts/NQBX_OUCH3.1.pdf.
LOVE IT, HATE IT? LET US KNOW
Jacob Loveless is the CEO of Lucera
© 2013 ACM
Originally published in Queue vol. 11, no. 8—
see this item in the ACM Digital Library
Tyler McMullen - It Probably Works
Probabilistic algorithms are all around us--not only are they acceptable, but some programmers actually seek out chances to use them.
Kate Matsudaira - The Science of Managing Data Science
Lessons learned managing a data science research team
Phelim Dowling, Kevin McGrath - Using Free and Open Source Tools to Manage Software Quality
An agile process implementation
Ivar Jacobson, Ed Seidewitz - A New Software Engineering
What happened to the promise of rigorous, disciplined, professional practices for software development?
Displaying 10 most recent comments. Read the full list hereFor those who oppose HFT, and are old enough, remember the 1/8 increments in bid ask? Today trading is much cheaper, both in commissions and and spread. If you are really a buy and hold investor, why are you quibbling over pennies? HFT's are like the guys in Superman who steal fractional pennies from savings accounts.
By the way, these machines don't read headlines. News is just noise. CNBC viewership is down because of people's experience following "expert" advice. There are only so many suckers born every minute.
So HFTs have a bit of overhead ... rent, electricity, salaries ... and beyond that this HFT business generates a lot of profit while benefiting the world not one bit. They make money but they create nothing, neither money nor goods nor services nor happiness nor love. Where does this money comes from? Some of their dollars might be mine, or yours. Or it could be from the FED and this is just a distribution system that effectively raises the velocity of money thereby increasing the GDP. That's good?
Are we going to start passing judgment now on who should make money? Tobacco, gambling, alcohol and prostitution ... Are we going to stop them?
most larger firms (over 50 traders) have some public ip's we dont advertise to the internet, but own with arin...
i cant think of one trading firm doing hft that uses nat... just not a very good technology... while the newer arista and Nexus 3548's do "hardware nat" I have not had to implement it for anyone... weird...
And isn't the market supposed to facilitate "investment"?
There's big money behind HFT so I don't expect regulation anytime soon. More trading by itself improves Wall Street's bottom line. But how bad will an HFT-triggered crash have to be before we effectively outlaw it?
Firms still go public and shares are still exchanged.
Regarding social value: That is subjective. One could ask, exactly what socially redeeming value do those who receive government largesse supply? For many, the answer is also "none". Yet they do, and will continue to do so.
It feels to me that HFT is more about electronic engineering or even physics instead of computer science, as the biggest challenge is to implement the simplest algorithm as fast as possible (which usually is done with hardware).
Displaying 10 most recent comments. Read the full list here