Welcome to Futures Industry
Will Acworth
Published 8/21/2007

Download PDF Version

Algorithmic trading in the U.S. futures industry is well past its infancy, with a host of tools available for use in a wide range of markets. Traders and vendors are now increasingly looking for ways to sharpen their edge, either by making their already fast trading engines interact with the market even more rapidly, or by tailoring their algorithms ever more closely to the specific characteristics of individual futures markets. In fact, there are signs that some traders are beginning to design algorithms that prey on other algorithms, or hide their presence from other algorithms behind a flurry of order messages never meant to be executed.

It is a difficult world to track, especially since so many of the algorithms in use in the futures industry were developed by proprietary traders, who naturally are reluctant to reveal any clues to their trading strategies. The field is getting crowded, however. Several independent software vendors have developed algorithmic trading tools of varying degrees of complexity for use in trading futures, and many of the broker-dealers that dominate algorithmic trading in the U.S. equities markets have adapted their models to the specific characteristics of the futures markets. Both type of vendors see futures trading firms such as commodity trading advisors as potential customers, and they are eager to talk about their offerings.

The Futures Industry Association’s annual OpTech conference is a good measure of the trend. The 2007 OpTech conference, which took place in mid-June, had four panels devoted to algorithmic trading, and many of the speakers on those panels are involved in building and marketing algorithmic trading systems specifically for the futures industry.

“Three years ago, most of the algorithmic trading systems in the futures industry were self-developed,” Jim Johanik, head of U.S. technology for Euronext.liffe, said at the OpTech conference. “Now we see a lot more off-the-shelf products being put to use.”

Although many observers assume that algorithmic trading began in equities and then migrated to futures and options, the reality is that algorithmic trading has a long history in the listed derivatives world. More than 15 years ago, traders were writing computer programs to automate their trading on Deutsche Terminbörse, the predecessor exchange to Eurex. Electronic options trading in Europe was another important incubator for automated trading engines, as market makers developed high-speed automated systems to cancel and replace quotes across dozens if not hundreds of instruments.

This has caused some confusion over what algorithmic trading actually means. Many people in the equities world separate the investment decision from the execution process and see algorithmic trading primarily as a tool for executing trades, typically against a benchmark such as volumeweighted average price. In contrast, people in the futures world often talk about algorithmic trading as a system for automating the entire trading process, including the search for trading opportunities. Under this approach, the “black box” reads the market data, analyzes the trading opportunities, and transmits the order messages, with little or no human intervention.

In reality, both approaches exist in both markets, and advances in one market are soon absorbed in the other. Arbitrage is especially well suited to algorithmic trading, mainly because the machines can operate so much more quickly than human beings. Russell Abramson, executive director at J.P. Morgan Futures, says some black box trading firms can transmit several thousand order messages to an exchange in less than a second, constantly canceling and replacing orders as the market changes, and quickly capturing any price discrepancy as soon as it emerges.

Black Box Trading in the Energy Markets

Energy is one of the hottest areas for algorithmic futures trading right now. Especially for the high-frequency trading community. Over the past two years, the IntercontinentalExchange has invested heavily in its technology platform to make its energy futures market more attractive to algorithmic traders, improving both its speed and capacity. Once the New York Mercantile Exchange listed its crude oil contract for round-the-clock trading on Globex in August 2006, the race was on.

“As soon as Nymex moved to Globex, we started getting calls,” says Jesper Alfredsson, head of algorithmic trading at Orc Software. “The Globex platform is well known to algorithmic traders, and the arbitrage with ICE is suited very well for that kind of trading.”

Both exchanges offer a FIX-based application programming interface and both permit trading firms to locate their servers close to their matching engines. (See “Co- Location Catches On”.) This gives electronic traders rapid access to their markets, with order message round-trip times measured in milliseconds.

In June of this year, Nymex gave another push to the trend, listing options on its energy and metals futures on the Globex platform. Alfredsson sees this as especially interesting for options market makers active in other electronic options markets, who can now apply the same automated quoting technology in a class of options that up to now have traded either on the Nymex floor or in the over-the-counter markets, neither of which is suited for algorithmic trading.

Multi-Market Arbitrage and the Challenge for Co-Location

The advances in energy futures highlight one of the current challenges facing algorithmic traders who are focused on speed. Most of the major futures exchanges offer co-location services, which allows trading firms to move their servers to an access point that is physically very close to the matching engine. This reduces the distance that their order messages have to travel to reach the exchange and can significantly reduce the latency in the order transmittal process. Three of the world’s largest futures exchanges —CME, ICE and Eurex—offer this type of service, and Euronext.liffe expects to do so by the end of the year.

The problem arises when a trading firm wants high-speed access to more than one exchange. If the firm locates its server closer to one exchange, it will be farther from the other, and vice versa. Either way, there will be some latency in the connection to the more distant exchange. Even if that latency is only a few milliseconds, it can make all the difference in the world of high-speed trading.

To overcome this problem, some firms are setting up separate trading engines at each exchange, in each case co-located with the matching engine, to ensure the fastest possible access. The various trading engines are programmed to interact with each other dynamically, with overall control residing at a single location. This requires an analysis of “network performance” according to Tayloe Draughon, vice president of e-trading IT strategy at Calyon Financial.He explains that firms are measuring the time it takes for an order message to travel from various points in the network and the internal latency of each exchange, then factoring thosemeasurements into their trading models.

Tuning the Engine

Co-location is hardly the only way that high-frequency traders gain an edge over the rest of the trading community. The players in this game have to make an all-around analysis of latency in the trading infrastructure to see if there is any part of the system where a few milliseconds can be shaved off. This is not a one-time exercise. Several speakers at FIA’s OpTech conference said trading firms have to monitor the performance of their networks on a real-time basis, and adjust their trading models for spikes in the latency of the network during the course of the day. In the old days, a trader might be watching four screens at the same time, commented one speaker. Now it’s two rows of four, with one row displaying the markets and the other row displaying network conditions.

Calyon Financial’s Draughon cites “garbage collection” as an example of how much attention trading firms are giving to these issues. In both C++ and Java-based programs, the systemruns a routine at certain intervals to “clean house” and reassign memory to the processing functions that are running at that time. This routine causes a very slight slowdown of the system, too small to be noticed by most users, but definitely noticeable for an automated trading machine that measures time in milliseconds, especially if it happens to coincide with a peak in market activity. This is one argument for using the C++ language, he says. In C++ the trader can control the timing of the routine, rather than leaving up to the machine to decide.

“It’s all about fine-tuning the performance of the system,” says Draughon.

From Software to Silicon

One of the effects of algorithmic trading is an explosion of market data. As traders use these tools to slice up their orders into smaller pieces to reduce market impact, it increases the number of trades. Average order size in the E-mini futures markets, one of the first U.S. futures markets to go electronic, has fallen to approximately two contracts, and many market participants are expecting the same result sooner or later in other futures markets. At the same time, the development of high frequency automated trading machines has contributed to a sharp increase in the order to fill ratio.Of the thousands of order messages that these machines transmit in a single second, only a few actually result in a trade.

What makes this flood of market data especially problematic is that many traders using algorithmic trading systems want every available piece of information. Just getting the bid and the ask is not enough; they need the full market depth with price and quantity, including every order resting in the order book and updated on a real-time basis. In addition, they want the data stored, processed, and available for analysis on a real-time basis. The faster their machines can receive the data, the faster they can identify a price discrepancy and act on it. This in turn creates a feedback effect. Each time a price change is transmitted by an exchange, the automated systems respond by sending a new batch of order messages, and the cycle starts all over again.

The bottom line is that exchanges are sending out far more market data than ever before. Euronext.liffe’s Johanik estimates that his exchange is now sending out four times as much data as it was two years ago, and in December Eurex quadrupled the capacity of its data feed to 1096k. Database vendors say their systems have to be capable of receiving many hundreds of thousands of order messages per second, and the one-million- message-per-second mark is not far around the corner.

Philippe Buhannic, president and chief executive officer of trading technology vendor TradingScreen, sees the increase in market data as one of the most critical issues facing the industry over the next several years. “Algorithmic trading is really quite small at present in the futures industry,” he says. “We are going to see a lot more trading by these systems, and it is going to have a very large impact on how we manage market data.”

Buhannic points to the U.S. equity options market as an example for the futures industry. Quote volume is rising extremely rapidly, and is expected to continue rising as algorithmic trading of options increases, especially if penny pricing is extended to additional series. The International Securities Exchange, the largest U.S. exchange for equity options, says market maker activity can result in up to 14 changes per second in the order book for a single option at a single strike price. Multiply that across all strike prices in all the different options series, and the result is a torrent of market data. The Options Price Reporting Authority currently has the capacity to transmit 573,000 market data messages per second and expects to raise that to 700,000 in January.

The futures industry is not yet generating message traffic at that level, but the exchanges are gearing up to transmit significantly more data at a higher speed. Eurex, which began transmitting real-time un-netted market data late last year, plans to provide additional market depth when it implements the next version of its trading technology platform in November, and ICE is now betatesting a high-speed data feed called Impact with several of its market participants. In addition, a number of firms are experimenting with technology to see if there are ways to dramatically improve their ability to process market data. One solution is to move the processing functions from software to silicon by using “field programmable gate arrays.” This is a type of computer chip that can be programmed on a very basic level to handle specific tasks. That is not a simple task, but firms using this technology report a huge performance jump. Activ Financial, a market data vendor used by a number of firms in the futures industry, estimates that using FPGA chips allows it to process 20 times the amount of data and reduce latency by 10 times compared to its previous software- based solutions.

TradingScreen’s Buhannic warns that it is very expensive to build systems capable of handling the torrent of market data necessary for algorithmic trading. “Only the top players can spend the money to create their own tick-by-tick database,” he says. “It is not enough to have a great model. You have to have the order management, connectivity and market data infrastructure to support that model, and that is becoming more difficult to do in-house.” As a result, he expects a handful of large banks and trading firms that have the resources to develop this infrastructure to gain an edge.

Buy vs. Build

The focus on technology raises an obvious question for everyone participating in these markets: is it better to buy the technology from a vendor, or build it oneself? A few years ago, many of the off-the-shelf solutions were designed mainly for the equities markets and required a lot of adaptation for use in the futures markets, according to several speakers at the OpTech conference. That is becoming less and less true, however, as the developers of algorithmic trading solutions have expanded into other asset classes.

Credit Suisse is one example of this trend. The firm’s Advanced Execution Services division is considered one of the leading providers of algorithmic trading tools in the equities world, but only recently has it recalibrated its tools for futures. For instance,many algorithms rely on statistical models to anticipate patterns in trading volume and determine the optimal timing for an order. An application in futures tradingmust take into account the special conditions that occur around expiration, when volume surges as positions are rolled into the next front month.

Andrew Yao, who heads sales to the futures industry for Credit Suisse’s AES division, says the firm spent two years building models for the futures markets, and now is focusing on marketing these algorithms to the futures industry. He says they have been designed for use not only with equity and interest rate futures, but also with commodity futures markets such as crude oil and corn. Equity traders that also trade futures naturally are among the first to adopt these tools, he says. For example, a mutual fund that uses S&P 500 futures might use an algorithmic trading tool to minimize the market impact of its trades. But he says Credit Suisse is also having success with commodity trading advisors, who see these algorithms as a way to improve the efficiency of their trading desks.

Another example is Goldman Sachs, which has retooled its Rediplus platform to make it more attractive for futures trading. Jana Hale, global head of algorithmic trading at Goldman Sachs, says the firm is responding to several broad trends. Institutional investors are more aggressively using derivatives products, including futures, and they are looking beyond equities to other asset classes, such as energy and other commodities. At the same time, the migration from floorbased trading to electronic trading has made more futures markets accessible for algorithmic trading, especially in the commodities area. As a result, Hale says the firm’s client base, predominantly asset managers and hedge funds, are increasingly using the Rediplus platform to trade non-equity futures. Goldman Sachs now offers 12 futures algorithms on this platform, she adds. The latest addition is a pegging algorithm. Rather than aiming for a specific price or benchmark, this algorithm targets the bid, the ask, or the midpoint, and continues to enter orders until the trade is completed. For example, a user wanting to buy 1,000 E-mini futures might tell the algorithm to display orders of five contracts at a time on the bid until the trade is completed.

The danger, of course, is that relying on a vendor for algorithmic trading solutions may expose a firm’s trading strategies to the vendor. “Information leakage” is a critical concern for hedge funds, commodity trading advisers and institutional investors worried about someone copying their trading strategies or front-running their orders. That is why the investment banks that market algorithmic trading solutions say they take great pains to separate their algorithmic services from the rest of the firm. At Credit Suisse, for example, AES client orders are routed through completely separate systems that have been examined by an independent auditing firm to ensure absolute anonymity, says Yao.

Rus Newton, co-head of Global Advisors, a commodity hedge fund based in London with about $150 million in assets under management, agrees that information leakage is a concern, but says the investment banks are well aware of the damage this would do to their reputations. In his view, the real issue is more a question of resources.

“Historically most commodity trading advisers have tended to be run like boutiques, with five to 25 people. So they haven’t had a massive budget to spend on algorithmic trading,” Newton said in an interview. “So the banks are thinking that if they spend a few million on building better algorithmic trading systems, they can differentiate themselves from the other brokers and get more of the order flow from the CTAs.”

Newton sees a problemwith this approach, however. CTAs live or die by their ability to generate profitable returns on their trading strategies, and typically emphasize the unique qualities of their strategies. “If you are telling investors that your trading system is truly unique, are you shooting yourself in the foot by using someone else’s algorithms?” he questions.

Phantom Liquidity

There are two sides to every story, of course. Just as the proprietary traders worry about predation by the investment banks, the investment banks are hearing complaints from their customers about predatory behavior by algorithmic traders. J.P. Morgan’s Abramson says there has been a noticeable increase in complaints about “phantom liquidity” over the last six months, and he sees this as resulting from the impact of algorithmic trading on the futures markets. “What is happening is that customers will see a quote displayed in the order book that they want to hit, but by the time they send in their order, the quote is gone,” he explains.

The suspicion among some firms, according to several other participants at FIA’s OpTech, is that some of the algorithms are programmed to coax liquidity into themarket, a practice called “fishing.” A firm will display an order on one side of the market, knowing that this will trigger a response from algorithmic traders, and then cancel the order and hit the other side of the market. This might get a slightly better price or more volume at the same price than was initially displayed.

This kind of behavior is hardly new to the futures industry. Itmay happen a lot faster than before, but industry veterans remember similar gamesmanship when trading took place on the floors. When one of the legendary figures in the business entered the pits, the other traders watched for any clues to their trading intentions. In order to disguise their moves, they would have their real trades executed by other brokers on the other side of the pit.

It does illustrate, however, why the development of algorithmic trading tools is often compared to an arms race. As soon as a particular algorithm is widely used, the trading community immediately creates a new algorithm that takes advantage of the predictable patterns of the older algorithm. Some vendors say they are now on their “third generation” of algorithms, and no doubt there will be many more as traders continue their never-ending search for an edge.

Will Acworth is the editor of Futures Industry magazine.
Change text size
Print this article
Email a link