There is a buzz around high frequency trading, to put it mildly, which attracts both envy and suspicion. "High frequency is trading at speeds that exceed human manual and/or visual control," says Peter van Kleef, CEO of Lakeview Capital Market Services. "While commercial software, like broker execution algorithms, is now catching up, it's nowhere close to ultra high frequency levels. We're talking thousands of messages a minute."
For some firms it is indeed fast and furious. "Currently we run roughly a thousand orders per second on our high frequency platforms in several proximity data centres across Europe," says Mark Holt, head of systematic implementation at BlueCrest Capital Management, "so there may be faster players, but it's fast enough for us."
Peter van Kleef
Fast, but not simple. "All high frequency algos have to adapt dynamically to real time flows," adds van Kleef. "So you not only react in milliseconds to an earnings report and compare it to analyst forecasts, but you look at the market reaction in the first few seconds of trading as well and co-ordinate it with other events."
Kee-Meng Tan, head of the European electronic trading group in Knight Equity Markets tries to put it into context. "The alpha horizon for high frequency traders is in milliseconds to seconds; for day traders it may be minutes to hours, for hedge funds it could be hours to days or weeks, for mutual funds it might be weeks to months," explains Tan. "It is the diversity of flow and variance in investment horizons that facilitate efficient matching of trades, and resilience of markets. Without this diversity, there would be significantly less liquidity in the market."
This is significant. The diversity within which high-frequency trading occurs is positive. "In most exchange and MTF displayed order books today," confirms Adam Toms, head of the market access group at Nomura, "the top five bid and offers will, in part, be made up of high frequency liquidity, you can't get away from that when over 40% of European volumes are related to some type of high frequency trading."
% of US Equities Volume from HFT and by Segment
Source: Celent, a member of the Oliver Wyman Group
The stakes are high. "Our platform is fully automated," says Tan at Knight. He explains how ten years ago in the US they had over 250 traders doing what they are doing now with a room full of machines and 90% fewer traders while handling 10 times more flow, and covering a much larger universe of names.
For others, alpha and risk are the drivers. "High frequency will diversify our strategies," says Rafael Molinero, CEO and founder of Molinero Capital Management. "Low frequency plays are often directional. With high frequency there's less directional bias, but there is however a constraint on capacity. These different behaviours would add diversification to our portfolio."
Composition of US Equities HFT Flow (3Q 2009: 6 billion shares)
Source: Celent, a member of the Oliver Wyman Group
No secret sauce
How do you succeed at high frequency trading? "There are fundamentally two sorts of model: mean reversion and trending, but there's no secret sauce," says Stuart Theakston, head of high frequency trading at GLC. "Often the alpha comes as much from how you implement it as the strategy itself: the technology, risk management, or testing approach." Theakston insists that the robustness of the overall architecture is vital, adding, "We're connected to more than a dozen European exchanges. There's plenty of potential for connectivity problems. How you cope with these problems makes a difference. We constantly have to challenge the market data, and ask what's normal or abnormal. It's not straightforward."
Fast technology is of course essential. "With the advent of smart order routing the ability to transmit and amend child orders between venues as fast as possible is essential," says Toms, explaining why Nomura purchased a pure dark fibre network ring to ensure they have the lowest latency solution possible to transmit order flow between liquidity venues.
Yet sometimes even the speed of light is not fast enough. "We're currently looking at colocation at the exchange," says Holt at BlueCrest, "which would bring access down below a millisecond. The latest multi-core chips will support our more sophisticated statistical strategies in a single box." He sees this as making colocation in restricted space feasible, but admits that it is very hard to model the benefits. "Once we roll out our new trading platform we'll try it and see," he concludes. "At these speeds you're never quite sure, in pure scientific terms, why you may be successful. Are you just faster, is there some other time effect or are your strategies better? To some extent you have to stay ahead of the pack just to ensure you keep your alpha."
Nevertheless, hardware alone is not the answer. "The latency between proximity data centres and markets keeps falling as brokers and communications companies compete," notes Holt. "This prompts you to update your software to exploit it." For example BlueCrest is moving from a two-tier architecture with daily trading algorithms driving high frequency engines close to the market, to three-tiers with an intermediate speed engine in the middle and potentially faster servers in colocation centres at the exchange. "This will allow us to do more cross-market or cross-asset trading," says Holt, "but you need the extra cores to do the calculations and to split the software models as well. It's a continuous learning process."
Theakston sees the software as key. "While big firms have a technology advantage, small firms can opt for niche strategies," he argues. "Technology costs are also coming down. Two guys in a garage can now compete. The markets are much more open than they ever were."
Facing up to the risks
There are many high frequency risks. Tan at Knight explains, for example, how you constantly have to monitor technical risks: a line goes down, data feeds are delayed, or outliers need to be ignored. "Our clients are very sophisticated and constantly monitoring us as well, questioning and challenging," says Tan. "The quality of your people is the real differentiator."
Then there are those unexpected regime changes. "Most algos must cope with at least two regimes, fast and normal markets," says van Kleef at Lakeview, "and switch dynamically between them. However, some traders will identify many different market conditions and react accordingly. You need to change your filters and adjust your signalling."
This means that at the very high end, everything possible is pre-calculated and you look it up on the fly. "There's no time to do calculations of even modest complexity, not even with graphics cards with hundreds of cores, which are also used," says van Kleef.
Theodoros Tsagaris, Senior Strategist, GSA Capital, will soon be publishing some academic research on these challenges. Speaking personally he says, "Alpha decays very quickly, so you have to act promptly on new information. You have to think in terms of good enough models when working in microseconds." Tsagaris uses techniques from machine learning, statistics, computing and engineering disciplines. "There has been lately a growing interest in the cross fertilisation of techniques from these areas," he says, explaining how inversion of large matrices can add latency to the system. "There are few numerical techniques that handle recursive computation. Batch approaches for estimation are best avoided. For large datasets, online descent approaches are much faster and converge under some conditions."
At least the capital commitment is somewhat reduced in the high frequency space. "The higher the frequency the less capital you actually need to deploy," says Theakston at GLC. "Moreover, these strategies generate more alpha than capacity. Often, the more capital you deploy the lower the return. These are niche opportunities."
So it's back to backtesting
"Models of course are constantly changing," says Theakston. "They work for a while, but then there's a regime shift, and you have to find a new idea." Theakston argues that as the mix of participants and strategies in the markets changes, so you adapt. "The speed of developing new models is key," he concludes.
As markets accelerate and more participants trade faster, conditions change, observes Holt at BlueCrest. "Backtesting with last year's data may no longer be relevant," he argues. "This leads to a need for continual benchmarking."
Yet markets are not deterministic, as Theakston points out. "So all the backtesting in the world doesn't prove anything. In different market conditions - a regime shift - it can all change." Still he notes that high frequency strategies give you quick confirmation if they are working in real markets. "What a backtest does tell you," explains Theakston, "is how long you might lose money for before you need to worry. If you're worried, then you need to find an explanation. It might be a technical failure or perhaps some unexpected market news. However, if you can't explain draw downs then you question the strategy itself. "
"For the players who have invested in low latency access to the markets, their competitive edge has shifted to finding better ideas for systematic trading algos," says Pierre-François Filet, CEO and co-founder of QuantHouse, a provider of end to end systematic trading solutions. He sees the bottleneck as backtesting. "One of our hedge fund clients, AlgoDeal, has therefore developed with QuantHouse a massive simulation grid to rapidly backtest algorithms against terabytes of market data and then benchmark their performance. They are making it available to independent quants using an innovative profit sharing scheme."
Filet believes this could lead to a rich harvest of new ideas and challenge many of the biggest players who rely on their deep pockets and dozens of in-house PhDs for competitive edge. "By comparison AlgoDeal could mobilise thousands of PhDs with equivalent tools," says Filet. "It's very exciting."
Filet also believes that these kinds of systems could open up a range of new electronic services for traders to analyse market behaviours and facilitate model creation and calibration. "The real limit is not the number of quants," says Filet, "but the compute power to test thousands of models with thousands of parameters. Like the aerospace industry, the key to innovation lies in simulation. These services can then be offered across extranets like our own to radically reduce lead times."
Arms race or democracy?
With such dramatic speed and innovation, is it any wonder that high frequency trading has stirred deep passions and attracted the interest of regulators? "We are likely to be reviewing high frequency trading as part of our 2010 MiFID review, along with other market developments like dark pools and broker crossing networks," says Tim Binning, policy officer for the securities markets unit at the European Commission Internal Markets. "We support innovation and competition but we always need to keep in mind their effect on the overall efficiency of markets and the general interests of investors in the EU."
"We are also aware that in the US the SEC is consulting on high frequency issues," says Binning, "and we will watch with interest the results of that consultation. In Europe people have urged us to consider both the benefits and risks of high frequency trading. Currently we have an open mind, but we will clearly want to explore those issues where people have expressed some concerns, such as risks to trading systems from high messaging volumes, potential rogue trading engines, controls on sponsored access and fairness issues such as what impact high frequency trading is having on other market users and investors. We also sometimes hear the suggestion that high frequency trading on lit venues is contributing to increased use of dark pools."
Yet many high frequency traders see more systemic opportunities than risks. "By definition, the ultra high frequency players are not really directional in the traditional sense, but rather arbitraging discrepancies or disseminating information from one market to another," says van Kleef at Lakeview. "They just add liquidity and predictability for the rest. The world is a lot more efficient for being more connected. Of course we have to have circuit breakers and other suitable safety features to prevent self-destructive shocks, but the market is still learning. It's a natural evolution."
Dark Liquidity vs. Overall US Equities Volume
Source: Celent, a member of the Oliver Wyman Group
"Even long only traders are now using low latency smart order routing from the brokers," points out Theakston, "but once these algos were the high tech property of a few specialists. It's all part of a continuum. MiFID fragmented the market, high frequency trading puts it back together again. High frequency techniques will eventually be commoditised as well, and the leading edge moves on. Fees are down, competition up. Technology has been good for the institutional investor."
Tan at Knight insists that high frequency flow in fact provides a bridge between other types of trading activity. He warns that regulation could lead to unexpected consequences. "The short selling ban in financial stocks instituted by regulators in the fall of 2008 forced high frequency traders and electronic market makers out of the market in these names," says Tan. "This in turn, caused spreads to widen and liquidity to plummet, but only in stocks affected by the ban. While inadvertent, the regulators in fact created a real life experiment of what a market looks like without high frequency trading, and it is not a market that any type of investor should hope or expect to see again. We hope we don't see the same mistake again with the institution of excessive restrictions on high frequency trading activity."
Clearly, with the world economic outlook still uncertain, the stakes have never been higher.