Way back in the sixties, as more and more powerful computers were being developed, computer scientists were beginning to ask themselves whether, one day, in the far distant future, perhaps in the eighties or even beyond, computers would be "fully conversant", by which they meant, capable of acting with "human-like" intelligence.
Here in the distant future, in the time beyond such science fiction, we know that, just as computers occasionally win chess games, so algorithms are increasingly sophisticated and increasingly being deployed in trading rooms to finesse increasingly sophisticated trades. A number of trading firms are also using "machine-learning" techniques to optimise trading strategies, says Graham Miller, CEO, Marketcetera. "Machine learning techniques are good at taking a trading strategy and squeezing an extra 40% to 50% out of it," Miller explains.
But Richard Balarkas, president and CEO of Instinet Europe, traces the rise of the machines in the trading room back to decimalisation of the US capital markets in 2000, rather than back to those clunky early thinking machines. The point about decimalisation is that it resulted in smaller tick and trade sizes, causing market data volumes to explode, and the point there is that machine development changed course towards number-handling, frequency, volume, and so on, and thus away from more intuitive, interpretative approaches. As Balarkas points out, the increased data volumes were way beyond the data assimilation capabilities of a human trader, whereas machines were able to process thousands of data points per second as inputs to high frequency auto/algo models without difficulty.
So the change of direction was self-reinforcing; once those machines had proved themselves better at handling data, that's how they had to develop. The machine became an integral part of every trader's kit, allowing them to analyse vast reams of market data and deploy increasingly sophisticated trading strategies or algorithms. So the evolution had to continue, further away from that old concept of "full conversance" and on into ever-faster data exploitation. "A lot of these trading strategies are time sensitive which meant five years ago any trading opportunity in the market place was only there for a few seconds," says Miller. "Today, those opportunities disappear in under a millisecond."
Machines are essential for traders to capitalise on such opportunities, of course, and they provide a rich array of "non-human" functionality. "[They] are also better with a certain complexity of calculation, things like ETF trades where you need to calculate fair value," explains Miller. "And to the extent that many risk metrics require significant computation, you can also put together a system that manages risk without much human intervention." Given the volatility of stock markets in recent months, with swings of 700 bps to 1200 bps in major market indices and spikes in trading volumes, some even venture to suggest that "having an autopilot [or machine-based model] at the controls has been a generally safer bet than being in one guided by human hands."
The road not travelled, in this context, is the one on which desktop trading systems evolved to the point at which they could talk back, discuss trade ideas, effectively "feel" the market in the way that a human trader might, and possibly even provide feedback. That's not totally science fiction, in a world where the competition between machine-readable news feeds is beginning to hinge on the capacity, first, to react to sentiment, and secondly, to measure the possible impact of non-market news (see Automated Trader Q4 2008, cover story). Bear that in mind as we discuss whether the ongoing economic meltdown has provided an effective tipping point in favour of machines and at the expense of human traders. Have the machines "won" the crisis, and if so, would they have won bigger if they'd been more human?
Initial evidence favours the machines. "In this challenging environment, selecting the right algorithm can help traders efficiently manage risk and volatility," says Hitesh Mittal, managing director and head of algorithmic trading at Investment Technology Group. ITG's analysis of 2008 client execution data for 3.5 billion share trades also indicates that choosing the 'right' algorithm during highly volatile market conditions saved trading customers "up to 60 bps in trading costs, as compared to a more modest but still significant 20 bps of cost savings during low volatility periods".
Volatility is good, of course, and not only if you're choosing the right algorithm every time. "I have a friend who is machine trading and he has been making more or less consistent profits all through the crisis while the rest of his trading desk - very experienced people, but all manual - have had a very difficult end of 2008," says Jonas Hansbo, CEO, Tbricks. "Given recent market conditions, a lot of people have quite successfully managed to continue using their existing code without major rewrites." But not everybody agrees with that, or sees 2008 as a win for the machines. Miles Kumaresan, head of Quantitative Trading at TransMarket Group, does not believe machines outperform humans, even in volatile market conditions. What is more important, he says, is the ability of the trading model to adapt to varied market conditions dynamically. "If you used vast amounts of data to create a model that can calibrate itself and respond optimally to significant changes such as market volatility, then the model will perform a lot better," Kumaresan says.
That almost sounds like using vast amounts of data to achieve full conversance. But not quite. In terms of risk adjusted returns, Kumaresan says quantitative trading strategies have performed significantly better than the alternatives in the current climate. In this respect, he argues that the benefit a machine brings is that it enables a trader to analyse data more accurately and more quickly. Back to the numbers. Frédéric Ponzo, managing director, NET2S, says that market conditions over the last few months have supported those trading models based on volatility. However, it is not because they have got more computers doing the job that these strategies are "winning", he says. "It is because they are backing the right strategy."
Again, volatility is good, and especially so if you're using the right "volatility-friendly" model. Going back to the man/machine debate, Bruce Bland, head of algorithmic research at Fidessa, argues machines have a number of factors in their favour. "They are not emotional about gains and losses. They simply maintain the trading style they were set, and continue, regardless of the fact that they are winning or losing." They never take lunch or a break, or get a phone call. "Humans are subject to many more sources of information which may help trading, but may also hinder trading performance." And they're fast. "With the growth in smart order routing, this is becoming more important, with machines able to pick up price improvement from non-primary markets especially in volatile periods."
There's another rather more depressing argument against filling up the trading floor with warm bodies. This is a time of staff cutbacks. "There is a case for increasing automation because it increases productivity," says Ponzo of NET2S. "Machines can execute more trades, look at more stocks and contemplate more scenarios." But what does this mean for the surviving human traders? Will machines control trading as well as execute trades, much as the on-board computer HAL controlled the space ship's operations in Stanley Kubrick's film,"2001: A Space Odyssey"?
No. "Due to the chaotic nature of markets, it is hard to imagine a trading world completely devoid of humans, although their role may lie in supervising models," says Bland. Others are less bullish about the pervasiveness of machines in the trading room, particularly for those trades that do not lend themselves to automation - and there still are some. Scott Eaton, former global head of Principal Trading at ABN Amro says there is something to be said for "warm" equity trading. "Trading credit, for example, is more art than science," he says. Even now, or perhaps, especially now.
"There is always a need for the human interpretation of
information," says Balarkas, recalling an interview he gave some
years back where he was asked about the use of artificial
intelligence in the trading room. "I said we are still looking
for the natural stuff. The sentiment I was trying to get across
is that the machines are
only as good as the humans that make them." And sometimes the human programmers make erroneous judgements, misread the market or fail to anticipate certain events. While a number of high frequency statistical arbitrage models have performed well in the current climate, Balarkas points out that some models were wrecked by high levels of volatility, which no one had predicted. "Although these models were designed to accommodate rigorous back testing and highly complex Monte Carlo simulations, inevitably nobody was considering the likelihood of scenarios that hadn't happened before," he says.
Yet, today's high velocity markets require seasoned traders to leverage fully the capabilities of automated systems, says Miller of Marketcetera. He cites that famous incident last year when a reporter re-posted a 2002 United Airlines bankruptcy filing, flagging the story as new and feeding it into a wire service published on thousands of Bloomberg terminals worldwide. "Both automated systems and their human counterparts got it wrong and triggered waves of selling," says Miller. "A carefully calibrated strategy, however, fully leveraging the power of the machine, would have searched archival records and caught the anomaly before executing a trade."
Sometimes the capabilities of the machine and the human programming it are so closely intertwined that it is difficult to determine where one begins and the other ends, but as Ponzo points out, the machine is merely an extension of the trader rather than vice-versa. Yet, without machines, traders would not be able to execute complex algorithms quickly enough to capitalise on market opportunities, let alone refine them. "The other benefit I think that is often overlooked is that when you use consistent models and algorithms to trade you have the ability to go back over trades and strategies in detail and examine the effectiveness and helpfulness of various inputs within the model," says Richard Wilson, a consultant with the Hedge Fund Group. "This is very hard to do with more qualitative or subjective investment research and decision making." Kumaresan agrees, saying that machines are good at analysing the effectiveness of trading strategies. "Even if you are a skilled trader, you are not going to be able to analyse some of these things as well as a machine, which can produce a more exact estimate," he says.
It is the human/machine interaction - not the machine on its own - that ultimately determines the performance of a particular trading model. Machines are an enabler, albeit a very useful and increasingly important one, and for that reason Eaton says we are likely to continue to see the use of computer modelling for trading and risk management, as well as the increased use of algorithmic trading. And as advances in technology make it easier and more cost effective to deploy machines, they are likely to be more widely used. "It's clear that humans are using machines more and more," says Hansbo of Tbricks. "It's also becoming more of a reality as there are new powerful systems available that you can use without making huge investments in development infrastructure."
"Man versus machine" is actually a far more fluid situation that the tag implies; in fact there's migration in both directions. As the race to minimise latency approaches its zero sum conclusion, some automated traders are looking for alternatives; just being fast is no longer necessarily enough to ensure their profitability. "We've noticed an interesting shift in the past three months," says David Knox, CEO of i-traders.com. "Before, the majority of our research clients were manual traders who were mostly operating intraday, but not at high frequency. Now we've noticed automated traders wanting to take an XML feed of our trade ideas and market levels - either to trade direct or as input to their own models."
The intriguing thing is that a significant number of these automated traders had a track record in very high frequency trading but were clearly looking to diversify into higher time frames and new methods. "Some of them also wanted to take a feed from us that they could use as a manual overlay to their automated models," says Knox. "By the same token, some of our manual trading clients have also moved in the opposite direction and started dipping a toe into automated trading."
The overall I-TRADERS' perspective is that the recent market upheavals have precipitated a more agnostic approach by traders, hence this migration between camps. "In the past, traders tended to classify themselves as automated or manual and also by the time frame in which they operated," says co-founder of I-TRADERS, Shaun Downey. "Now circumstances have driven a more pragmatic and flexible approach: 'Whatever works, I'll do that!'"