Listen to this article
This is an experimental feature. Give us your feedback. Thank you for your feedback.
What do you think?
For as long as there have been exchanges for shares, currencies or commodities, investors have looked for ways to beat the markets and produce above-average returns.
Highly-skilled traders and fund managers command premium salaries for their abilities to second-guess market moves and return profits to their clients. But developments in computer science mean that automated trading systems could well outperform even experienced traders across global financial markets.
Researchers at HP’s European labs in Bristol, England have found that international financial institutions are increasingly showing interest in their work on automated trading agents - despite the fact that the agents were not originally developed for financial markets. HP Labs’ complex adaptive systems group first started working on trading algorithms in the mid-1990s. But they were originally developed to help large companies allocate computing resources in data centres.
That work looked at allocating scarce computing resources, such as CPU power, memory and network bandwidth among users’ projects. Researcher Dave Cliff, who worked at HP Labs between 1998 and 2005, developed a trading algorithm known as Zero Intelligence Plus (Zip).
In 2001, testing carried out at IBM’s Watson research labs, in New York, found that both IBM’s MGD trading algorithm and Zip beat human traders, with Zip coming out on top in the tests. This prompted interest in the technology from financial institutions. Zip is now used as a benchmark for measuring other trading algorithms.
The Zip algorithm works by calculating the best trading strategy for continuous double auctions (CDAs), the trading basis of most financial markets. In theoretical microeconomics, supply and demand curves are represented by smooth lines on a graph, with markets moving quickly to an equilibrium between supply and demand.
In a real CDA market, the supply and demand curves are stepped, not smooth, as traders have to deal in fixed units - such as a share, a barrel of oil or a unit of currency - or their multiples. And, unlike theoretical markets where buyers and sellers have perfect information, in a CDA buyers and sellers will lie to strengthen their own positions.
Earlier research into automated trading had suggested that markets, rather than traders, held most of the intelligence. This led to the development of a number of “zero intelligence” trading algorithms. But research carried out by Mr Cliff found that under some circumstances, zero intelligence trading agents would fail to reach an equilibrium. In a real exchange, this would mean market failure.
Zip traders have the ability to “learn” from their actions, using simple machine learning rules. This function allows the trading algorithms to improve their own behaviour. As a result, Zip algorithms succeed in trading where zero intelligence algorithms fail.
According to Andrew Byde, who is now managing the automated trading research at HP Labs, trading algorithms work best at the execution end of the financial services business, such as buying or selling stocks or commodities according to trading rules. The algorithms cannot, in themselves, decide on the stocks an investor should be buying or which markets they should trade in.
“The decision on stocks themselves is not the realm of the trading agent,” he says. “That depends on the underlying fundamentals of the company whose shares you are trading in. Modelling that would be extremely difficult.”
Instead, HP’s research suggests that once a fund manager has made the decision either to invest in or to divest of a particular share, automated trading agents are better placed than humans to execute the deal in the most profitable way. “Once you have taken the decision to get into or out of a stock, you have to instruct a broker to do that for you,” says Mr Byde. “That is where most automatic trading algorithms come in. Once you have taken the decision on where to put your money, there are good and bad ways to carry out that trade.
Banks, Mr Byde notes, already make significant use of automated trading algorithms: around 30 to 40 per cent of trades are based on computer-generated decisions. HP is working with a number of tier-one financial services institutions on automated trading research, although the company will not discuss the details.
However, much of the current focus of HP’s research is to make the automated trading systems more relevant to the way financial services institutions carry out their trading.
Both HP and IBM’s initial research in the field focused on ways of allocating resources to maximise the total welfare of society and traders. This works perfectly well in a closed, internal market - such as a company’s data centre - but less well in an open market where the objective of each trader is to maximise profit.
“All the studies of Zip we have done have focused on the ‘global good’ question,” says Mr Byde. “That is quite different from the perspective of a fund manager or a bank, who wants to make a good deal. But one reason the banks are interested is that if we can produce an algorithm that maximises efficiency, we can also produce one that is good at maximising profit.”
Central bankers and regulators might also be interested in using the trading algorithms to simulate markets and to model the possible impact of any market intervention.
The next phase of the research, Mr Byde says, is to make what is essentially an abstract model of markets “more concrete”, so it can be deployed by large financial institutions. But nor is the research limited to financial services.
HP has already used market-based trading to allocate computing resources among digital animators. BT is also looking at how market-based mechanisms could be used to control complex communications networks, and HP’s research into data centres is ongoing. “Our research is now focusing on making the models more interesting and complex, but still informative,” says Mr Byde.