Here is an example of a “neural network.” I suspect that there will be many improvements down the line, but certainly the financial markets are trying to move away from historical data with multivariate analysis on regressions, to real time, data adjustment.
Traders in financial markets know when to cut their losses, or to push aggressively for a deal. Now a new Gordon Gekko-like software trader can do the same.
By becoming more or less aggressive as market conditions change, it can be 5% more profitable than existing trading programs.
Trading software is increasingly replacing human traders in markets such as foreign exchange because programs can react to market events faster.
However, they typically react to market conditions at any moment as if they were static, says Krishnen Vytelingum at the University of Southampton, UK.
In reality, markets like the NASDAQ are very dynamic with frequent, sudden changes in prices and trading behaviour. If trading algorithms could recognise the dynamics of the market they could become more profitable, he says.
Vytelingum has developed a new trading program that can adjust how aggressively it trades to match market conditions, together with Southampton colleague Nick Jennings and trading algorithm pioneer Dave Cliff at the University of Bristol, also in the UK.
Tough tradingIf acting aggressively, the agent will sacrifice more to beat a competitor to a deal, just like a human in an online auction. “If you are on eBay and you really want to get something, you would bid higher and higher, rather than passively waiting for a good price,” says Vytelingum.
The new agent responds to market changes in two ways. Firstly, it can change its aggressiveness by monitoring other traders’ behaviour.
If other traders are being aggressive – for example, by attempting to undercut others – it raises its game to trade even more aggressively. If trading is less competitive, the software acts less aggressively and calmly aims for the biggest profits available.
Secondly, the software can also use past market trends to try to forecast future conditions. If a period of volatility seems likely, the software changes its behaviour more frequently, meaning it is more likely to be ready to exploit any sudden switches in conditions.
“The majority of share trading in Europe is now handled by algorithms,” says Richard Balarkas, CEO of Instinet Europe a leading algorithmic trading firm. Trading software that is able to read and respond to market behaviour like a human is very desirable, he says.
Secret strategiesIn fact some firms may already be using software like Vytelingum’s, says Balarkas. “They don’t tell anyone about it, that’s how they make money.”
New strategies cannot be tested properly in a real market because it is impossible to know everyone else’s hand. Simulated markets provide a more comprehensive test, a method Vytelingum used to develop his program.
“We benchmarked it against the best agents,” he says. Not only could it make 5% more profit on average then more conventional agents, but when all traders used the strategy, the whole market was more efficient.
Almost all of the available market profit was obtained. “This means that as a group everyone benefits,” says Vytelingum.
If the software works this well, the researchers should probably keep it to themselves, says Balarkas. “The real acid test here is whether or not the next time Dave Cliff calls me, it is from his 60-foot yacht.”
Journal reference: Artificial Intelligence (DOI: 10.1016j.artint.2008.06.001)