From your social media and music streaming accounts to ride sharing and online news, algorithms are processing your personal information, deciding what you will see and experience next based on digital rules.
The new breed of machine learning or artificial intelligence (AI) based algorithms behind everyday services are far more complex than any program we knew a decade ago. These algorithms are being put to work in more than just your Facebook feed, their lightning-fast processing and iterative capabilities are giving an edge to high frequency traders by analysing and acting on financial market data.
Whilst retail traders may be familiar with basic trading algorithms such as the stop-loss facility, which buys and sells a security at a predetermined point, high frequency traders use more adaptable algorithms to buy and sell securities without direct human input.
Some of these algorithms enter and exit a position in under a millisecond to profit from tiny spreads in the market; some analyse the relationship between securities, waiting for small deviations to trade on; others focus on ‘momentum trading’ by identifying short term trends and market movements. However, these capability have also been used to scam and ‘spoof’ markets into thinking prices and volumes are not what they are.
Whilst the use of trading algorithms is nothing new, the complexity of new programs has reduced the transparency with which they operate. This complexity presents several risks for financial markets, perhaps the most dangerous of which is known as the algorithm contagion. Think the US DOW’s ‘Flash Crash’ in May 2010, where the index fell 9.2 percent in seconds.
Algorithms used for momentum trading are particularly susceptible to contagions. By focusing on trends in the market rather than the intrinsic value of the underlying asset there is the potential for these algorithms to distort prices, undermining the integrity of the market.
Even before the digital age, momentum trading generated significant financial instability. It was momentum trading that saw the price of rare tulip bulbs inflate to over USD 1000 each (adjusted for inflation) at the height of the bubble in Holland in the 1636, and saw their price crash to one hundredth of its peak in a matter of days. Whilst I am not suggesting that there is an impending algorithm-driven Tulip mania, the risk of contagions must be recognised and understood to minimise their impact on financial markets.
As the name suggests, a contagion requires two or more algorithms to interact, trading on the same security to push its price away from the intrinsic value. This can involve sophisticated algorithms interacting with each other, or interacting with simpler programs such as stop-loss facilities, chasing one another to an inaccurate valuation. This is not a theoretical risk, the interaction of algorithms has caused small but rapid price deviations in stocks, currencies and derivatives in Australian and foreign financial markets in the past.
Regulators attempting to limit the impact of misbehaving algorithms have responded with a number of safeguards to stop large deviations, however the complexity and speed at which they work can quickly affect prices before human intervention is possible. In August 2012, Knight Capital’s [not so] ‘smart order router’ malfunctioned and sent millions of erroneous orders to the NYSE. The result, USD 460 million in losses for Knight Capital, shares down 75 percent, and a bail out that saw it bought by another company. For Knight Capital, a programming error cost the firm its own existence.
Regardless of regulatory measures, the most important restraint remains the design and testing of the algorithm. We must remember that whilst these AI algorithms can give the illusion of independent and intelligent automation, their failure to mitigate the risk of contagions is still human error. Indeed, the more automated the process, the smarter the people managing it must be to identify and fix issues or stop a program when it does something unexpected. But given the speed these programs run, human intervention can be too late, so a fall back ‘auto kill’ ability is needed.
In ASIC’s most recent review of high frequency trading and dark liquidity the ASX noted that the concentration of high frequency traders was increasing in Australia, with the 20 largest traders accounting for 95 percent of turnover volume. As these firms have consolidates, their algorithms have become better at avoiding others, reducing the risk of contagions. However, as computing power improves and the complexity of programs develops, there will continue to be a risk of new and unpredictable interactions.