Robo-traders and robo-labour

Banks and buy-siders are starting to harvest the benefits of machine learning beyond the front office

Mark Long, NBillustration.co.uk

The application of machine learning to generate trading strategies might grab the headlines, but its use elsewhere within financial firms also deserves attention.

Machines are being put to work in areas such as model validation and credit risk, in addition to activities such as trading execution. Teaching a computer how to make money remains a work in progress, but teaching it, say, to spot errors in overnight risk calculations could be both easier and less hazardous.  

In trading, Bloomberg is one firm employing tools to analyse fixed-income liquidity that draw on machine-learning techniques. The vendor uses cluster analysis – a machine-learning method – to estimate transaction costs for illiquid bonds where relevant trading data is sparse. Machine learning helps determine the similarity of apparently unrelated bonds.

Natixis, meanwhile, is using machine learning to alert it to anomalies in its stress testing results. Regulatory pressure has loaded model validators with greater workloads, and machine-learning techniques help ease that burden, freeing staff up to focus on more pressing issues.

Machine learning is really just sophisticated statistical analysis for large libraries of data. It uses an algorithm like any other program, but rather than being told explicitly what to do, it is told how to construct and revise hypotheses using statistical methods.

Greater volumes of data, increasingly powerful hardware and the development of sophisticated algorithms have all contributed to its rise.

But one of machine learning’s greatest strengths is also one of its greatest weaknesses. By allowing a computer to learn for itself, without an explicit authoritative code, it becomes harder to explain the computer’s conclusions.

This is the notorious ‘black box’ problem: “Data goes in and data comes out, but what happens in between?” critics ask. Detractors worry that by dint of relying so heavily on historical data, machine learning’s statistical results are invariably overfitted. They fret that such methods will be applied to problems where data is too scarce or where relationships between data are changing.

Quants are right to be wary – but wariness should be applied to any tool. Machine learning can point to demonstrable successes when aptly used. Firms are steadily working out that its application is well-suited to areas where resource-stretched teams are drowning in data – problems that are equally likely to be found in the back office as in the front.