In October last year, fintech company NMRQL launched its first unit trust in South Africa. The NMRQL SCI Balanced Fund is the first local unit trust to exclusively use machine learning for managing its portfolio.
Using computers in investment management is certainly not a new idea, as there are already a number of funds run using quantitative models and many active managers use these systems in their decision making. This is however the first fund giving South African investors the chance to see if machines on their own can make better investment decisions than humans.
Proponents of artificial intelligence argue that this won’t be much of a contest, given what has already been achieved. Computers have been built to play better chess than the best human masters, have beaten the top-ranked players at the far more complex Chinese strategy game Go, and last year overcame the world’s best poker players and achieved the highest possible scores in dozens of Atari games.
It is inevitable, they argue, that computers will also become better than humans at managing money.
So how does it work?
Machine learning differs from quant investing in that quant models start off with prior assumptions. In other words, they are programmed to pick stocks based on already-identified criteria.
With machine learning, the computers themselves decide which criteria should be used, and when. They do this based on interpreting vast amounts of data and using algorithms to make decisions based on patterns they have picked up.
The arguments in favour of machine learning essentially focus on two key points. The first is that computers are able to analyse far more data and do it much quicker than humans can. The second is that they can act on what they pick up without the emotional biases any human has to deal with.
“At the end of the day, it’s maths, statistics and computer science,” says Tom Schlebusch, NMRQL’s CEO. “There’s no magic involved.”
The computers are coming
A number of people have however expressed concern about letting computers loose on the investment world. The ‘Flash Crash’ in May 2010 was caused by an algorithm that essentially spiralled out of control as it began acting based on its own trades.
The Long-Term Capital Management hedge fund also famously came close to bankruptcy back in 1998 and had to be bailed out by the US Federal Reserve. That fund was using quant models to take highly-leveraged positions.
Schlebusch however insists that the NMRQL SIC Balanced Fund is very different to either of these.
Firstly, it is not a hedge fund and employs no leverage. It also is not a quant fund that looks to benefit only be making specific kinds of trades.
“In addition, we are not high-frequency traders,” says Schlebusch. “We don’t trade every day. We only trade on a weekly basis and turn the portfolio about three times a year.”
It is therefore not likely that it will influence its own behaviour.
Better, but not infallible
NMRQL’s chief engineer, Stuart Reid, adds that it’s important to appreciate that artificial intelligence doesn’t pretend to be perfect. In this respect, self-driving cars make for an interesting comparison with what NMRQL is doing.
“What people don’t realise is that self-driving cars are expected to be three to four times better than a human driver, but that doesn’t mean that they won’t have crashes,” Reid says. “They will almost certainly have crashes, but they will still be much safer.”
A self-driving car can’t anticipate every scenario. If another car swerves out directly in front of one of them, for instance, it can’t avoid an accident. However, it does take human frailty completely out of the equation, such as having a driver with poor eyesight, slow reaction speed or who is under the influence of alcohol.
The South African market has already produced an example of something similar in terms of investing. NMRQL’s algorithms could not pick up ahead of time that a Viceroy report on Capitec would hurt the company’s share price, but it did pick up on problems at Steinhoff and therefore the fund did not hold the share in December last year.
“It’s the difference between interpolation and extrapolation,” Reid explains. “Interpolation is predicting based on things that are similar to what you have seen before. Extrapolation is what happens when you try to predict something that you’ve never seen before. Generally machine-learning models are only good at extrapolation to a certain point, and then they start to fail.”
NMRQL therefore continually monitors its algorithms to see how well they are predicting the current market environment. When they are not predicting well and their errors are growing larger, this is an indication that they are relying too much on extrapolation. At these points, the system is designed to pull out of the market.
“For example, 2008 was a completely new situation,” Reid says. “The behaviour we saw in equity markets had never been seen before, so the models had no reference point for understanding that behaviour and deteriorated quite rapidly. When the algorithms start to deteriorate and make bad decisions like that because we’ve transitioned from one regime to another and we don’t have as much predictive power, we pull out and put money into risk-free assets.”