You are here
The robots are coming for fund management jobs
REMEMBER Aibo, the computerised dog Sony Corp started selling in 1999 as the first personal robot? Hiro Mizuno, the chief investment officer of Japan's Government Pension Investment Fund, does. So he asked Sony's computer science lab unit to build him a cyberhound using artificial intelligence to help oversee the external fund managers who manage GPIF's US$1.6 trillion in assets.
If the training programme succeeds, the software watchdog could catch investors who are straying from their comfort zones, help screen potential portfolio managers based on their previous track records, and even distinguish between luck and skill in generating returns. But there's a catch - and it's a big one that's not unique to this particular use of AI.
The project, which Mr Mizuno says is part of his experiments in improving the way money is managed, will run through March, but the Sony team recently issued an interim report.
The system uses what Sony calls deep learning systems. It was initially trained on 1,000 stocks, using seven computer-generated investment styles including strategies that favour cash flows, dividend yields, and both positive and negative momentum to generate long and short positions.
It was subsequently given actual portfolio data from GPIF's existing external managers to analyse. Here's where it gets interesting.
One way the system tracks portfolios is by analysing their self-resemblance over time, on the theory that a manager pursuing a particular strategy should stay true to whatever investment parameters the fund has deemed most likely to generate returns even as market conditions change.
In one of its case studies, the Sony software identified a plunge in a fund manager's self-resemblance measure that persisted for some time.
Subsequent analysis found the shift in buying and selling patterns away from the norms previously observed reflected efforts to chase performance - in contravention of the fund's stated goals.
In another example, a fund stopped resembling its previous behaviour and started instead to match one, then another, of the other funds being studied.
The researchers conclude that the fund must have a policy of "opportunistically switching between different investment strategies based on market conditions". While that may be a reasonable plan of market attack, it's only legitimate if the fund is upfront with customers about its intention to zag and zig in its approach. Otherwise, such randomness should raise a red flag.
Catching instances of style drift depends on comparing current portfolio selection with historical trading over lengthy time periods, involving the sort of data crunching that computers are much better at than human overseers.
Sony also says the system can compensate for the very human tendency to focus on what's going wrong rather than things that seem to be working fine. "The inevitable reality is that relatively worse-performing funds receive the most attention," the report says.
A fund delivering stellar returns but only by trading outside of its pre-agreed area of expertise may be storing up trouble for the future, a deviation that may go unnoticed by human overseers until profits become losses, but which the AI program can detect as style drift.
Japan's GPIF has its own particularities including a reputation for being somewhat bureaucratic. The Sony researchers call its demands on its chosen fund managers as "onerous".
They suggest that automating the analysis by allowing the system to scrutinise existing investment behaviour could pave the way for smaller firms to become available as potential candidates for inclusion in the fund's roster of external managers.
So what's not to love about a virtual guard dog written in code that can scrutinise how billions of dollars of assets are being invested better than a human can? The issue is that when it woofs, its handlers can't determine the steps that triggered the bark - a problem known in AI research as the black box problem.
As Sony says, potential users of its system have to "accept the operation of the neural network itself as a black box". In other words, while the inputs and outputs of the system may be clear, what happens in between remains opaque. That's not desirable for institutions such as GPIF, which are fiduciaries of the public interest and have an obligation to be transparent.
Because of its size as the world's biggest pension fund, GPIF is known domestically as "The Whale". Sony's hope, expressed in the research paper, is that its computer-driven augmentation will mean "cybernetic whale will be able to see what humans may not be able to see without the assistance of AI".
Mr Mizuno of the GPIF went a step further at a seminar earlier this month at Oxford's Said Business School, saying he'd like the programming to develop and mimic the style of the active investment managers it analyses, basically computerising the investment process. We're not there yet, but the robots are coming for jobs in all kinds of areas of employment. BLOOMBERG