Case study : How is Sanlam’s AI-managed fund holding up?

Running a portfolio using an artificially intelligent system has encouraged one asset manager to expand into wider mandates.

In 2017, South African head quartered asset manager Sanlam enhanced the Sanlam Managed Risk (SMR) UCITS Fund with risk management signals from an artificially intelligent system called the Predictive Investment Engine, designed by software firm A.I. Machines. That year the fund delivered a 16.8% return, against the 12.1% return of its benchmark, the Morningstar USD Flexible Allocation category of funds.

The SMR is an Irish Stock Exchange-listed fund that dynamically adjusts its net equity exposure to developed markets on a weekly basis between 90% and 10%. It achieves this by passively tracking the MSCI World Equity Index in US dollars combined with an active risk overlay. The risk overlay is supplied by the AI system that autonomously adapts the fund’s weekly net equity exposure through short equity futures, rather than by trading cash equities in order to minimise transaction costs.

Although the fund had always been systematically managed it had historically functioned best in trending markets.


“It liked bull markets and bear markets, but it didn’t like V-shaped markets,” says Gideon Nell, global head of business development, at Sanlam Global Investment Solutions. “The market became very V-shaped in 2014, and it persisted until Brexit in 2016. The only way to manage that aversion to V-shaped markets had historically been to employ lots of PhD-qualified investment specialists, to follow the movements of data series, note the changes in patterns then rewrite the code to manage that change.”

However, not being a big quant shop, that was not a very efficient process for Sanlam. Checking data, reviewing findings then making a decision based on those findings was relatively manual and time consuming.

“We engaged with A.I. Machines, and learned the machines can rewrite their own algorithms,” says Nell. “It took nearly two years to work through the required approvals internally and externally.”

Today the assets under management (AUM) of the fund is US$89 million, and its performance since the AI was included stands at 10.4% against 3.7% of its benchmark.

A working model

The AI fund manager has a point each week at which it makes a decision, based on its daily observations of the market data. Nell notes that having 52 asset allocation decisions a year is significantly more than a human portfolio manager would typically make. It also allows the fund to make decisions that a human might not make, due to the apparent disparity in changing direction so often.

The firm says it enjoyed a wonderful participation in the 2017 rally then, two weeks before the markets became unstuck at the end of January 2018, the fund went from being heavily risk-on to risk-off.

“No trend following or momentum driven strategy would have done that, and they all looked terrible at that time because they did not do that,” Nell observes. “That was unbelievable for us. Although its ‘walk-forward’ tests had demonstrated what it was capable of in testing, because it was not live we hadn’t appreciated it. Our first example was in January 2018 when it went from 82% equity exposure to 19% equity. A week later the markets started falling. Then, the week after that it went back up to 80+%.”

He notes that aggressive asset allocation from a human might involve moving from 70% to 50%, far tamer that the 82% to 19% shift.

“When it reversed that the following week it made a call that no human asset allocator would have the confidence to make.”

Limits and risks

The more frequently the fund can adjust to data, the more accurately it can respond to changes it sees in the data. Sanlam has decided it wants the fund to work as fast as possible. However, at the same time it wants to sell the mandate into the pension space and that carries cost limits.

“In the UK the maximum fee is typically 99 basis points total cost to client or you do not get considered for a mandate,” Nell says. “We have a fund that has an annual management charge of 75bps and of course if we trade too often it won’t take much to push the cost too far.”

Using a short future overlay helped to manage that concern and the fund’s current TER is 0.84%.

“Instrument selection was one of the most important things we found when considering the use of AI for a fund,” Nell says, adding that the use of a low cost MSCI World Equity Index tracker and what are very cost efficient and liquid futures reduces cost and risks as the fund grows.

“In the systematic space people often have to go into weird and wonderful instruments to carve out their value,” he notes. “However, with a very simple mandate and the right instruments, an AI has few constraints.”

The engine

Alessandro Di Soccio, co-CEO at A.I. Machines, says the firm is currently working with both conventional asset managers and wealth managers and has been in talks for several months with two hedge funds, regarding long-short equity strategies. At present he reports there are 14 managed accounts being run on PIE and two funds. He is reluctant to discuss the technology underpinning the system due to its commercial sensitivity.


“There is a lot of proprietary know-how related to applied artificial intelligence,” he says. “Large quantitative shops and large asset managers are struggling to get their hands onto something that really works. The development of an industrial grade AI engine requires a multi-year, multi-million dollar R&D programme, with no guarantee of success at the end of it. Once the engine has been built, its effectiveness needs to be demonstrated over multiple years of live trading.”

However, he says that the system uses multiple disciplines in order to function effectively, and flexibly.

It’s a matter of doing it right in order to create a scaleable process,” he says. “And a scaleable process is never linked to a single algorithm. That is what PIE is all about. If a developer falls in love with one particular learning algorithm you will have a strong bias to that type of algorithm. So, in order to diversify and reduce your single point of failure of risk in terms of technology, you should create an architecture that is based on multiple types of machine learning methods.”

Sanlam has been pleased with the results and the firm believes the platform is flexible enough to handle other mandates.

“The AI Machine engine was given a proof of concept with [the SMR] mandate,” says Nell. “We have since given them more mandates. We have more plans with them. And Sanlam has traditionally been a very conservative firm.”



Leave a Reply

Your email address will not be published. Required fields are marked *