What can UK pension schemes learn from GPIF’s adoption of AI? Mona Dohle takes a look.
When Hiromichi Mizuno joined GPIF as chief investment officer in 2014, he was taking on a big challenge. The Japanese Government Pension Investment Fund, which manages ¥159trn (£1.1trn) was in the midst of a significant switch into riskier assets. Between 2012 and 2015, GPIF reduced its allocation to domestic bonds to 38% from 67%. By March 2019, Japanese bonds only accounted for 26% of the scheme’s assets while exposure to foreign debt and global equities increased significantly.
Over the past 10 years, in the context of Abenomics, the scheme has faced negative returns across its equity and bond portfolios whilst forking out more than ¥70bn (£500m) in management fees between 2013 and 2016 alone. In the final quarter of 2018, the pension fund giant reported losses of ¥15trn (£106bn). GPIF argues that the culprit for such a disappointing performance was not just low yields but also the failure of active managers to outperform. GPIF’s leadership then made a bold statement: artificial intelligence (AI) could play a key role in addressing this problem.
In 2017, GPIF commissioned Sony Computer Science Laboratories to investigate how the pension fund could use AI to transform its performance. By assessing fund manager styles, it hopes to improve its understanding of alpha generation and investment performance. As of January 2020, GPIF confirmed that it had set up a system to monitor funds across a universe of 1,000 Japanese and foreign stocks. By analysing trading data, the system aims to identify resemblances in the characteristics of various fund managers. It is now at the stage of implementing the system across its portfolio, initially on an experimental basis.
While the Japanese investment giant might be at the forefront of AI implementation, there are early indications that its European and American counterparts might follow suit. More than half of institutional investors globally plan to use AI as part of their investment research within the next five to 10 years, according to a recent poll by Thomson Reuters and Greenwich Associates. This would mark a significant change as only 17% of respondents are currently using AI in their investment process.
Machine learning
Yet the extent of AI adaptation could easily be misinterpreted by conflating it with a more efficient use of financial data. For John Beckett, author of the book “The New Fund Order”, three key criteria define AI.
“Of course, fund managers and investors are used to accessing data since the early days of Bloomberg and Archipelago,” he adds. “But data as an information point doesn’t constitute AI. For AI to be present either of three criteria must apply.
“First, it should involve an algorithm or programme built in layers, what we call pillars which make decisions, so there must be a clear learning aspect to it.
“The second form of AI is where you develop an algorithm which can move left or right based on the information you enter. And the third is what we would call the naïve Bayes classification based on Bayesian inference. If you confront the system with a set of factors, it can predict the probability of a certain outcome. If you don’t have one of these types of machine learning all you have is data,” Beckett stresses.
His impression is that most professional fund investors are not yet engaging with AI on a conscious basis. “Are fund selectors taking the lead on this? The answer is no. I know only of a few, mainly quant-driven, professional investors with a mathematical background who are looking for tools in the market that empower them to incorporate AI in their decision-making process.
“Overall, AI integration happens more as an osmosis. A lot of fund analysts are not fully aware of the change in the way they are doing things,” he adds. “In the same way that 20 years ago, it would have been hard to explain to fund selectors what Morningstar or Lipper would do for them today.”
Morningstar on the other hand is already using AI in its fund ratings, which in turn also affects institutional investors who use the data provider. In 2018, it launched a new quant rating system which uses a machine-learning model that enables it to rate six times as many funds than an analyst would.
Gavin Corr, director manager selection services at Morningstar, explains that the tool helps Morningstar to tackle the effects of a rapidly growing asset management industry. Morningstar’s 125 fund analysts currently screen funds based on their investment process, people employed, past performance, the parent company and price, subsequently awarding funds with a bronze, silver or gold rating. The machine learning model replicates this screening process across a broader fund range and can independently award funds with the usual bronze, silver or gold rating.
Corr is keen to stress that the introduction of machine learning has not led to it employing fewer fund analysts but instead is aimed at helping the firm grapple with the dramatic growth of the fund universe and the persistent presence of orphaned funds with smaller assets and high fees who might have previously not been analysed.
AI and ESG
Another field where AI has already been actively implemented is in ESG investing. Dutch pension giant APG, which manages €399bn (£339bn), was among the first pension schemes to apply AI in this field.
Since 2016, it uses its tech subsidiary Entis to scan 10,000 listed companies for their compliance with the United Nations’ Sustainable Development Goals (SDG). It has since been joined by €234bn (£199bn) pension fund PGGM. Both funds plan to launch a sustainable development investment asset owner platform by the end of March.
Humans are ultimately very expensive fleshy robots.
John Beckett
In the UK, Brunel Pension Partnership, a local government pool for £30bn worth of investments from 10 partner funds, appointed AI firm TrueValue Labs in October to screen its listed investments for ESG and reputational risks.
For Faith Ward, Brunel’s chief responsible investment officer, using AI data provided by TrueValue adds an element of objectivity. “TruValue Labs’ data isn’t dependent upon what companies publish about themselves. Their timely material ESG data helps us to continually monitor the managers in our client partners funds and to evaluate and select new managers,” she says.
Peak human?
The factors which drove GPIF to consider using AI in its fund research might not be entirely unfamiliar to many UK funds. As a relatively young fund, launched in 2006, GPIF faced the effects of the first wave of quantitative easing in Japan. The yield compression was painful its portfolio, which was then heavily reliant on fixed income and forced it to shift its money into riskier assets.
This is indicative of a global trend. Between 2004 and 2014, global assets under management doubled from $37.3trn (£28.6trn) to $78trn (£60trn) and could reach as high as $112trn (£860trn) by the end of 2020, PwC predicts. This growth is linked to the impact of quantitative easing worldwide, with $4.5trn (£3.4trn) in assets bought by the US Federal Reserve. Simultaneously, the growth of passive investments throughout that period shone a light not just on the inefficiencies of active managers, but also on those working to select investment funds who failed to outperform.
Becket argues that throughout the same period, the number of professional fund investors has increased from more than 50 in the early 2000s to more than 5,000 last year. For Beckett, this means that we may have reached “peak human” and that investors should brace themselves for consolidation.
Trends at GPIF might offer a taste of things to come. Despite being the world’s largest pension fund with assets worth more than £1trn, the fund only employs 139 members of staff, a similar amount to the UK’s largest retirement scheme USS, which has around £63bn in assets and employs 140.
For John Southall, head of solutions research at LGIM, the growing importance of AI in fund analysis is more likely to require a different skill set, rather than an outright reduction of investors.
He predicts that the ability to run suitable algorithms, to extract relevant information and sorting the signal from the noise will be of increasing importance. “An algorithm might flag a connection but it will often need human oversight to establish the risk of data-mining, the validity of an output on a forward-looking basis due to events, how the strategy would fit into the broader portfolio and suitable risk limits based on the plausibility of the strategy.”
The impact of AI is more likely to be a shift in skill-sets rather than a reduction of staff.
John Southall, LGIM
Christine Chow, director and head of Asia and global emerging markets stewardship at Hermes, and her colleague Janet Wong, also stressed in a recent paper for Hermes Investment Management that there will still be need for a human element. “For AI to be useful to investors, analysts need to establish hypotheses on the possible relationships between data and the results obtained when they are used in investment analysis.” But fund analysts should still brace themselves for increased scrutiny.
The way forward
How could AI affect fund research over the next 10 years? For Morningstar’s Corr it is clear that while AI might not lead to a reduction of staff members at Morningstar, pension funds and other institutional investors will certainly consider re-sizing their research teams, certainly in the light of a growing focus on investment research costs under Mifid II. He predicts that fund research could evolve in a similar way as credit rating, with a handful of institutions dominating the ratings market. While Beckett is perhaps the most outspoken on the impact of AI on fund research, stating that “humans are ultimately very expensive fleshy robots,” he also predicts that AI could have a positive role to play in the fund landscape of the future. He sees opportunities for AI to influence the passive fund industry, for example, by shaping the way indices are being designed and creating more objective criteria for index creation.
In the best-case scenario, integrating AI should lead to better outcomes for scheme members. “The time saved could be devoted to a much more enriched level of engagement with customers and communication will be a much bigger role of their job than it is now,” he predicts.
