In 2018, it feels like the use of artificial intelligence (AI) is poised to accelerate unhindered. CB Insights reports that funding deals to AI start-ups increased from 150 in 2012 ($559m invested) to 698 in 2016 ($4.8bn invested). In June 2017, McKinsey ranked financial services first in ‘future AI demand trajectory’ – measured by the estimated percentage change in AI spending over the next three years.
So are recent applications of AI truly revolutionary, or are they just a way of improving on an already existing process?
Hedge funds are embracing AI, but wide adoption is unlikely
In the world of hedge funds, AI is nothing new. According to David Andre, who has a PhD in AI and is CEO of Cerebellum Capital in San Francisco: “[In the past] it has mostly been about fitting models to data … what’s really changed recently are new machine learning (ML) algorithms, such as deep learning, and the improvements in computer power.”
Deep learning involves trying to model the brain, not the subject being analysed. The most quoted example of this is giving a system millions of examples of cats so that the system, on its own, can learn to identify a cat. In an investment context, a deep learning system wouldn’t be given pictures of cats but carefully defined 'pictures' of ‘attractive investments’ that it would then try and find.
New York-based Rebellion Research, a hedge fund that launched in 2007, has been applying these new techniques. In the early days of the fund, when analysing the Australian economy, its systems automatically adjusted for the rapidly increasing importance of, and correlation with, Chinese industrial output. CEO Alexander Fleiss says: “[Today,] ML not only learns how the economy adjusts for things like commodity prices going up and down; its factors for rating individual assets change over time. So not only will its predictions be different [due to changes in inputs like prices], but the factors that create the prediction will be very different.”
The machines appear to have delivered. Rebellion reported a cumulative return of 157% between January 2007 and August 2017, compared to the 111% gain of the S&P 500. Other AI hedge funds have also performed well but not spectacularly. Eurekahedge has constructed an index from the performance of 23 AI/ML hedge funds. This shows a cumulative return of 63% between 2011 and 2016, outperforming traditional ‘quant’ funds and hedge funds generally, but not the S&P 500, with a cumulative return of 97%.
Advances are about analytical capability and
the sophistication of input data. Algorithms now scan the web, read and ‘understand’ text and speech, and extract sentiment and meaning. Satellite images of crops or shopping centre car parks are analysed to forecast yields or detect early patterns in consumer behaviour.
How quickly AI is embraced as a normal part of the investment process is likely to depend on the sophistication of the investor. At Rebellion Research, the majority of retail investors work in high technology companies. “They are engineers or managers who have done ML work on their own; they believe in it,” says Alexander.
But David reckons that, for most traditional investors, trust and “understandability” are going to be the biggest roadblocks to growth in the use of AI across the investment sector.
AI complements fundamental research for asset managers
Cardiff-based technology firm Amplyfi is attacking this space. Algorithms access and analyse both the ‘surface web’ and the ‘deep web’. The surface web is smaller and accessible using standard search engines. The deep web has more information in richer datasets such as academic journals, government databases and financial records. It is less accessible and a vast, relatively untapped resource.
Amplyfi’s technology is used to find, analyse and discover weak signals and ‘unknown-unknowns’, correlations, trends and patterns that influence a market or investment. It enables new input metrics to be created for asset managers’ forecasting models that are often limited. (See box out.)
An Amplyfi example
Amplyfi conducted a study to try to improve inflation (CPI) forecasting. A correlation – strong enough to warrant inclusion in forecasting models – was found between CPI and perceptions of water and wastewater mismanagement in the mining industry. These perceptions were driving increased government regulation that, in turn, increased costs for mines, downstream industries and, ultimately, the price of goods used to calculate CPI. Perception, tracked by previously ‘unmeasurable’ metrics such as the volume and content of newspaper articles in a local area, became forecasting model inputs.
Mark Woods, Amplyfi's chief technology officer, says this should worry some consultants: “We don’t doubt that you could pay a large consultancy to have 200 people locked away in a room for a few weeks and produce something like this. But then in a month’s time you need to repeat the exercise again. And again.”
Should researchers and analysts be worried? Probably not. The acid test for many AI techniques used by hedge funds and asset managers will be the ability to demonstrate performance over longer periods and under different market conditions. Andrew Lo, Professor of Finance at MIT Sloan School of Management, in his ‘adaptive markets’ theory, has argued that investment strategies must change over time as markets are neither efficient nor inefficient but ‘adaptive’ and go through periods where the degree of efficiency varies.
Degrees of market efficiency can be influenced by human irrationality (bubbles and crashes) and by new technologies. Rapid smartphone adoption has resulted in the same information being available to all investors concurrently, boosting market efficiency. But advanced analysis of information from satellite images may only be affordable by a small group of institutional investors, resulting in information asymmetry and less efficient markets. Machines will need to ‘learn’ to cope with a range of such dynamic conditions.
AI threatens significant disruption for financial advisers
New York-based Pefin claims to be an AI financial adviser at “1/20th the cost” of a human adviser. The platform went live in October 2017 after running in beta with 4,000 users since early 2017. Pefin automates client-adviser interaction, asset allocation and investment decisions.
Algorithms track individual clients’ income and spending patterns. They detect changes in these patterns and ‘learn’ about client behaviour. They also track factors like tax changes and the prices of goods affecting a client.
This is used to generate a financial plan that is continuously and automatically updated. The system warns clients if they are overspending or if market conditions warrant a change to their behaviours.
Should financial advisers be worried? Probably
Should financial advisers be worried? Probably. Pefin has its own direct-to-consumer offering but also has fast-track growth opportunities such as white labelling – where an incumbent advisory firm would pay to use Pefin’s technology but rebrand the offering as its own.
But Jeff McMillan, chief analytics and data officer of Morgan Stanley Wealth Management, which has 16,000 financial advisers, sees a ‘humans versus machines’ argument as overly simplistic: “The reality is it’s the blending of the two that’s going to drive the model for the next ten years.” He stresses how good machines are at analysing large amounts of data and how ML can provide valuable insights, such as identifying when markets or clients are exhibiting unusual behaviour that doesn’t match past patterns.
Jeff equates a strong AI capability to having advisers being supported by “800 people with expertise in every financial product in the entire world”, leaving advisers more time to think about complex client issues. “When you look at what clients want from a financial adviser, asset allocation is expected [but] they also want to know how to deal with an autistic child that has just been born, or deal with a parent that has early stage dementia, or how to navigate a divorce. AI is not there yet.”
Stories such as those of Rebellion Research, Amplyfi and Pefin start to clear some of the fog surrounding the future vision of AI. A new level of sophistication is being introduced. Whether humans will be comfortable taking financial advice from machines remains to be seen. The picture is nuanced.
This article was originally published in the Q4 2017 print edition of The Review and entitled 'How artificial intelligence is changing investing'. The print edition is available to all members who opt in to receive it, except student members. All eligible members who would like to receive future editions in the post should log in to MyCISI, click on My Account/Communications and set their preference to 'Yes'.
Seen a blog, news story or discussion online that you think might interest CISI members? Email email@example.com.