Fund management companies hoping artificial intelligence will help them produce fabulously better investment returns may be deluding themselves, according to a new study.
Back-testing of their hypothetical AI models is hugely overstating the performance likely to be achieved in the real world, according to a paper from Scientific Beta, a research house whose clients include BlackRock, Legal & General and UBS.
Most of the apparent gains are illusory because they are made in small company stocks that are too small to actually trade because of the lack of liquidity, it argues. Another major bias was that of hindsight: the AI choices were based on insights from data not available at the start date of the hypothetical investment.
Felix Goltz, research head at Scientific Beta and co-author of the report, said one of the claims made for AI models was that they could produce outperformance compared with benchmarks of as much as 40 per cent a year.
But after taking account of these unrealistic biases, the outperformance was “more like 3 per cent a year”, he said.
Small companies or “microcaps” are very difficult to trade in practice without the price moving adversely, whether buying or selling, because of the lack of liquidity — a factor that was not taken into account in back-testing studies.
Fund managers and hedge funds have for years engaged in “blackbox investing”, or quant funds, in which they set the initial factors used for stock selection — such as value, company size and so on — but then leave it to the computer to make the decisions.
AI and machine learning goes much further in enabling the computer to train itself on the data it collects and make stock buying and selling decisions based on patterns that may never become apparent to humans.
• Who should you trust, an AI chatbot or a financial adviser?
Goltz found that almost all the AI benefit came from allowing the algorithm to feed on additional data. Giving it more computational power achieved very little.
The report, published on Monday, concludes: “The benefits of machine learning signals are not as ubiquitous as advertised and are substantially reduced when leaving the highly stylised settings that are standard in the literature. Investors need to evaluate these techniques under their specific implementation constraints rather than assume that more complex models will universally outperform.”
The research was based on the behaviour of US stock prices between 1993 and 2021.
Two-thirds of asset managers now use artificial intelligence to boost efficiency, with more applying AI and machine learning to optimise trading and investment strategies, according to another study in September from Neudata.