AI Stockpicking Claims Challenged by New Research Study Exposes Limits of Artificial Intelligence in Fund Management

Claims that artificial intelligence will revolutionise investment returns are facing scrutiny as a new academic study warns that hoped-for gains may be exaggerated. According to findings from Scientific Beta, a research firm with an extensive clientele including BlackRock, Legal and General, and UBS, back-tested AI-driven fund management models often overstate their effectiveness in real-world conditions.

The report highlights systemic biases within AI models, including reliance on data points unavailable at the original investment date, a tendency known as hindsight bias. Moreover, many models appear to achieve impressive outperformance by targeting microcap stocks that are simply too illiquid to trade efficiently. This means the returns generated during historical simulations could not be realised in practice, since buying or selling such shares would likely result in adverse price movement.

Felix Goltz, head of research and co-author of the Scientific Beta paper, notes that some proponents have suggested AI models could beat conventional benchmarks by up to 40 percent annually. After factoring in trading constraints and removing data that would not actually have been known at the time, the study found the realistic outperformance is closer to three percent per year.

Blackbox investing, already popular among hedge funds and quantitative managers, is now joined by machine learning approaches that enable algorithms to identify patterns and make decisions autonomously. However, the study shows that the leap from back-tested results to live performance shrinks sharply when more practical trading realities are applied.

The research, based on US stock price data spanning 1993 to 2021, cautions that investors should not assume sophisticated models will automatically deliver robust results. Goltz’s team found that increased computing power added little to performance compared with simply granting the algorithm access to more data. The report concludes machine learning methods are not universally beneficial and that success depends on the specifics of their implementation.

Sector-wide, the adoption of artificial intelligence continues to rise, with two-thirds of asset managers deploying AI to boost operational efficiency. More are exploring machine learning to refine trading and investment strategies, as indicated by a separate Neudata study published in September.

As the debate continues, the message for investors is clear: critical assessment and a focus on realistic trading conditions remain vital when considering AI-based investment solutions.