A review of machine learning experiments in equity investment decision-making: why most published research findings do not live up to their promise in real life

Abstract

The numerical nature of financial markets makes market forecasting and portfolio construction a good use case for machine learning (ML), a branch of artificial intelligence (AI). Over the past two decades, a number of academics worldwide (mostly from the field of computer science) produced a sizeable body of experimental research. Many publications claim highly accurate forecasts or highly profitable investment strategies. At the same time, the picture of real-world AI-driven investments is ambiguous and conspicuously lacking in high-profile success cases (while it is not lacking in high-profile failures). We conducted a literature review of 27 academic experiments spanning over two decades and contrasted them with real-life examples of machine learning-driven funds to try to explain this apparent contradiction. The specific contributions our article will make are as follows: (1) A comprehensive, thematic review (quantitative and qualitative) of multiple academic experiments from the investment management perspective. (2) A critical evaluation of running multiple versions of the same models in parallel and disclosing the best-performing ones only (“cherry-picking”). (3) Recommendations on how to approach future experiments so that their outcomes are unambiguously measurable and useful for the investment industry. (4) An in-depth comparison of real-life cases of ML-driven funds versus academic experiments. We will discuss whether present-day ML algorithms could make feasible and profitable investments in the equity markets.