Scientific Beta

In this research overview, we show that machine learning methods can generate substantial outperformance in realworld equity investments, but only when strategies are built from the ground up with implementation discipline and information breadth as core design principles.

In this research overview, we show that machine learning methods can generate substantial outperformance in realworld equity investments, but only when strategies are built from the ground up with implementation discipline and information breadth as core design principles.

Machine learning models are different from traditional factor models because they can add more information and are more flexible when capturing relationships between firm characteristics and stock returns. We systematically compare linear versus nonlinear models using both traditional factor sets and expanded information sets. Importantly, we evaluate performance not just in the stylised conditions used in popular studies, but also in realistic investment conditions that account for transaction costs, exclude microcaps, and eliminate factor hindsight bias.