Essays in Empirical Asset Pricing with Machine Learning

Student thesis: Doctoral ThesisDoctor of Philosophy


This thesis consists of two papers on topics in empirical asset pricing with machine learning. The first paper introduces a nonlinear and non-parametric generalisation of the seminal Fama-MacBeth cross-sectional regressions through partial derivatives of an arbitrary estimator function with respect to its input. We show that our novel methodology is universally applicable to a large number of models and can be applied to estimate risk exposures and prices of risk. Furthermore, we introduce the concept of Jacobian regularisation to the field of asset pricing and show that the penalisation of the input gradients is economically meaningful. Our estimators outperform all considered linear benchmarks. The paper further introduces the notion of sensitivity-sorted portfolios, a novel technique that leverages the input gradients for portfolio construction that deliver economic gains for investors. Most importantly, we move elements of explainability and interpretability to the foreground of discussion, contributing to the emerging literature that connects interpretable machine learning and asset pricing. The second paper studies the marginal conditions under which out-of-the-box machine learning starts and stops adding value to institutional investors – a critical question for practitioners. We refer to those conditions as performance borders. We contribute to the literature investigating equity returnsimulations in the cross-section of returns by introducing a novel methodology for simulatingfirm characteristics and returns. In addition, we discuss the notion of goodness-of-simulation,which quantifies the quality of a simulated investment universe. An empirical study based ona standard Fama-French three-factor data-generating model finds that out-of-the-box machinelearning fails to perform competitively out-of-sample. The recovered performances point todomain-specific model adjustments necessary for machine learning to thrive in asset pricing.

Date of Award1 Aug 2022
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorGeorge Kapetanios (Supervisor) & Fotios Papailias (Supervisor)

Cite this