Causal Discovery Algorithms in Factor Investing: Applications and Insights from Optimal Transport

Abstract

Factor investing research has identified numerous return-predictive "factors," yet many of these relationships may be correlational rather than causal. This thesis investigates whether advanced causal discovery methods can distinguish genuine factor causality from spurious associations in financial markets. We develop a synthetic dataset of stock returns with pre-defined causal relationships among six common equity factors (value, size, quality, volatility, SMB, HML) to rigorously test and compare several modern causal discovery algorithms. Subsequently, we apply this framework to real Fama-French factor data from 1990 to 2023. Our study employs a comprehensive suite of causal discovery algorithms, including the constraint-based PC algorithm, functional-based Additive Noise Models (ANM), and the optimal transport-based DIVOT method. Our synthetic data results demonstrate that all three methods perform similarly, each achieving 66.7% accuracy. However, the application to real financial data reveals a key finding of this thesis: a performance reversal of the methods. In the complex real-world environment, the DIVOT method proves to be the most robust, achieving an accuracy of 66.7%, while the PC algorithm and ANM struggle significantly (16.7% and 0.0% accuracy, respectively). This disagreement between methods and environments highlights the nuanced, time-varying nature of financial data. The study concludes that while algorithmic causal discovery shows promise, a multi-method approach is essential for validation, and that distributional methods like DIVOT may be better suited for the complexities of real financial markets.

Date of Award	2025
Original language	American English
Supervisor	Yerkin Kitapbayev (Supervisor)