Reproduce “Empirical Asset Pricing via Ensemble Gaussian Process Regression” on OpenResearch

About this paper

Damir Filipović, Puneet Pasricha

Researchers developed an ensemble Gaussian Process Regression method to predict conditional expected stock returns, providing both point estimates and quantifiable epistemic uncertainty. The approach significantly improved portfolio performance and statistical predictive accuracy by incorporating these uncertainty estimates, while also overcoming the computational limitations of traditional GPR for large financial datasets.

Key insights

• The Bayesian non-parametric nature of GPR allows for the inherent quantification of prediction uncertainty, providing valuable confidence intervals essential for risk-aware financial applications.
• An ensemble approach effectively scales GPR for large datasets, making it computationally feasible for empirical asset pricing and adaptable for online learning in dynamic markets.
• Incorporating epistemic uncertainty directly into portfolio construction significantly enhances both the statistical predictive accuracy and the economic performance of investment strategies.

Summary by alphaXiv.