Dynamic shrinkage estimation for portfolio estimation : a supervised learning and reinforcement learning approach

Details

Supervisors

Lassance, Nathan

Faculty

Louvain School of Management

Degree label

Master [120] : ingénieur de gestion, à finalité spécialisée

Abstract

enThis master thesis investigates whether supervised learning and reinforcement learning can improve covariance matrix estimation for Global Minimum Variance portfolios by learning dynamic shrinkage intensities. Instead of relying on a fixed shrinkage parameter, the thesis models shrinkage as a time-varying decision that adapts to market conditions. Supervised models are trained to predict an ex-post oracle shrinkage intensity from volatility, correlation, eigenvalue and covariance-stability features, while reinforcement learning agents learn shrinkage policies directly from realised portfolio outcomes. The results show that shrinkage intensity varies significantly over time and is partly predictable, with nonlinear models and reinforcement learning approaches providing the strongest improvements over the sample covariance benchmark. However, robustness checks indicate that performance depends on the estimation window, dataset and modelling framework. Overall, the thesis shows that data-driven dynamic shrinkage can improve portfolio stability and connects classical covariance shrinkage with adaptive machine learning-based portfolio construction.