Multi-Asset Dynamic Portfolio Choice under Transaction Costs: A Machine Learning Approach

Département: Finance

Intervenant:  Fabio Trojani  (University of Geneva)

Salle:T004

Title:

Multi-Asset Dynamic Portfolio Choice under Transaction Costs: A Machine Learning Approach

Abstract:

We develop a general computational framework for solving high-dimensional dynamic portfolio choice problems with proportional transaction costs. The methodology exploits the endogenous geometry of the optimal no-trade region to decompose the Bellman equation into economically distinct local approximation problems. Gaussian process regression, combined with Bayesian active learning, is then used to construct mesh-free surrogate models for the resulting local value-function components, substantially reducing the computational burden of dynamic programming. The framework accommodates multivariate state-dependent investment opportunity sets, including regime-switching expected returns and covariance structures, stochastic proportional transaction costs, and portfolio constraints. Numerical experiments demonstrate that the methodology accurately recovers value functions, optimal trading policies, and no-trade regions while scaling to substantially richer stochastic environments and larger asset universes than existing solution methods.