Risk Budgeting Allocation for Dynamic Risk Measures.

This article focuses on developing time-consistent dynamic risk budgeting portfolio strategies using coherent distortion dynamic risk measures (DRMs). It introduces a notion of dynamic risk contributions based on Gâteaux derivatives, proving that these contributions satisfy a full allocation property and can be explicitly computed for coherent distortion DRMs. The authors characterize self-financing dynamic risk budgeting strategies as unique solutions to a sequence of strictly convex recursive optimization problems and propose an actor-critic deep reinforcement learning algorithm to approximate these strategies efficiently. Numerical illustrations demonstrate the approach in a stochastic volatility market model with multiple assets and time steps, highlighting convergence and interpretability of the learned strategies. The work leverages the elicitability of conditional risk measures to enable practical estimation without nested simulations.