Data-Driven Dynamic Portfolio Selection with Vine Copula Simulation and Reinforcement Learning

01 ABSTRACT

The authors propose a data-driven dynamic portfolio selection method combining vine copula simulation and reinforcement learning, maximizing expected terminal utility under CRRA with a running CVaR penalty on one-period losses. They estimate a conditional vine model from historical returns to generate joint scenarios for CVaR evaluation and train a recurrent policy on conditional paths. Under local derivative bounds they derive power-law stability bounds in Hellinger distance and total variation, and in a finite-horizon model combine these with approximation budgets to bound policy values, optimal values, and true-market loss.

02 KEY FINDINGS

1. The method integrates vine copula conditional simulation with reinforcement learning to train a recurrent portfolio policy under a CVaR penalty.
2. Theoretically, edgewise pair-copula parameter errors are propagated to obtain power-law stability bounds in Hellinger distance and total variation.
3. On a seven-asset panel, an ensemble of ten policies achieves 34.13% CAGR and 2.31 Sharpe over 22 complete monthly holding periods, with a 41.5% reduction in target-to-target turnover.