A2029
Title: Recovering no-trade regions: Pontryagin-guided policy projection for transaction-cost control
Authors: Jeonggyu Huh - Sungkyunkwan University (Korea, South) [presenting]
Hojin Ko - Sungkyunkwan University (Korea, South)
Abstract: Transaction-cost portfolio choice is a canonical high-dimensional nonsmooth control problem: proportional costs induce no-trade regions and boundary interventions that neural policies can fail to preserve. Regularized PG-DPO, a Pontryagin-guided policy-projection method, learns this structure without solving a global value function or variational inequality. The method warm-starts a policy by direct Monte Carlo optimization, estimates fixed-policy marginal values by BPTT through continuation rollouts, and applies a closed-form dead-zone projection derived from a quadratically Regularized transaction-cost Hamiltonian. The projection exactly preserves the no-action wedge and converts marginal-value ratios into buy, hold, or sell decisions. Local guarantees are provided for fixed-policy marginal-value recovery, projection stability, and margin-based buy/hold/sell regime recovery. Experiments on multi-asset portfolio benchmarks show recovery of no-trade regions, correlation-dependent intervention geometry, mixed trading regimes, and dimension-robust local optimality diagnostics up to 50 assets.
