NVIDIA Portfolio Execution Gym: GPU-Accelerated Markets for Robust Trade Execution

Evaluating trade execution policies requires balancing microstructural realism with high-throughput simulation. To address this, we present NV-PEG, a GPU-accelerated framework for learning and evaluating execution policies on synthetic limit order book data. Using Tokyo equity market-by-order and top-of-book data, NV-PEG integrates parallel data preprocessing, a factorized Multidimensional Deep Queue-Reactive (MDQR) market model, vectorized order matching, and an RL environment for accelerated policy evaluation. MDQR models event timing, type, location, and size, generating a sequence of market events which NV-PEG converts into individual orders and cancellations for replay. An optimized LOB matching engine processes these alongside the execution policy’s orders, determining when they trade, at what price, and for how many shares. A reinforcement learning pipeline enables the training and evaluation of learned trade execution policies alongside deterministic execution schedules and an open language model based policy by considering implementation shortfall, completion, and participation. We provide an extension to multi-agent settings where heterogeneous, dynamic policies interact in a shared order book. By reducing matching overhead across independent market simulations, NV-PEG aims to support larger experiments that expose execution failures, compare policies under different market conditions, and provide iterations for reinforcement learning. In a profiled prototype generation-and-replay workflow, optimized GPU order matching reduced matching’s share of runtime from 13.7% with GPU JAX-LOB to below 0.2%, effectively eliminating matching as a bottleneck and shifting computational throughput entirely to generative market modeling and policy optimization.