Abbade and Reali Costa 2026
“Realistic Market Impact Modeling for Reinforcement Learning Trading Environments” by Lucas Riera Abbade and Anna Helena Reali Costa (2026, arXiv) introduces a suite of Gymnasium-compatible RL trading environments (MACE stock trading, margin, portfolio optimisation) with pluggable nonlinear market-impact models grounded in Almgren–Chriss and the square-root law. Comparing a flat 10 bp fee against an impact model, the authors show flat fees let agents learn unrealistic high-turnover behaviour: one agent’s daily costs fall from 8k (turnover 19% to 1%) and a tuned PPO drops from 34% to 25% out-of-sample return under realistic costs. It appears in this vault as direct evidence that for MDP/RL trading, the cost model is part of the environment and changes the learned policy itself.
Connections
- Transaction Costs and Slippage — includes_costs, source: https://arxiv.org/html/2603.29086
- Markov Decision Process Trading Model — tests_strategy, source: https://arxiv.org/html/2603.29086
- Reinforcement Learning Trading Policy — tests_strategy, source: https://arxiv.org/html/2603.29086
- Square-Root Law of Market Impact — uses_dataset, calibrates environments with the square-root impact law, source: https://arxiv.org/html/2603.29086
- Almgren Chriss 2000 — uses_dataset, calibrates environments with the Almgren-Chriss framework, source: https://arxiv.org/html/2603.29086