FR–LUX: FRICTION–AWARE, REGIME–CONDITIONED POLICY OPTIMIZATION FOR IMPLEMENTABLE PORTFOLIO MANAGEMENT

Battle Count: 91

This paper is directly and highly relevant to quantitative trading. It addresses the core challenge of translating ML-based return forecasts into implementable trading policies under realistic transaction costs. The framework explicitly models execution costs (proportional + impact), controls turnover via trade-space trust regions, conditions on volatility-liquidity regimes, and provides formal guarantees on after-cost performance. The evaluation protocol with multiple-testing controls, cost sensitivity curves, and regime-stratified aggregation mirrors institutional-grade backtesting standards. The method is designed for production deployment with OMS/EMS integration guidance and capacity management.

The implementation requires: (1) a PPO/TRPO training loop adapted for continuous portfolio weights with feasibility projections (simplex or L1-ball); (2) regime construction from volatility and liquidity diagnostics with rolling quantile thresholds; (3) cost calibration from spread/ILLIQ proxies with regime-dependent scaling; (4) CVaR or smoothed MDD risk penalties with alternating optimization; (5) trade-space trust region penalties; (6) regime-balanced objective with mixture-of-experts or regime embeddings; (7) GAE advantage estimation on after-cost rewards; (8) multiple-testing inference (Romano-Wolf stepdown, HAC-aware Sharpe comparisons). The algorithmic template is provided but requires careful hyperparameter tuning (KL budget, trade penalty, clip epsilon, risk weight) and proper cost calibration infrastructure.

The paper explicitly states that all figures and tables are reproducible from artifacts. They release data pre-processing scripts, exact configuration files for each scenario and seed, training logs and random seeds, and plotting code. All numerical tables are generated from the same artifacts with hashes and timestamps. However, no specific GitHub URL is provided in the extract, and no rendered numerical tables are included (results referenced to Figures 1-7). The evaluation protocol is well-documented with implementation parity across methods.