Paper: arXiv 2510.02986

Authors: Jian’an Zhang

Abstract

Transaction costs and regime shifts are major reasons why paper portfolios fail in live trading. We introduce FR-LUX (Friction-aware, Regime-conditioned Learning under eXecution costs), a reinforcement learning framework that learns after-cost trading policies and remains robust across volatility-liquidity regimes. FR-LUX integrates three ingredients: (i) a microstructure-consistent execution model combining proportional and impact costs, directly embedded in the reward; (ii) a trade-space trust region that constrains changes in inventory flow rather than logits, yielding stable low-turnover updates; and (iii) explicit regime conditioning so the policy specializes to LL/LH/HL/HH states without fragmenting the data. On a 4 x 5 grid of regimes and cost levels with multiple random seeds, FR-LUX achieves the top average Sharpe ratio with narrow bootstrap confidence intervals, maintains a flatter cost-performance slope than strong baselines, and attains superior risk-return efficiency for a given turnover budget. Pairwise scenario-level improvements are strictly positive and remain statistically significant after multiple-testing corrections. We provide formal guarantees on optimality under convex frictions, monotonic improvement under a KL trust region, long-run turnover bounds and induced inaction bands due to proportional costs, positive value advantage for regime-conditioned policies, and robustness to cost misspecification. The methodology is implementable: costs are calibrated from standard liquidity proxies, scenario-level inference avoids pseudo-replication, and all figures and tables are reproducible from released artifacts.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents advanced mathematical formulations including MDP modeling, convex optimization theory, and formal proofs for policy improvement, scoring high on math complexity. It also demonstrates strong empirical rigor with a comprehensive backtest framework on 20 scenarios, robust statistical inference (Romano-Wolf corrections), and reproducible artifacts.

Research Flowchart

  flowchart TD
  A["Research Goal<br>How to build robust, implementable trading policies<br>that account for execution costs and market regimes?"] --> B["Methodology: FR-LUX Framework<br>(Friction-aware, Regime-conditioned RL)"]
  B --> C["Key Components<br>1. Microstructure Cost Model (Proportional + Impact)<br>2. Trade-Space Trust Region (Inventory Flow Constraints)<br>3. Explicit Regime Conditioning (LL/LH/HL/HH)"]
  C --> D["Data & Inputs<br>• Multi-regime scenarios (4x5 grid)<br>• Calibrated transaction costs<br>• Liquidity proxies"]
  D --> E["Computational Process<br>Reinforcement Learning Optimization<br>with Cost-Aware Rewards & Trust Region"]
  E --> F["Key Findings & Outcomes<br>• Top Sharpe with tight confidence intervals<br>• Flatter cost-performance slope vs. baselines<br>• Positive, significant scenario improvements<br>• Theoretical guarantees (optimality, robustness)"]
  F --> G["Outcome: Implementable, Robust Portfolio Policies<br>Reproducible artifacts released"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes FR‑LUX, a reinforcement‑learning framework that learns after‑cost trading policies with regime conditioning and a trade‑space trust region.

Source sentence
We develop FR–LUX (Friction–aware, Regime–conditioned Learning under eXecution costs), a reinforcement‑learning framework that learns after‑cost trading policies and remains robust across volatility‑liquidity regimes.

Assumes 5

  • Action set W is non‑empty, convex, compact. model
    Source sentence
    (i) Action set W ⊂Rd is nonempty, convex, compact.
  • Execution cost C_z(Δw) is convex, lower semicontinuous, zero at no trade, and bounded below by a linear term. execution
    Source sentence
    The execution cost C_z(Δw) is convex, lower semicontinuous, C_z(0) = 0, and satisfies C_z(u) ≥κ1(z)∥u∥1 for some κ1(z) > 0.
  • Downside‑risk proxy Ψ is non‑negative and Lipschitz on compact sets. model
    Source sentence
    The downside‑risk proxy Ψ in (1) is nonnegative and Lipschitz in the portfolio loss on compact sets.
  • The controlled process (s_t, z_t) is Markov and β‑mixing under any stationary policy. statistical
    Source sentence
    The controlled process (st, zt) is Markov and β‑mixing under any stationary policy.
Show 1 more
  • Policy class π_θ is continuously differentiable and either a mixture‑of‑experts with regime‑gated experts or a single head with learned regime embedding; the induced action map is Lipschitz. model
    Source sentence
    πθ(· | s, z) is continuously differentiable in θ, and either (i) a mixture‑of‑experts (MoE) with regime‑gated experts, or (ii) a single head with a learned regime embedding e(z); the induced action map a = ProjW(gθ(s, z)) is Lipschitz.

Data

No dataset named in the text.

Costs

Modeled Costs are part of the reported headline results.

A microstructure‑consistent execution model combines proportional and impact costs and is embedded directly in the reward.

Source sentence
a microstructure‑consistent execution model combining proportional and impact costs, directly embedded in the reward

Finds 5

  • FR‑LUX achieves the top average Sharpe ratio across all 20 regime‑cost scenarios with narrow bootstrap confidence intervals. performance
    Source sentence
    FR‑LUX achieves the top average Sharpe across all 20 scenarios with narrow bootstrap confidence intervals
  • FR‑LUX maintains a flatter cost‑performance slope than strong baselines such as vanilla PPO and risk‑parity. performance
    Source sentence
    maintains a flatter cost‑performance slope than strong baselines (vanilla PPO, mean‑variance with/without caps, risk‑parity)
  • FR‑LUX attains superior risk‑return efficiency for a given turnover budget. performance
    Source sentence
    attains superior risk‑return efficiency for a given turnover budget
  • Pairwise scenario‑level improvements of FR‑LUX over benchmarks are strictly positive and remain statistically significant after Romano‑Wolf stepdown and HAC‑aware Sharpe comparisons. statistical
    Source sentence
    Pairwise scenario‑level improvements are strictly positive and remain statistically significant after Romano‑Wolf stepdown and HAC‑aware Sharpe comparisons
  • The method provides theoretical guarantees including existence of an optimal stationary policy, monotonic improvement under a KL trust region, and long‑run turnover bounds. theoretical
    Source sentence
    We provide formal guarantees: existence of an optimal stationary policy under convex frictions; a monotonic improvement lower bound under a KL trust region with explicit remainder terms; an upper bound on long‑run turnover and an induced inaction band due to proportional costs

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 47 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive (1 by independent authors)

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (52)

  1. Portfolio Selection (1952)
  2. Optimal Execution of Portfolio Transactions (2001)
  3. Continuous Auctions and Insider Trading (1985)
  4. Illiquidity and Stock Returns: Cross-Section and Time-Series Effects (2002)
  5. Liquidity Risk and Expected Stock Returns (2003)
  6. Trading Costs and Returns for U.S. Equities: Estimating Effective Costs from Daily Data (2009) — link
  7. Do Liquidity Measures Measure Liquidity? (2009)
  8. What Are the Best Liquidity Proxies for Global Research? (2017)
and 44 more
  1. Optimal Trading Strategy and Supply/Demand Dynamics (2013)
  2. Dynamic Trading with Predictable Returns and Transaction Costs (2013) — link
  3. A Taxonomy of Anomalies and Their Trading Costs (2016)
  4. Volatility-Managed Portfolios (2017) — link
  5. Nonlinear Shrinkage of the Covariance Matrix for Portfolio Selection: Markowitz Meets Goldilocks (2017)
  6. Comparing Asset Pricing Models (2018)
  7. Model Comparison with Sharpe Ratios (2020) — link
  8. Empirical Asset Pricing via Machine Learning (2020)
  9. Dissecting Characteristics Nonparametrically (2020)
  10. Shrinking the Cross Section (2020)
  11. Deep Learning in Asset Pricing (2024) — link
  12. Importance of Transaction Costs for Asset Allocation in Foreign Exchange Markets (2024) — link
  13. Size Discount and Size Penalty: Trading Costs in Bond Markets (2024) — link
  14. International Asset Allocation with Regime Shifts (2002)
  15. Parametric Portfolio Policies: Exploiting Characteristics in the Cross-Section of Equity Returns (2009)
  16. A Reality Check for Data Snooping (2000)
  17. Exact and Approximate Stepdown Methods for Multiple Hypothesis Testing (2005)
  18. The Statistics of Sharpe Ratios (2002)
  19. Approximately Optimal Approximate Reinforcement Learning (2002)
  20. Trust Region Policy Optimization (2015)
  21. Proximal Policy Optimization Algorithms (2017)
  22. Constrained Policy Optimization (2017)
  23. A Transaction-Cost Perspective on the Multitude of Firm Characteristics (2020)
  24. A Review of Reinforcement Learning in Financial Applications — → 2411.12746
  25. Machine Learning and International Stock Returns (2025) — link
  26. Markowitz portfolios under transaction costs (2025) — link
  27. A Review of Reinforcement Learning in Financial Applications — → 2411.12746
  28. Liquidity Regimes and Optimal Dynamic Asset Allocation (2020) — link
  29. Algorithmic and High-Frequency Trading (2015)
  30. High-Dimensional Continuous Control Using Generalized Advantage Estimation (2016)
  31. Algorithms for CVaR Optimization in MDPs (2014)
  32. Efficient Risk-Averse Reinforcement Learning (2022)
  33. Optimization of Conditional Value-at-Risk (2000)
  34. Conditional Value-at-Risk for General Loss Distributions (2002)
  35. Expected Shortfall: A Natural Coherent Alternative to Value at Risk (2002)
  36. Efficient Projections onto the $\ell_1$-Ball for Learning in High Dimensions (2008)
  37. Projection onto the Probability Simplex: An Efficient Algorithm with a Simple Proof, and an Application (2013)
  38. Adaptive Mixtures of Local Experts (1991)
  39. Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer (2017)
  40. Markov Decision Processes: Discrete Stochastic Dynamic Programming (1994)
  41. Safe Policy Iteration (2013)
  42. Infinite Dimensional Analysis: A Hitchhiker’s Guide (2006)
  43. Elements of Information Theory (2006)
  44. Nonlinear Programming (1999)

2 of 52 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.