Paper: arXiv 2509.14385

Authors: Gabriel Nixon Raj

Abstract

This study proposes a regime-aware reinforcement learning framework for long-horizon portfolio optimization. Moving beyond traditional feedforward and GARCH-based models, we design realistic environments where agents dynamically reallocate capital in response to latent macroeconomic regime shifts. Agents receive hybrid observations and are trained using constrained reward functions that incorporate volatility penalties, capital resets, and tail-risk shocks. We benchmark multiple architectures, including PPO, LSTM-based PPO, and Transformer PPO, against classical baselines such as equal-weight and Sharpe-optimized portfolios. Our agents demonstrate robust performance under financial stress. While Transformer PPO achieves the highest risk-adjusted returns, LSTM variants offer a favorable trade-off between interpretability and training cost. The framework promotes regime-adaptive, explainable reinforcement learning for dynamic asset allocation.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 6.0/10 (capped from 7.5: transaction costs not modelled; no out-of-sample test; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.5)

Why this score: The paper employs advanced mathematics in regime detection (HMM, GMM), Monte Carlo simulations, and RL architectures (Transformer, LSTM), while demonstrating empirical rigor through multi-decade financial datasets, stress-test alignment with historical crises, and comparative benchmarking against classical baselines.

Research Flowchart

  flowchart TD
  A["Research Goal: Adaptive Portfolio Optimization<br>under Macroeconomic Regime Shifts"] --> B["Data & Inputs<br>Multi-Asset Market Data"]
  B --> C["Methodology<br>Regime-Aware RL Environment Construction"]
  C --> D["Computational Process<br>Agent Training: PPO, LSTM-PPO, Transformer-PPO"]
  D --> E["Key Findings & Outcomes<br>1. Transformer-PPO: Highest Risk-Adjusted Returns<br>2. LSTM-PPO: Best Interpretability/Cost Trade-off<br>3. Robust Performance under Stress"]

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 a regime-aware reinforcement learning framework for long-horizon portfolio optimization.

Source sentence
This study proposes a regime-aware reinforcement learning framework for long-horizon portfolio optimization.

Assumes 2

  • The framework assumes stationarity in regime transition probabilities. model
    Source sentence
    our current framework assumes stationarity in regime transition probabilities; relaxing this assumption using adaptive or Bayesian models may improve real-world alignment
  • The simulation environment abstracts away real‑world frictions such as slippage, liquidity constraints, and execution latency. execution
    Source sentence
    the simulation environment, while stylized and stress-informed, abstracts away real-world frictions such as slippage, liquidity constraints, and execution latency.

Data

  • Public annual returns dataset · multiple vendors (Bloomberg, Morningstar, Capital IQ, Compus market-prices public
    several decades covering 1931, 1974, 198 · equities, fixed income, real estate, commodities, short‑term risk‑free assets
    Source sentence
    ... It contains annual historical returns across several decades and includes a diverse set of asset classes such as equities (e.g., S&P 500, Small Cap), fixed income (10-Year Treasuries, Baa Corporate Bonds), real estate (REIT proxies), commodities (Gold), and short-term risk-free assets (3-Month T-Bills). The data sp

Costs

Unclear - sources disagree The platform classifier and the paper-text extraction give different answers, so no value is asserted.

Platform classifier: Not modeled

Paper text extraction: Partial - A transaction penalty is included in the reward to discourage excessive turnover, but no explicit cost model (e.g., bps) is reported.

Source sentence (the authors' words)
Transaction penalty: discourages excessive portfolio turnover by penalizing abrupt weight shifts.

Finds 5

  • Transformer‑based PPO achieves the highest risk‑adjusted returns, with a Sharpe ratio of 1.07 and Sortino of 1.20, outperforming prior RL baselines. performance
    Source sentence
    Ours (PPO + Regime) 1.07 1.20 –72.58% $1.1×1012
  • LSTM‑based PPO offers a favorable trade‑off between interpretability and training cost compared to Transformer PPO. performance
    Source sentence
    While Transformer PPO achieves the highest risk‑adjusted returns, LSTM variants offer a favorable trade‑off between interpretability and training cost.
  • The agents demonstrate robust performance under financial stress, with improved learning stability and drawdown resilience due to regime‑aware features. performance
    Source sentence
    ... incorporating regime context significantly improves learning stability, financial performance, and decision
  • The PPO agent exhibits higher drawdown in long‑horizon scenarios, reflecting an aggressive growth posture. negative
    Source sentence
    the PPO agent’s higher drawdown in long‑horizon scenarios reflects an aggressive growth posture that may be unsuitable for capital‑constrained institutional mandates.
  • SHAP interpretability analysis validates that the agent learns from meaningful economic signals rather than noise. other
    Source sentence
    Through interpretable diagnostics like SHAP, we further validated that the agent was not merely fitting noise but learning from meaningful economic signals.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 54 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

https://github.com/GabrielNixon/RegimeAware-PPO — the authors' own repository

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