Paper: arXiv 2511.17963

Authors: Jun Kevin, Pujianto Yugopuspito

Abstract

This paper introduces a hybrid framework for portfolio optimization that fuses Long Short-Term Memory (LSTM) forecasting with a Proximal Policy Optimization (PPO) reinforcement learning strategy. The proposed system leverages the predictive power of deep recurrent networks to capture temporal dependencies, while the PPO agent adaptively refines portfolio allocations in continuous action spaces, allowing the system to anticipate trends while adjusting dynamically to market shifts. Using multi-asset datasets covering U.S. and Indonesian equities, U.S. Treasuries, and major cryptocurrencies from January 2018 to December 2024, the model is evaluated against several baselines, including equal-weight, index-style, and single-model variants (LSTM-only and PPO-only). The framework’s performance is benchmarked against equal-weighted, index-based, and single-model approaches (LSTM-only and PPO-only) using annualized return, volatility, Sharpe ratio, and maximum drawdown metrics, each adjusted for transaction costs. The results indicate that the hybrid architecture delivers higher returns and stronger resilience under non-stationary market regimes, suggesting its promise as a robust, AI-driven framework for dynamic portfolio optimization.

Complexity vs Empirical Score

  • Math Complexity: 8.0/10
  • Empirical Rigor: 7.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a sophisticated hybrid model combining LSTM and PPO with multiple derivations and complex loss functions, while also detailing a multi-asset dataset, specific backtesting metrics, and transaction cost adjustments, indicating strong empirical implementation.

Research Flowchart

  flowchart TD
  A["Research Goal: Dynamic Portfolio Optimization"] --> B["Hybrid LSTM-PPO Framework"]
  
  B --> C["Data Input<br>Multi-Asset (Jan 2018 - Dec 2024)<br>US/Indonesia Equities, Treasuries, Crypto"]
  
  C --> D["Process: LSTM Forecasting<br>Deep Recurrent Networks<br>Temporal Trend Prediction"]
  
  C --> E["Process: PPO Agent<br>Proximal Policy Optimization<br>Continuous Action Space Allocation"]
  
  D --> F["Fusion & Adaptation<br>Dynamic Rebalancing + Transaction Cost Adjustment"]
  E --> F
  
  F --> G["Benchmarking<br>vs. Equal-weight, Index, LSTM-only, PPO-only"]
  
  G --> H["Key Outcomes:<br>↑ Higher Annualized Returns<br>↑ Sharpe Ratio<br>↓ Max Drawdown<br>↑ Market Regime Resilience"]

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 hybrid portfolio optimization framework that combines LSTM forecasting with PPO reinforcement learning for dynamic allocation.

Source sentence
This paper introduces a hybrid framework for portfolio optimization that fuses Long Short-Term Memory (LSTM) forecasting with a Proximal Policy Optimization (PPO) reinforcement learning strategy.

Assumes 5

  • Transaction costs are modeled as a flat 0.1% turnover cost applied during portfolio evaluation. execution
    Source sentence
    Transaction costs are not applied at this stage; a 0.1% turnover cost is later included during portfolio evaluation to isolate the LSTM’s predictive effect.
  • Data are split chronologically into 70% training and 30% testing for strict out‑of‑sample evaluation. data
    Source sentence
    The dataset is then chronologically divided into 70% training and 30% testing subsets to maintain strict out-of-sample evaluation.
  • Missing values are forward‑filled and returns are computed as log‑returns to improve stationarity. data
    Source sentence
    Missing values were forward-filled, trading calendars synchronized, and closing prices transformed into log-returns
  • All price data are sourced from Yahoo Finance and resampled to weekly intervals. market
    Source sentence
    All data were sourced from Yahoo Finance (2018–2024) at daily frequency and resampled to weekly intervals.
Show 1 more
  • The PPO agent operates in continuous action spaces and uses a clipped‑surrogate reward based on cumulative return and Sharpe ratio. model
    Source sentence
    these signals into sparse portfolio weights that maximize cumulative net return under transaction cost and turnover constraints.

Data

Backtest sample: 2024-01 – 2024-12 · multi-asset (U.S. equities, Indonesian equities, bonds, cryp · multi-asset · weekly

  • U.S. equities (Nasdaq‑100) · Yahoo Finance market-prices public
    2018–2024 · Nasdaq‑100 constituents
    Source sentence
    ... All data were sourced from Yahoo Finance (2018–2024) at daily frequency
  • Indonesian equities (IDX30) · Yahoo Finance market-prices public
    2018–2024 · IDX30 constituents
    Source sentence
    ... All data were sourced from Yahoo Finance (2018–2024) at daily frequency
  • U.S. 10‑Year Treasury yields (ˆTNX) · Yahoo Finance market-prices public
    2018–2024 · U.S. Treasury yields
    Source sentence
    ... All data were sourced from Yahoo Finance (2018–2024) at daily frequency
  • Top 10 cryptocurrencies by market cap · Yahoo Finance crypto-onchain public
    2018–2024 · Bitcoin, Ethereum, XRP, Binance Coin, etc.
    Source sentence
    ... All data were sourced from Yahoo Finance (2018–2024) at daily frequency

Recognised sources: Yahoo Finance

Costs

Modeled Costs are part of the reported headline results.

A flat 0.1% turnover cost is applied during portfolio evaluation; results are reported net of transaction costs where applicable.

Source sentence
Transaction costs are not applied at this stage; a 0.1% turnover cost is later included during portfolio evaluation to isolate the LSTM’s predictive effect.

Finds 3

Reported performance (Hybrid LSTM+PPO (Top-5)): Sharpe 0.96 · annual return 25.4% · max drawdown 13.7%

  • The Hybrid LSTM+PPO Top‑5 portfolio achieves the highest annualized return of 25.4%, outperforming all single‑model and benchmark strategies. performance
    Source sentence
    Hybrid LSTM+PPO (Top-5) portfolio achieves the highest annualized return (25.4%), outperforming both stand-alone LSTM and PPO models as well as passive benchmarks.
  • Increasing the Top‑K diversification level reduces maximum drawdown, with Top‑30 achieving the shallowest drawdown of –10.6%. performance
    Source sentence
    The maximum drawdowns (MDD) occur around mid-2024, reaching approximately −13.7%, −12.0%, and −10.6% for Top-5, Top-10, and Top-30 respectively.
  • LSTM‑only strategies perform poorly at low K values, showing negative returns and high drawdowns, highlighting the limitation of static predictive models without adaptive allocation. performance
    Source sentence
    The LSTM-only strategies, while capable of capturing temporal dependencies, perform poorly at smaller K values due to limited diversification and their static allocation nature.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 40 related papers). Coverage is partial, so this is neither replication nor refutation.

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 (30)

  1. Forecasting cryptocurrency returns using classical statistical and deep learning techniques (2024) — link
  2. Cryptocurrency price forecasting – A comparative analysis of ensemble learning and deep learning methods (2024) — link
  3. Optimizing Portfolio with Two-Sided Transactions and Lending: A Reinforcement Learning Framework — → 2408.05382
  4. Bitcoin replication using machine learning (2024) — link
  5. A Comprehensive Analysis of Machine Learning Models for Algorithmic Trading of Bitcoin — → 2407.18334
  6. Collective Dynamics, Diversification and Optimal Portfolio Construction for Cryptocurrencies — → 2304.08902
  7. Machine learning for cryptocurrency market prediction and trading (2022) — link
  8. Deep Learning Forecasting in Cryptocurrency High-Frequency Trading (2021) — link
and 22 more
  1. A Contrastive Deep Learning Approach to Cryptocurrency Portfolio with US Treasuries (2024) — link
  2. Predictive crypto-asset automated market maker architecture for decentralized finance using deep reinforcement learning (2024) — link
  3. Dynamic portfolio choice with uncertain rare-events risk in stock and cryptocurrency markets (2023) — link
  4. Using Deep Reinforcement Learning with Hierarchical Risk Parity for Portfolio Optimization (2023) — link
  5. Deep Learning for High-Frequency Cryptocurrency Trend Detection Incorporating Technical Indicators and a New Approach For Data Stationarity (2024) — link
  6. Extending Deep Reinforcement Learning Frameworks in Cryptocurrency Market Making (2020)
  7. Forecasting Cryptocurrency Prices Using LSTM, GRU, and Bi-Directional LSTM: A Deep Learning Approach (2023) — link
  8. Deep graph convolutional reinforcement learning for financial portfolio management – DeepPocket (2021) — link
  9. Reinforcement learning in optimisation of financial market trading strategy parameters (2024) — link
  10. Deep Reinforcement Learning: A Survey (2024) — link
  11. DeepTrader: A Deep Reinforcement Learning Approach for Risk-Return Balanced Portfolio Management with Market Conditions Embedding (2021) — link
  12. Applications of Markov Decision Process Model and Deep Learning in Quantitative Portfolio Management during the COVID-19 Pandemic (2022) — link
  13. Application of Deep Reinforcement Learning for Cryptocurrency Market Trend Forecasting and Risk Management (2024) — link
  14. Maximizing portfolio profitability during a cryptocurrency downtrend: A Bitcoin Blockchain transaction-based approach (2023)
  15. Top 100 Cryptocurrencies by Market Capitalization (2025) — link
  16. Portfolio Optimization through Hybrid Deep Learning and Genetic Algorithms: Vine Copula–GARCH–EVT–CVaR Model (2023) — link
  17. Reinforcement-Learning-Based Portfolio Management with Augmented Asset Movement Prediction States (2020) — link
  18. Multi-Period Portfolio Optimization Using a Deep Reinforcement Learning Hyper-Heuristic Approach (2023) — link
  19. Outperforming the Tutor: Expert-Infused Deep Reinforcement Learning for Dynamic Portfolio Selection of Diverse Assets (2024) — link
  20. A Deep Learning Framework for Financial Time Series Using Stacked Autoencoders and Long Short-Term Memory (2017) — link
  21. Decision-Making for Financial Trading: A Fusion Approach of Machine Learning and Portfolio Optimization (2019) — link
  22. A Hybrid Deep Learning Model for Stock Index Prediction Using LSTM and GARCH (2020) — link

3 of 30 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.