Paper: arXiv 2412.18563
Authors: Gang Huang, Xiaohua Zhou, Qingyang Song
Abstract
Artificial intelligence is transforming financial investment decision-making frameworks, with deep reinforcement learning demonstrating substantial potential in robo-advisory applications. This paper addresses the limitations of traditional portfolio optimization methods in dynamic asset weight adjustment through the development of a deep reinforcement learning-based dynamic optimization model grounded in practical trading processes. The research advances two key innovations: first, the introduction of a novel Sharpe ratio reward function engineered for Actor-Critic deep reinforcement learning algorithms, which ensures stable convergence during training while consistently achieving positive average Sharpe ratios; second, the development of an innovative comprehensive approach to portfolio optimization utilizing deep reinforcement learning, which significantly enhances model optimization capability through the integration of random sampling strategies during training with image-based deep neural network architectures for multi-dimensional financial time series data processing, average Sharpe ratio reward functions, and deep reinforcement learning algorithms. The empirical analysis validates the model using randomly selected constituent stocks from the CSI 300 Index, benchmarking against established financial econometric optimization models. Backtesting results demonstrate the model’s efficacy in optimizing portfolio allocation and mitigating investment risk, yielding superior comprehensive performance metrics.
Complexity vs Empirical Score
- Math Complexity: 7.0/10
- Empirical Rigor: 8.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced mathematical concepts typical of deep reinforcement learning and financial econometrics (like Sharpe ratio derivations and DRL algorithms), while providing a backtest-ready empirical validation using real market data (CSI 300 stocks) with benchmarks against established models.
Research Flowchart
flowchart TD A["Research Goal<br>Dynamic Portfolio Optimization<br>vs Traditional Methods"] --> B["Data & Inputs<br>CSI 300 Constituent Stocks<br>Financial Time Series"] B --> C["Methodology: Deep RL Framework<br>Actor-Critic Algorithms<br>with Novel Sharpe Ratio Reward"] C --> D["Computational Process<br>Image-based DNN for Data Processing<br>Random Sampling for Training"] D --> E["Backtesting & Optimization<br>Dynamic Asset Weight Adjustment"] E --> F["Key Outcomes<br>Superior Performance & Risk Mitigation<br>Positive Sharpe Ratios Achieved"]
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 novel Sharpe ratio reward function for Actor-Critic DRL algorithms and a comprehensive DRL methodology integrating VGG-based networks and random sampling for dynamic portfolio optimization.
Source sentence
First, we introduce a novel Sharpe ratio reward function specifically designed for Actor-Critic DRL algorithms, achieving stable convergence during training and consistently positive Sharpe ratios in practice.
Assumes 6
- Short‑selling is prohibited, i.e., only long‑only positions are allowed. market
Source sentence
Under short-selling restrictions (long-only positions), the reward function in current DRL portfolio weight optimization models primarily consists of portfolio returns
- Only one transaction per stock per day is allowed. execution
Source sentence
The study restricts trading to one transaction per stock per day.
- All price data are forward‑adjusted daily records from the Wind database. data
Source sentence
The data is obtained from the Wind database's daily trading records, with all prices forward-adjusted.
- The portfolio contains one risk‑free asset (cash) and ten risky assets. model
Source sentence
The portfolio consists of 1 risk-free asset (cash) and 10 risky assets.
Show 2 more
- Backtesting uses exclusively out‑of‑sample data that were not seen during training. execution
Source sentence
The backtesting period encompasses exclusively out-of-sample data, independent from the training dataset.
- Randomly selected constituent stocks from the CSI 300 index represent a generic portfolio for testing DRL adaptability. other
Source sentence
This study constructs an investment portfolio using randomly selected constituent stocks from the CSI300 index for empirical analysis.
Data
2023-08 – 2024-02 · CSI 300 constituent stocks · equities · daily
- CSI300 constituent stocks (random selection) market-prices commercial
Source sentence
This study constructs an investment portfolio using randomly selected constituent stocks from the CSI300 index for empirical analysis. The portfolio consists of 1 risk-free asset (cash) and 10 risky assets. The data is obtained from the Wind database's daily trading records
Wind
Costs
Modeled
Initial allocation incurred ~0.22% cost; subsequent daily costs ranged between 0.002% and 0.01% on most trading days.
Source sentence
Substantial transaction costs (approximately 0.22%) were incurred solely during initial capital allocation, with subsequent transaction cost rates maintaining minimal levels between 0.002% and 0.01% on most trading days.
Finds 2
Sharpe 1.55 · annual return 19.6% · total return 12.6% · max drawdown 5.8%
- The DRL model achieved the highest annualized average return of 19.56%, a Sharpe ratio of 1.5550, Sortino ratio of 2.9567, maximum drawdown of 5.85% and Calmar ratio of 3.3395, outperforming all benchmark models. performance
Source sentence
The model achieved the highest annualized average return of 19.56%, maintained moderate volatility (12.58%), and notably outperformed other models with a Sharpe ratio of 1.5550
- Weight adjustments remained stable (standard deviation 0.031‑0.033) with cash holdings between 0.11 and 0.12, indicating effective diversification and liquidity management. performance
Source sentence
weight standard deviation stabilizing between 0.031 and 0.033, demonstrating effective diversification characteristics. The cash position fluctuated within a narrow range of 0.11 to 0.12, ensuring adequate portfolio liquidity.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 26 related papers). Coverage is partial, so this is neither replication nor refutation.
2 papers in the archive
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.