Comparative Evaluation of Modern Deep Learning Methodologies for Portfolio Optimization

1. Problem Statement

Traditional portfolio optimization methods, such as Modern Portfolio Theory (MPT), the Capital Asset Pricing Model (CAPM), and the Black-Litterman model, rely heavily on static assumptions (e.g., normally distributed returns, static covariance matrices) and linear relationships. These models struggle in modern financial environments characterized by:

• Non-stationarity and Non-linearity: Markets exhibit time-varying correlations and complex, non-linear dependencies that static models cannot capture.
• High Dimensionality: The sheer volume of assets and data sources (news, sentiment, fundamentals) overwhelms traditional feature engineering.
• Estimation Error: MVO is highly sensitive to input parameters; small errors in expected return or covariance estimates can lead to unstable and suboptimal portfolios.

The study addresses the gap in literature regarding the integration of advanced deep learning architectures with traditional financial theory to create robust, adaptive, and risk-aware portfolio strategies.

2. Methodology

The study employs a comparative framework using historical daily data (2015–2023) across diverse asset classes (Equities, ETFs, Bonds, Gold). The methodology involves constructing, training, and backtesting seven distinct strategies.

A. Data & Preprocessing

• Assets: AAPL, TLT, AMZN, SPY, EFA, GLD, MSFT.
• Features: Daily returns, technical indicators (volatility, moving averages, momentum), standardized to ensure model stability.
• Backtesting: Simulated trading with transaction costs, dynamic rebalancing, and risk constraints (max volatility/drawdown).

B. Model Architectures

The research evaluates four primary deep learning paradigms and their hybrid combinations:

1. Deep Reinforcement Learning (DRL):

   • Goal: Learn optimal portfolio weights dynamically to maximize the Sharpe ratio.
   • Mechanism: Uses a PPO (Proximal Policy Optimization) agent interacting with the market environment.
   • Limitation Tested: Standalone DRL often suffers from sample inefficiency and lack of structural awareness.

2. Autoencoders (AE):

   • Goal: Dimensionality reduction and noise filtering.
   • Mechanism: Encodes high-dimensional return streams into a lower-dimensional latent space ($Z$) and reconstructs them.
   • Usage: Used as a feature compressor for DRL inputs or as a standalone covariance estimator.

3. Graph Neural Networks (GNN):

   • Goal: Model inter-asset dependencies and evolving correlation structures.
   • Mechanism: Treats assets as nodes in a graph ($\mathcal{G} = (\mathcal{V}, \mathcal{E})$) and uses message passing to update node representations based on neighbors.
   • Usage: Specifically used to estimate the covariance matrix for optimization.

4. Transformers:

   • Goal: Capture long-range temporal dependencies in asset returns.
   • Mechanism: Uses self-attention mechanisms to weigh the importance of past time steps for future return prediction.
   • Usage: Generates predicted return vectors ($\mu$).

C. Hybrid Strategies

• Autoencoder + DRL: Uses latent features from the AE to stabilize DRL training.
• Transformer + GNN: Combines Transformer-predicted returns with GNN-estimated covariance matrices, feeding both into a Mean-Variance Optimization (MVO) framework.

D. Benchmarks

The deep learning models are compared against:

• Markowitz MVO: Using predicted inputs.
• Equal-Weighted (1/N): A simple static baseline.
• 60/40 Portfolio: A traditional fixed-allocation strategy.

E. Evaluation Metrics

• Return: Cumulative Return, Annualized Return (CAGR).
• Risk: Annualized Volatility, Maximum Drawdown (MDD).
• Efficiency: Sharpe Ratio.

3. Key Contributions

1. Hybrid Architecture Validation: Demonstrates that combining complementary deep learning models (e.g., Transformers for time-series and GNNs for relational data) yields superior risk control compared to standalone models.
2. Re-evaluation of Traditional Models: Challenges the narrative that MVO is obsolete. The study shows that MVO remains highly competitive (and often superior in raw returns) when paired with high-quality, deep-learning-derived inputs.
3. Feature Compression Impact: Proves that dimensionality reduction via Autoencoders significantly stabilizes DRL training, preventing the "naive" DRL from underperforming.
4. Comprehensive Backtesting Framework: Provides a rigorous comparison of seven strategies under realistic constraints (transaction costs, rebalancing) over an 8-year period including various market regimes.

4. Results

The empirical analysis yielded the following key findings (based on Table 2 in the paper):

Strategy	Cumulative Return (%)	Volatility (%)	Sharpe Ratio	Max Drawdown (%)
MVO (with DL inputs)	461.31 (Highest)	19.38	1.0885 (Highest)	-31.82
Autoencoder	287.79	15.45	1.0546	-27.70
Autoencoder + DRL	273.76	15.26	1.0390	-26.52
Transformer + GNN	269.71	14.67 (Lowest)	1.0659	-18.81 (Lowest)
60/40	332.89	21.09	0.8798	-32.64
Equal-Weight	287.79	15.45	1.0546	-27.70
Standalone DRL	50.06	15.71	0.3671	-37.36

• Return Performance: The MVO strategy utilizing deep learning inputs achieved the highest cumulative return (461.31%) and Sharpe ratio (1.09), outperforming all deep learning-only strategies.
• Risk Management: The Transformer + GNN hybrid demonstrated the best risk control, achieving the lowest volatility (14.67%) and the lowest maximum drawdown (-18.81%).
• DRL Performance: Standalone DRL performed poorly (Sharpe 0.37), confirming that without structural priors or feature compression, DRL struggles to learn optimal policies in noisy financial data.
• Autoencoder Behavior: The standalone Autoencoder strategy mirrored the Equal-Weight portfolio, indicating that feature compression alone does not generate alpha without a dynamic allocation mechanism.

5. Significance and Conclusion

The study concludes that no single methodology universally dominates. Instead, the optimal approach lies in the synergy between classical financial theory and modern deep learning:

• Complementarity: Deep learning excels at feature extraction (Transformers for time, GNNs for structure, Autoencoders for noise reduction), while traditional optimization (MVO) excels at efficient portfolio construction given those inputs.
• Robustness: Hybrid models (specifically Transformer+GNN) offer superior stability and downside protection, making them ideal for risk-sensitive institutional environments.
• Practicality: The research validates that traditional models are not obsolete; rather, their performance is contingent on the quality of input data. By replacing manual estimates with deep learning predictions, MVO regains its competitiveness.

Future Directions: The authors suggest exploring the integration of latent financial representations directly into traditional optimization frameworks to further enhance scalability and interpretability in large-scale institutional deployments.