Reinforcement Learning-Based Cryptocurrency Portfolio Management Using Soft Actor-Critic and Deep Deterministic Policy Gradient Algorithms

Battle Count: 87

The paper is directly relevant to quantitative trading, addressing cryptocurrency portfolio optimization using state-of-the-art DRL methods. It provides concrete performance comparisons (SAC vs DDPG vs Markowitz), detailed hyperparameter configurations, and a realistic backtesting framework. The focus on risk-adjusted metrics (Sharpe, Sortino, VaR, CVaR, MDD) aligns with institutional quantitative trading practices. The entropy-regularized SAC approach and LSTM-based feature extraction are directly applicable to real-world trading systems. However, the limited asset universe (4 coins) and absence of live trading validation slightly reduce practical applicability.

Implementation requires: (1) custom gym-based trading environment with transaction cost modeling, (2) LSTM-based feature extraction networks for multiple assets, (3) actor-critic architectures for both DDPG and SAC with target networks and replay buffers, (4) entropy-regularized training for SAC with automatic temperature tuning, (5) OU noise exploration for DDPG, (6) softmax normalization for portfolio weights, (7) Markowitz optimization via SciPy as benchmark, and (8) a separate forecasting network for validation. The multi-component architecture and careful hyperparameter tuning add significant complexity, though individual algorithms are well-documented in literature.

The paper provides detailed hyperparameter tables for SAC, DDPG, and the forecasting network. Data sources (yfinance) and time periods (2016-2022 training, 2023-2024 testing) are specified. A custom gym-based simulator is mentioned but no code repository is provided. Network architectures are described but exact layer configurations beyond hidden units are not fully specified. The absence of a public code repository limits full reproducibility.