Purpose: This study aims to propose and evaluate a dynamic framework for portfolio optimization in cryptocurrency markets that ensures an effective balance between return maximization and risk control under conditions of high volatility. The primary objective is to develop a risk-sensitive adaptive agent that integrates multiple baseline deep reinforcement learning agents with a dynamic switching mechanism to adjust asset weight allocations over time. This adaptive structure enables conservative behavior during highly volatile market regimes and more aggressive positioning during trending periods. The empirical analysis is conducted on a selected set of five major cryptocurrencies—Bitcoin, Ethereum, Solana, Ripple, and Tether—to assess the effectiveness and practical applicability of the proposed framework in real-world market environments.

Method: Daily data spanning five years were collected from Yahoo Finance and, after data cleaning, were divided into training and testing sets using a 70%–30% split. The state representation comprised asset prices, trading volumes, and commonly used technical indicators. To capture long-term temporal dependencies, state encoding was implemented using a Transformer-based architecture. Two categories of methods were examined: (1) individual deep reinforcement learning agents—namely PPO, A2C, and DQN—each independently managing the portfolio; and (2) a risk-sensitive adaptive framework that dynamically switches among agents by monitoring short-term performance indicators such as rolling cumulative profit and loss and five-day moving volatility. The reward function incorporated risk-sensitive components, including a logarithmic utility of returns, a volatility penalty, and the Calmar ratio. Performance evaluation was conducted using cumulative return, annualized return, Sharpe ratio, Sortino ratio, Calmar ratio, and maximum drawdown.

Findings: On the out-of-sample test dataset, the proposed risk-sensitive adaptive framework consistently outperformed each individual deep reinforcement learning agent. During non-trending market periods, the proposed model achieved higher terminal portfolio value and superior risk-adjusted performance compared to alternative methods. For instance, it recorded a Sharpe ratio of approximately 1.13 and a Sortino ratio of about 1.81. In bullish market regimes, the framework delivered markedly stronger results, achieving a Sharpe ratio of approximately 1.73, a Calmar ratio of around 3.34, and higher annualized returns, such that the final portfolio value during the uptrend was nearly twice that of the best-performing standalone agent. The quantitative results indicate that dynamic switching among aggressive, balanced, and conservative agents enhances the exploitation of favorable market trends while mitigating losses during highly volatile periods. Although a temporary increase in maximum drawdown was observed in certain scenarios, this effect was effectively offset by a reversion mechanism that shifts the agent back toward conservative behavior.

Conclusion: Integrating a Transformer architecture for temporal feature extraction with a risk-sensitive adaptive agent framework provides an effective solution for dynamic portfolio management in cryptocurrency markets, particularly when the objective is to simultaneously enhance returns and control downside risk. The findings indicate that employing a dynamic switching mechanism among multiple deep reinforcement learning agents can significantly improve portfolio performance and stability when facing shifting market regimes. Furthermore, statistical significance analysis using the Wilcoxon Signed-Rank Test across ten independent random seeds confirmed that the proposed model significantly outperformed PPO, A2C, and DQN in all major performance metrics (p < 0.01), demonstrating both robustness and reproducibility of the results.