Optimization of Dynamic Crypto Portfolios via Deep Reinforcement Learning Under Transaction Costs (Research Square)

Summary

• The DRL-PPO strategy generated an annualized return of 37.22% under a 0.1% transaction cost constraint.p.2
• The tested universe comprised 10 liquid cryptocurrencies analyzed over daily price series from January 2021 to January 2024.p.6
• DRL-PPO achieved the lowest maximum drawdown among all evaluated strategies at -28.74%.p.10

Key insights

1. DRL-PPO achieved market neutrality with a beta of -0.08 and generated a positive Jensen's alpha of 46.41%.p.2
2. Supervised predictive models exhibited market betas near unity, leaving portfolios exposed to systematic market downturns.p.2
3. Cryptocurrency market assets display cross-asset correlation coefficients frequently exceeding 0.60, limiting the benefits of passive diversification.p.9
4. Adding a zero-return risk-free CASH asset to the action space provides endogenous portfolio hedging during severe bear markets.p.8
5. Explicitly penalizing transaction frictions within the DRL reward function prevents over-trading and mitigates turnover drag.p.3
6. Traditional predict-then-optimize frameworks degrade net performance because point-forecasting errors spur excessive rebalancing.p.3