CVaR-based risk parity model with machine learning

Abstract

This study proposes a risk parity model based on conditional value-at-risk (CVaR), enhanced by integrating machine learning techniques into dynamic portfolio optimization. The CVaR-based risk parity (CVaR-RP) model allocates portfolio tail risk among assets evenly to mitigate downside risk. To enhance the CVaR-RP's predicting accuracy and adaptability to changing market conditions, we use a two-stage training approach within machine learning algorithms to forecast asset price movements. Portfolios are dynamically rebalanced based on these predictions to optimize the trade-off between risk mitigation and return maximization. Numerical analysis shows that the CVaR-RP strategy outperforms volatility-based risk parity and equal-weight strategies. Specifically, with machine learning-driven predictions and dynamic weight adjustments, the CVaR-RP achieves a higher Sharpe ratio, reduced maximum drawdown, and improved Calmar ratio. This research highlights the effectiveness of integrating machine learning methods into CVaR-RP strategies in enhancing returns and mitigating downside risk.