Abstract Risk-based portfolio allocation strategies using the so-called hierarchical risk clustering have recently attracted attention from both academics and practitioners, mainly because of their ability to construct well-diversified portfolios through machine learning algorithms without the need to invert the covariance matrix. However, despite this innovative approach, the existing literature remains inconclusive regarding the outperformance of this methodology compared to traditional risk-based strategies in empirical applications and, additionally, there is uncertainty regarding how sensitive this new approach is to covariance matrix estimation. This paper addresses these gaps by providing out-of-sample comparisons using real-world data sets. The findings suggest that hierarchical risk clustering strategies are sensitive to covariance matrix estimation. Furthermore, there is no evidence that hierarchical risk clustering techniques outperform established risk-based approaches in the datasets analysed, even when alternative covariance matrix estimators are used.
Carlos Trucíos (Sun,) studied this question.