New clustering method uses stochastic dominance for risk-based asset allocation

Stochastic Dominance (SD) theory has long been used to rank assets according to investor risk preferences, but traditional stock clustering methods rely on metrics like Euclidean distance that ignore dominance relationships. A new preprint by Hua Li, Xue Jia, Yilin Kang, and Wing-Keung Wong (arXiv:2605.24422) bridges this gap by incorporating SD test statistics directly into clustering algorithms. They build a 'Stochastic Dominance Coefficient Matrix' from first-, second-, and third-order SD tests, then modify K-means and Hierarchical Clustering to produce 12 distinct algorithm variants. To evaluate clustering quality, they introduce new validity indices: SD-SC and SD-DBI. The method is tested on the US NASDAQ and China's CSI 100 indices, showing robust performance across developed and emerging markets.

Beyond clustering, the authors apply their results to portfolio construction. By integrating SD-based clusters into the Single Index Model and Global Minimum Variance Portfolios (GMVP), they enable customized asset allocation that aligns with specific risk attitudes—risk-averse, risk-seeking, or risk-neutral. This approach outperforms traditional clustering in capturing risk dominance relationships, offering a more theoretically grounded tool for quantitative finance. The paper represents a meaningful step toward combining rigorous economic theory with machine learning for practical investment decision-making.