Adaptive Seriational Risk Parity and Other Extensions for Heuristic Portfolio Construction Using Machine Learning and Graph Theory

In this article, the authors present a conceptual framework named adaptive seriational risk parity (ASRP) to extend hierarchical parity (HRP) as an asset allocation heuristic. The first step of HRP (quasi-diagonalization), determining hierarchy assets, is required for actual done in second (recursive bisectioning). original scheme, found using single-linkage clustering correlation matrix, which static tree-based method. compare performance standard with other and adaptive methods, well seriation-based methods that do not rely on trees. Seriation broader concept allowing reordering rows or columns matrix best express similarities between elements. Each discussed variation leads different time series reflecting portfolio 20-year backtest multi-asset futures universe. Unsupervised learningbased these time-series creates taxonomy groups strategies high correspondence construction various types ASRP. Performance analysis variations shows most alternatives outperform used risk-adjusted basis. Adaptive tree show mixed results, generic approaches underperform. Key Findings ▪ introduce decisions implement quasi-diagonalization single linkage. Tree-based are further separated versions. Altogether, 57 connected literature. Backtests HRP-type applied universe lead resulting return series. This can be visualized dendrogram clustering. reflected by similar step. Most seem underperform HRP, whereas results. presented fit into triple artificial intelligence approach connect synthetic data generation explainable machine learning. generates market applies article. third uses model-agnostic explanation such SHAP explain features model selection