Hierarchical risk parity variants for cryptocurrency portfolios: denoising, detoning, tail dependence, embeddings and statistical inference, 2020–2026

We extend the 2025 conference study on Hierarchical Risk Parity (HRP) by addressing its two stated methodological follow-ups (Marchenko-Pastur denoising and spectral detoning) and by evaluating abroader cross-section of portfolio construction choices on a survivorship-bias-corrected, point-in-time Binance universe (146 ever-included symbols, 2020–2026). The comparison includes 27 constructed strategies spanning HRP-family variants, risk-based allocators, nested-clustering optimisers, momentum rules, and signal-aware extensions. Across 76 monthly rebalances, the HRP-family variants remain tightly clustered in risk-adjusted performance (annualized Sharpe 0.69 - 0.75), and this result persists under a 547- cell hyperparameter sweep (combined Hansen Superior Predictive Ability (SPA) pconsistent = 0.91). Strategies that alter the allocation step produce larger dispersion (notably the Minimum Variance Portfolio (MVP) , the Correlation-Regularised Iterative Shrinkage Portfolio (CRISP), the Nested Clustered Optimization with CRISP allocation (NCO–CRISP), and Nested Clustered Optimization (NCO)), with CRISP and NCO–CRISP significant in pairwise Ledoit–Wolf tests versus base line HRP in both full and post-COVID windows. However, after accounting for multiple comparisons, the headline SPA does not reject (pconsistent = 0.51), and the Deflated Sharpe Ratio (DSR) is directionally consistent with that conclusion. Post-COVID results indicate strong regime sensitivity: diversified HRPvariants lose 51–64% of capital while Bitcoin (BTC) and BTC-concentrating allocators hold up better. We also revise a nearlier interpretation on cluster stability: over the full panel, Pearson-based clustering is more stable than lower- tail dependence at monthly frequency. Overall, the evidence supports a cautious interpretation: within this universe, horizon, and search space, allocation choices are more strongly associated with performance differences than clustering perturbations, but those edges remain statistically fragile under search-adjusted inference.