Hierarchical Risk Parity (HRP)

• Hierarchical Risk Parity (HRP) is a portfolio construction method that leverages tree-based clustering and recursive risk splitting to avoid full covariance matrix inversion.
• It mitigates instability and overconcentration by bypassing noisy covariance estimates, providing more robust risk management than classical techniques.
• HRP’s three-stage process—clustering, quasi-diagonalization, and recursive bisection—facilitates robust out-of-sample diversification in various market conditions.

Searching arXiv for recent and foundational papers on Hierarchical Risk Parity and related hierarchical portfolio methods. Hierarchical Risk Parity (HRP) is a portfolio-allocation methodology that organizes assets through a hierarchy of dependence and then allocates capital by recursive risk splitting rather than by direct inversion of the full covariance matrix. In the literature summarized here, HRP is consistently presented as a response to instability, concentration, and poor out-of-sample behavior in classical mean-variance optimization, especially when covariance estimates are noisy or ill-conditioned. Its canonical workflow is the three-stage sequence introduced by López de Prado—tree clustering, quasi-diagonalization, and recursive bisection—while more recent work recasts HRP as a particular information-restricted or minimum-variance-only point within broader hierarchical and Schur-complement frameworks (Sen et al., 2022).

1. Definition and conceptual role

HRP is a risk-based portfolio construction method that uses the dependence geometry of asset returns to build a hierarchical ordering and then allocates capital across that hierarchy. Across the cited studies, its defining practical characteristics are that it does not require expected-return forecasts and avoids direct covariance-matrix inversion, which is repeatedly contrasted with the sensitivity of Markowitz-style optimization to estimation error, concentration, and instability (Sen et al., 2023).

The basic intuition is that diversification should be organized around clusters of correlated assets rather than around a single global optimizer. In applications to Indian equities, Latin American regional equities, and high-dimensional synthetic portfolios, HRP is described as producing more diversified portfolios, being more robust to estimation errors in the covariance matrix, and often exhibiting stronger out-of-sample stability than classical comparators such as CLA, MVP, or maximum-Sharpe portfolios, although not necessarily higher raw return in every setting (Sen et al., 2022).

Several papers also emphasize that HRP is not naturally a general mean-variance method. One recent formulation states this explicitly: standard HRP is “signal-blind” and “minimum-variance-only” because it does not accept an arbitrary expected-return signal $\mu$ and is pinned to the covariance-only setting (Wuebben, 26 Apr 2026). A mathematically related interpretation places HRP within “Heuristic Portfolio Optimization,” where it is treated as an information-restricted map from $(\boldsymbol{\mu}_e,\mathbf{\Sigma})$ to the simplex, using only a tree and recursively aggregated cluster variances rather than the full tangency solution (Alonso, 10 Jun 2026). This suggests that HRP is best understood as a structured regularization of portfolio choice rather than as a fully general optimizer.

2. Canonical algorithmic structure

The canonical HRP pipeline consists of three steps: hierarchical clustering, quasi-diagonalization, and recursive bisection. This three-stage description appears consistently across empirical and theoretical papers, including studies on machine-learning strategy selection, Indian sector portfolios, and broad expositions of HRP’s mathematical form (Kisiel et al., 2021).

The first step begins from a covariance or correlation representation of returns. A standard formulation converts correlations into distances via

$d_{ij}=\sqrt{\frac{1-\rho_{ij}}{2}},$

which appears explicitly in several papers, including studies of the Meta Portfolio Method, theoretical HPO analysis, and the NUAM-market application (Kisiel et al., 2021). One implementation then computes a second distance matrix,

$$\tilde d_{i,j}=\sqrt{\sum_{n=1}^{N}(d_{n,i}-d_{n,j})^{2}},$$

interpreted as representing the correlation of each asset with the rest of the portfolio, and uses $\tilde d_{ij}$ to build the hierarchical tree (Kisiel et al., 2021). Other papers specify agglomerative clustering with particular linkage rules, but the reported choices differ: one text states “single linkage clustering,” while another explicitly uses Ward linkage, and at least one study notes an internal inconsistency between text and figure captions on this point (Sen et al., 2023).

The second step, quasi-diagonalization or seriation, reorders the covariance or correlation matrix so that similar assets are adjacent and the matrix is approximately block diagonal. In matrix form this is often written as

$\Sigma^{(q)} = P \Sigma P^\top$

or

$C^* = P C P^\top,$

where $P$ is the permutation implied by the dendrogram order (Sen et al., 2022). The procedural aim is the same across papers: place larger covariances near the diagonal and separate more dissimilar assets.

The third step, recursive bisection, allocates capital top-down across sibling clusters. A standard cluster-internal inverse-variance portfolio is

$$\tilde w_{K,i} = \frac{\sigma_i^{-2}}{\sum_{j\in K}\sigma_j^{-2}},$$

with cluster variance

$$\widetilde V_K=\tilde{\mathbf w}_K^\top \mathbf{\Sigma}_K \tilde{\mathbf w}_K.$$

At a split $(\boldsymbol{\mu}_e,\mathbf{\Sigma})$0, HRP assigns

$(\boldsymbol{\mu}_e,\mathbf{\Sigma})$1

and multiplies all running weights in $(\boldsymbol{\mu}_e,\mathbf{\Sigma})$2 and $(\boldsymbol{\mu}_e,\mathbf{\Sigma})$3 by these fractions before recursing further (Alonso, 10 Jun 2026). In expository and empirical papers, this is often summarized more simply as allocating more weight to the less risky cluster (Ramirez-Carrillo et al., 3 Sep 2025). Some application papers do not print the full recursive equations line by line; where that occurs, only the top-down inverse-cluster-variance logic is directly supported (Kisiel et al., 2021).

3. Mathematical interpretation and links to Markowitz

A major strand of recent work argues that HRP is best understood through its relationship to minimum-variance optimization. One paper presents the exact generic top-down recursion

$(\boldsymbol{\mu}_e,\mathbf{\Sigma})$4

after partitioning a seriated covariance matrix as

$(\boldsymbol{\mu}_e,\mathbf{\Sigma})$5

and then shows that standard hierarchical allocation corresponds to a recursion that largely discards the off-block covariance $(\boldsymbol{\mu}_e,\mathbf{\Sigma})$6 (Cotton, 2024). In this view, HRP is a divide-and-conquer approximation to global minimum variance.

The same paper derives a Schur-complement recursion in which the exact global minimum-variance portfolio

$(\boldsymbol{\mu}_e,\mathbf{\Sigma})$7

can itself be written hierarchically using

$(\boldsymbol{\mu}_e,\mathbf{\Sigma})$8

together with modified budget vectors. This yields a continuum indexed by $(\boldsymbol{\mu}_e,\mathbf{\Sigma})$9, where $d_{ij}=\sqrt{\frac{1-\rho_{ij}}{2}},$0 recovers the “usual hierarchical family of methods” and $d_{ij}=\sqrt{\frac{1-\rho_{ij}}{2}},$1 recovers exact minimum variance (Cotton, 2024). A plausible implication is that classical HRP can be interpreted as the zero-Schur-correction endpoint of a broader hierarchical minimum-variance family.

A more abstract treatment places HRP inside a general “Heuristic Portfolio Optimization” theory. There the tangency portfolio is

$d_{ij}=\sqrt{\frac{1-\rho_{ij}}{2}},$2

and the central optimality condition is

$d_{ij}=\sqrt{\frac{1-\rho_{ij}}{2}},$3

For HRP, this means it is exactly tangency-optimal only for the implied-return ray

$d_{ij}=\sqrt{\frac{1-\rho_{ij}}{2}},$4

which the paper treats as a precise characterization of when a heuristic coincides with the optimizer (Alonso, 10 Jun 2026).

The same HPO analysis identifies the specific substitutions behind HRP’s heuristic nature. Relative to exact Schur-recursive minimum variance, HRP replaces the Schur complement $d_{ij}=\sqrt{\frac{1-\rho_{ij}}{2}},$5 by the raw block $d_{ij}=\sqrt{\frac{1-\rho_{ij}}{2}},$6, replaces the tilted budget $d_{ij}=\sqrt{\frac{1-\rho_{ij}}{2}},$7 by $d_{ij}=\sqrt{\frac{1-\rho_{ij}}{2}},$8, and replaces within-cluster minimum-variance weights by inverse-variance weights (Alonso, 10 Jun 2026). This gives a theorem-level explanation for a widely noted empirical property: HRP gains robustness by deleting information that is often unstable to estimate.

4. Empirical behavior across markets and regimes

Empirical studies do not support the claim that HRP always dominates, but they do support a recurring pattern: HRP is usually more defensive, more diversified, and more stable out of sample, while more aggressive alternatives can produce higher cumulative return in favorable regimes. In the Meta Portfolio Method study, HRP is explicitly described as producing “a portfolio with a much lower variance” and “a return stream with heavily damped tails,” while Naïve Risk Parity is said to have “heavy tails” and to be “considerably more profitable than HRP during certain market conditions” (Kisiel et al., 2021).

That same study reports out-of-sample HRP Sharpe ratios across ten ETF universes of $d_{ij}=\sqrt{\frac{1-\rho_{ij}}{2}},$9 and $$\tilde d_{i,j}=\sqrt{\sum_{n=1}^{N}(d_{n,i}-d_{n,j})^{2}},$$0, with cumulative returns of $$\tilde d_{i,j}=\sqrt{\sum_{n=1}^{N}(d_{n,i}-d_{n,j})^{2}},$$1 and $$\tilde d_{i,j}=\sqrt{\sum_{n=1}^{N}(d_{n,i}-d_{n,j})^{2}},$$2 (Kisiel et al., 2021). The paper’s own interpretation is that HRP “stabilizes while NRP accelerates,” and that HRP is particularly valuable around stress episodes such as the onset of the COVID-19 pandemic in 2020 (Kisiel et al., 2021).

Indian-market evidence is mixed but broadly supportive of HRP’s out-of-sample role. In the NIFTY 50 and sector study against CLA, HRP wins on test-set Sharpe in 7 of 8 universes—Auto, Banking, FMCG, IT, Metal, Realty, and NIFTY 50—while CLA wins only in Pharma (Sen et al., 2022). In another Indian study comparing HRP with Eigen portfolios, HRP delivers higher test Sharpe in five of seven universes—consumer durable, healthcare, information technology, oil and gas, and NIFTY 50—while Eigen wins in auto and financial services (Sen et al., 2022). A separate comparison with MVP and reinforcement learning on Indian sectors presents less clean evidence because some IT and Finance/Financial tables are internally inconsistent, but still treats HRP as a credible diversification-oriented benchmark that can outperform MVP in some sectors such as Auto and FMCG (Sen et al., 2023).

In the Latin American NUAM market, HRP does not beat the maximum-Sharpe benchmark on return or Sharpe, but it records lower volatility than $$\tilde d_{i,j}=\sqrt{\sum_{n=1}^{N}(d_{n,i}-d_{n,j})^{2}},$$3, lower drawdown than both benchmarks, and much lower tracking error than the maximum-Sharpe portfolio. The paper reports annual return, volatility, and Sharpe of $$\tilde d_{i,j}=\sqrt{\sum_{n=1}^{N}(d_{n,i}-d_{n,j})^{2}},$$4, $$\tilde d_{i,j}=\sqrt{\sum_{n=1}^{N}(d_{n,i}-d_{n,j})^{2}},$$5, and $$\tilde d_{i,j}=\sqrt{\sum_{n=1}^{N}(d_{n,i}-d_{n,j})^{2}},$$6 for HRP, compared with $$\tilde d_{i,j}=\sqrt{\sum_{n=1}^{N}(d_{n,i}-d_{n,j})^{2}},$$7, $$\tilde d_{i,j}=\sqrt{\sum_{n=1}^{N}(d_{n,i}-d_{n,j})^{2}},$$8, and $$\tilde d_{i,j}=\sqrt{\sum_{n=1}^{N}(d_{n,i}-d_{n,j})^{2}},$$9 for Max Sharpe and $\tilde d_{ij}$0, $\tilde d_{ij}$1, and $\tilde d_{ij}$2 for $\tilde d_{ij}$3 (Ramirez-Carrillo et al., 3 Sep 2025). The substantive conclusion is explicitly conditional: HRP is attractive less as a return maximizer than as a resilient framework for risk control in volatile integrated emerging markets.

5. HRP in high-dimensional estimation and machine-learning systems

HRP appears in the recent literature not only as a standalone allocator but also as a component inside broader estimation and learning systems. In high-dimensional covariance-estimation studies, HRP is repeatedly the least sensitive portfolio rule to the choice of covariance cleaner. One such paper compares unconstrained MVP, long-only MVP, and HRP across hierarchical nested, one-factor, and diagonal covariance structures and reports that “HRP is the least sensitive to high-dimensional noise,” while covariance cleaning has only minimal impact on its performance (García-Medina, 2024).

That study implements HRP using the standard pipeline

$\tilde d_{ij}$4

followed by Single Linkage clustering and recursive inverse-variance logic (García-Medina, 2024). Its simulation and walk-forward results show that HRP consistently improves diversification relative to MVP-type strategies, but sophisticated covariance cleaners improve MVP more than they improve HRP. This suggests that one source of HRP’s practical appeal is precisely that its performance depends less on very fine covariance estimation.

HRP is also used as a component strategy in adaptive meta-allocation. In the Meta Portfolio Method, a GARCH-DCC covariance estimate is used both to build HRP and NRP and to generate features for an XGBoost model that predicts the future Sharpe-ratio spread

$\tilde d_{ij}$5

The allocation rule is binary:

• use HRP if $\tilde d_{ij}$6,
• use NRP otherwise (Kisiel et al., 2021).

In that framework, HRP functions as the defensive leg. Feature-importance analysis highlights hrp_down_dev, intra_cluster_var, and cophenetic_average, with the latter two interpreted as quantifying “the level and the complexity of the hierarchical structure in a given investment universe” (Kisiel et al., 2021). This suggests that HRP is most relevant when the cross-sectional dependence structure is meaningfully hierarchical and when downside-risk conditions make its tail-dampening properties valuable.

A distinct extension is RL-BHRP, which is HRP-inspired but not canonical HRP. There the hierarchy is fixed by sectors rather than learned from clustering, portfolio weights factor as

$\tilde d_{ij}$7

and reinforcement learning adapts sector and within-sector weights while penalizing deviations from hierarchical equal-risk contributions (Kang et al., 16 Aug 2025). The paper explicitly frames its contribution as “Bayesian Hierarchical Risk Parity” rather than classical López de Prado HRP, but it preserves the core principle that risk should be diversified across levels of a hierarchy (Kang et al., 16 Aug 2025).

6. Extensions, generalizations, and competing hierarchical frameworks

Recent theoretical work extends HRP in two main directions: signal-aware hierarchical allocation and exact hierarchical minimum-variance recursion. One 2026 paper introduces HRP-$\tilde d_{ij}$8 and HRP-$\tilde d_{ij}$9, two tree-based methods that accept arbitrary expected-return signals $\Sigma^{(q)} = P \Sigma P^\top$0. HRP-$\Sigma^{(q)} = P \Sigma P^\top$1 injects the signal through signed inverse-variance representatives, while HRP-$\Sigma^{(q)} = P \Sigma P^\top$2 replaces those representatives with recursive local mean-variance optima and retains $\Sigma^{(q)} = P \Sigma P^\top$3 complexity (Wuebben, 26 Apr 2026). The same paper states that HRP-$\Sigma^{(q)} = P \Sigma P^\top$4 and HRP-$\Sigma^{(q)} = P \Sigma P^\top$5 both outperform plain HRP not only for heterogeneous expected-return forecasts but also on the minimum-variance problem $\Sigma^{(q)} = P \Sigma P^\top$6, and that HRP-$\Sigma^{(q)} = P \Sigma P^\top$7 consistently improves on HRP-$\Sigma^{(q)} = P \Sigma P^\top$8 (Wuebben, 26 Apr 2026).

The same work introduces CRISP, a non-hierarchical iterative solver for

$\Sigma^{(q)} = P \Sigma P^\top$9

which interpolates between a diagonal rule and full Markowitz and is reported to dominate HRP, Cotton, Ledoit-Wolf shrinkage, direct Markowitz, and the new signal-aware hierarchical methods at intermediate $C^* = P C P^\top,$0 values (Wuebben, 26 Apr 2026). The paper’s conclusion is sharp: once alpha signals are available, plain HRP is no longer the relevant baseline because it discards the most informative input.

A complementary mathematical treatment develops return-adjusted HRP, or RA-HRP, within the HPO framework. It defines local cluster-Sharpe scores

$C^* = P C P^\top,$1

and then branch probabilities

$C^* = P C P^\top,$2

together with a unit-free interpolation

$C^* = P C P^\top,$3

between HRP and RA-HRP (Alonso, 10 Jun 2026). That paper shows that the marginal value of adding return information to HRP can be decomposed node by node through first-order Sharpe calculus, with nodewise alphas measuring the gain from deviating from HRP’s risk-only splits (Alonso, 10 Jun 2026).

Other alternatives retain the broad hierarchical spirit while departing sharply from classical HRP. Hierarchical Minimum Variance Portfolios assumes the covariance matrix itself has a recursive hierarchical graph structure and solves

$C^* = P C P^\top,$4

exactly through Schur-complement reductions rather than clustering heuristics (Mograby, 16 Mar 2025). Topological Risk Parity, by contrast, is designed for long/short, market-neutral, factor-aware portfolios; it uses a rooted sparse topology rather than a binary dendrogram and propagates signed signals through coefficients

$C^* = P C P^\top,$5

instead of performing variance-based recursive bisection (Nayar et al., 18 Apr 2026). These frameworks do not replace HRP in the narrow sense, but they clarify its scope by showing which of its properties depend on long-only recursive risk splitting and which belong more generally to tree-based portfolio design.

7. Limitations, controversies, and enduring significance

The modern HRP literature is explicit about several limitations. First, many empirical papers provide incomplete algorithmic specifications. Some do not report the clustering linkage, others omit the exact recursive-bisection formulas, rebalancing frequency, or window lengths, and some report internal inconsistencies in tables or figure labels (Sen et al., 2023). Second, results are often sample- and market-specific: several studies rely on a single train/test split, relatively short out-of-sample windows, and limited discussion of transaction costs, slippage, or turnover (Sen et al., 2022).

Third, HRP is not a universal performance maximizer. The evidence summarized above repeatedly shows that more aggressive strategies can earn higher cumulative return or higher Sharpe in benign regimes, and that HRP’s main comparative strength is defensive robustness rather than unconditional dominance (Kisiel et al., 2021). The NUAM study makes this especially clear by showing better drawdown and tracking-error properties for HRP alongside inferior return-based metrics relative to a maximum-Sharpe portfolio (Ramirez-Carrillo et al., 3 Sep 2025).

Fourth, newer theory challenges the view of HRP as a fully satisfactory endpoint. The HPO analysis argues that exact optimality of HRP is generically false except on a restricted implied-return ray, and the Schur-complement analysis shows that HRP ignores precisely the cross-cluster hedge terms that the exact optimizer would retain (Alonso, 10 Jun 2026). Signal-aware generalizations then go further and argue that once expected-return information is available, a method that cannot use $C^* = P C P^\top,$6 is structurally incomplete (Wuebben, 26 Apr 2026).

Yet HRP remains significant because it occupies a distinctive point in the design space of portfolio rules. It is tree-based, inversion-avoiding, interpretable, long-only in standard form, and comparatively robust to covariance-estimation noise. In empirical work it often generalizes well out of sample; in theoretical work it has become a reference object for studying information restriction, hierarchical regularization, and the trade-off between robustness and optimality. Across both application papers and recent mathematical treatments, the enduring message is not that HRP is universally optimal, but that it is a principled and highly influential way to regularize portfolio allocation through dependence structure rather than through full global optimization (Cotton, 2024).