Hierarchical Risk Parity (HRP)

HRP begins by computing a distance matrix from the asset correlation matrix, typically using the transformation d = √(1 - ρ). A linkage method (single, average, or Ward) is then applied to build a hierarchical tree (dendrogram) that groups assets by similarity. This step identifies the natural nested structure in the data without imposing a rigid factor model, making it robust to the curse of dimensionality.

The rows and columns of the covariance matrix are reordered to match the leaf order of the dendrogram. This produces a quasi-diagonal matrix where large covariances cluster near the diagonal. This step organizes assets so that similar investments are adjacent, enabling the recursive bisection algorithm to allocate capital efficiently without needing to invert the full covariance matrix.

Starting from the top of the clustered tree, capital is split recursively at each node using inverse-variance weighting on the two sub-clusters. The allocation to each branch is proportional to the inverse of its cluster variance, ensuring that risk is distributed evenly across the hierarchy. This top-down approach avoids the concentration and instability issues of traditional mean-variance optimization.

Unlike Markowitz's critical line algorithm or standard risk parity, HRP never requires inverting the covariance matrix. This is a critical advantage when the number of assets N approaches or exceeds the number of observations T (the N > T problem). By avoiding inversion, HRP eliminates the primary source of numerical instability and estimation error that plagues traditional portfolio optimizers.

Empirical studies, including those by López de Prado (2016), demonstrate that HRP delivers lower out-of-sample variance and higher Sharpe ratios than inverse-variance weighting, equal risk contribution, and mean-variance portfolios. The hierarchical structure acts as a regularizer, preventing the optimizer from chasing spurious correlations and producing more stable weights across rebalancing periods.

The clustering step can incorporate alternative distance measures beyond the standard correlation-based metric. Practitioners can use tail-dependence coefficients, mutual information from information theory, or partial correlations to capture non-linear dependencies. This flexibility allows HRP to adapt to specific market regimes or asset classes where linear correlation is an insufficient measure of co-movement.

The mathematical backbone of all risk parity strategies. Euler's theorem for homogeneous functions allows total portfolio risk to be perfectly decomposed into additive contributions from each asset. For a portfolio volatility $\sigma(w)$, the marginal risk contribution (MRC) is the partial derivative $\partial \sigma / \partial w_i$, and the risk contribution (RC) is $w_i \times MRC_i$. The sum of all RCs exactly equals total portfolio volatility, enabling the precise balancing that HRP seeks to achieve without inverting the covariance matrix.

A statistical remedy for the estimation error that plagues sample covariance matrices, especially in high-dimensional portfolios. Shrinkage combines the noisy sample matrix with a highly structured target, like constant correlation, using a shrinkage intensity parameter. This pulls extreme eigenvalues toward the center, dramatically improving out-of-sample performance. While HRP avoids direct inversion, robust covariance estimation remains critical for the quasi-diagonalization and recursive bisection steps that define the algorithm.

An evolution of asset-based parity that balances risk across underlying economic drivers rather than individual securities. Instead of equalizing the risk contribution of stocks and bonds, factor parity targets uncorrelated factors like growth, inflation, and volatility. This addresses a key limitation of naive risk parity: assets that cluster around the same factor can create hidden concentration risk. HRP's clustering step implicitly addresses this by grouping assets with similar return drivers before allocation.

A dimensionality reduction approach that allocates risk equally across the uncorrelated principal components of asset returns rather than the correlated assets themselves. By diagonalizing the covariance matrix, PCA identifies truly independent sources of risk. HRP achieves a similar goal through hierarchical clustering on the correlation matrix, but PCA parity uses a mathematically orthogonal basis. Both methods solve the problem of spurious diversification where seemingly different assets share common risk drivers.

A diagnostic metric that quantifies true diversification by calculating the exponential of the entropy of risk contributions. An equally weighted portfolio of 10 assets might have an ENB of only 3 if most risk concentrates in a few positions. HRP explicitly maximizes this measure by ensuring capital is distributed across genuinely distinct clusters. ENB provides a single, interpretable number to compare the diversification quality of HRP against naive risk parity or mean-variance optimization.

A time-series model that captures how correlations between assets evolve through market regimes. Unlike static covariance, DCC estimates allow risk parity weights to adapt as relationships shift during crises or bull markets. Integrating DCC with HRP creates a time-varying hierarchical structure where the clustering dendrogram itself updates as correlations change, preventing the portfolio from relying on stale diversification assumptions during periods of market stress.