Minimum Correlation Portfolio Construction Using Hierarchical Clustering

Minimum Correlation Portfolio Construction Using Hierarchical Clustering is a heuristic portfolio optimization technique that first groups assets into clusters based on their pairwise correlation structure-typically via hierarchical clustering-and then allocates risk weights across and within clusters to minimize effective correlation and improve diversification. The method decouples asset grouping from risk allocation: clustering identifies latent risk factors through correlation topology, while risk allocation (e.g., inverse-variance, risk parity, or minimum-variance within clusters) ensures no single cluster or asset dominates portfolio risk. This approach is designed to mitigate the estimation error and instability of traditional mean-variance optimization, especially in high-dimensional or noisy correlation matrices.

The technique leverages graph-theoretic and tree-based structures to approximate the underlying risk manifold. By constructing a dendrogram from a distance metric derived from the correlation matrix (e.g., 1 − ρ), the algorithm recursively partitions assets into homogeneous risk groups, enabling more robust allocation than direct covariance inversion. Empirical studies show that portfolios built this way achieve lower out-of-sample volatility and higher Sharpe ratios than naive diversification or full covariance-based methods, particularly when correlation matrices are estimated from limited or non-stationary return data.

The construction process consists of four core steps: (1) compute a correlation-based distance matrix, (2) build a hierarchical clustering tree (dendrogram), (3) compute cluster-level risk contributions, and (4) allocate risk weights recursively using a bisection rule. Each step is deterministic and avoids numerical optimization of the full covariance matrix.

• Distance matrix construction: Transform the correlation matrix ρ into a metric distance matrix d where dᵢⱼ = √(2(1 − ρᵢⱼ)), ensuring triangle inequality and enabling meaningful clustering.
• Hierarchical clustering: Apply agglomerative clustering (e.g., Ward’s or single linkage) to form a binary tree (dendrogram) where leaves are assets and internal nodes represent clusters of increasing risk homogeneity.
• Risk contribution estimation: For each internal node, compute the local minimum-variance or risk-parity weights for its two child subclusters using only the submatrix of those assets-avoiding full-matrix inversion.
• Recursive weight allocation: Starting at the root, assign weights to left and right subclusters proportional to the inverse of their risk contributions (e.g., in HRP the left subcluster receives σ²_right / (σ²_left + σ²_right) and the right subcluster the remainder), then recurse down the tree until leaf-level weights are assigned.

This workflow ensures that assets with high within-cluster correlation receive lower aggregate weights, directly targeting correlation reduction. The method is non-parametric and does not assume a specific return distribution, and the canonical HRP algorithm produces long-only, fully invested weights.

Several variants of hierarchical clustering-based minimum correlation construction exist, differing primarily in clustering linkage criteria, risk allocation rules, and clustering input metrics:

• Hierarchical Risk Parity (HRP): Uses variance-based risk contributions and inverse-variance weighting within clusters; the canonical formulation introduced by López de Prado.
• Hierarchical Minimum Variance (HMV): Directly minimizes portfolio variance within each bisection using the Schur complement to preserve full covariance information while reducing computational complexity.
• Correlation-based Minimum Spanning Tree (MST) selection: Uses the MST as a clustering scaffold, then allocates risk inversely to edge-weighted centrality-often combined with peripheral asset exclusion for selection.
• Network Risk Parity: Extends clustering by incorporating edge centrality or community detection in correlation graphs to weight clusters by systemic risk exposure.

All variants share the same high-level objective: reduce effective correlation by decoupling asset grouping from risk allocation. The choice of variant depends on data quality, investment constraints (e.g., long-only), and computational resources. HMV, for instance, is preferred when precise covariance estimation is feasible, while HRP is more robust in low-data regimes.

In practice, minimum correlation portfolio construction using hierarchical clustering is employed primarily for risk-based asset allocation in multi-asset and multi-security portfolios, especially where traditional covariance-based optimization fails due to estimation error or ill-conditioning. It is commonly applied in:

• Multi-asset class allocation: Grouping asset classes (e.g., equities, bonds, commodities) by correlation before assigning risk budgets.
• Factor-based portfolio construction: Clustering factors (value, momentum, quality) by correlation to avoid overexposure to correlated risk drivers.
• Security selection within sectors: Within a sector, clustering stocks by return correlation and allocating risk inversely to within-cluster volatility.

Implementation typically follows a backtest-first workflow: historical returns are clustered periodically (e.g., monthly) using a rolling window, weights are computed recursively, and the resulting portfolio is rebalanced at fixed intervals. Out-of-sample performance is evaluated using Sharpe ratio, turnover, and drawdown metrics. Empirical studies report lower out-of-sample volatility compared to equal-weight or naive risk-parity benchmarks, especially during periods of high correlation regime shifts.

Suppose a portfolio contains four assets: A, B, C, D, with the following correlation matrix (ρ):

1. Distance matrix: dᵢⱼ = √(2(1 − ρᵢⱼ)) yields d(A,B) ≈ 0.632, d(C,D) ≈ 0.632, and cross-pair distances > 1.2.
2. Clustering: Agglomerative clustering merges A-B and C-D first, then joins the two clusters-forming a dendrogram with two top-level subclusters: {A,B} and {C,D}.
3. Within-cluster risk: Assume equal volatilities (σ = 10% = 0.1). For {A,B} with ρ_AB = 0.8 and equal weights (0.5, 0.5), portfolio variance = (0.5²·0.1² + 0.5²·0.1² + 2·0.5·0.5·0.8·0.1·0.1) = 0.0025 + 0.0025 + 0.004 = 0.009 = 0.9%. Similarly, for {C,D} with ρ_CD = 0.7, variance = 0.0025 + 0.0025 + 2·0.5·0.5·0.7·0.01 = 0.005 + 0.0035 = 0.0085 = 0.85%. Thus, σ₁ = √0.009 ≈ 0.0949 (≈9.49%), σ₂ = √0.0085 ≈ 0.0922 (≈9.22%).
4. Cluster-level risk: The covariance between cluster portfolios {A,B} and {C,D} is computed from the asset-level covariances: cov = Σᵢ∈{A,B} Σⱼ∈{C,D} wᵢ wⱼ cov(i,j) = 0.5·0.5·0.2·0.01 + 0.5·0.5·0.1·0.01 + 0.5·0.5·0.1·0.01 + 0.5·0.5·0.2·0.01 = 0.25·(0.002 + 0.001 + 0.001 + 0.002) = 0.25·0.006 = 0.0015 = 0.15%. The correlation between clusters is therefore ρ₁₂ = cov / (σ₁σ₂) = 0.0015 / (√0.009 · √0.0085) ≈ 0.0015 / (0.0949·0.0922) ≈ 0.0015 / 0.00875 ≈ 0.171. Total portfolio variance = w₁²·0.009 + w₂²·0.0085 + 2w₁w₂·0.0015.
5. Recursive allocation: Using an inverse-volatility split (HRP’s canonical inverse-variance rule would give w₁ = 0.0085/0.0175 ≈ 0.486), the condition is w₁σ₁ = w₂σ₂ ⇒ w₁/w₂ = σ₂/σ₁ ≈ 0.0922/0.0949 ≈ 0.972. Normalizing, w₁ = 0.972 / (1 + 0.972) ≈ 0.493, w₂ = 1 − w₁ ≈ 0.507. Final leaf weights: w_A = w_B = 0.493·0.5 ≈ 0.2465, w_C = w_D = 0.507·0.5 ≈ 0.2535.

Result: Weights are nearly equal across all assets, but derived from correlation topology—not naive diversification. In contrast, mean-variance optimization might concentrate in C or D due to lower marginal risk, increasing exposure to the C-D correlation cluster.

While hierarchical clustering improves robustness, it introduces specific limitations:

• Sensitivity to distance metric choice: Using 1 − ρ assumes linear correlation is sufficient; nonlinear dependencies (e.g., tail correlation) may be missed unless a different distance (e.g., rank correlation or mutual information) is used.
• Discrete clustering artifacts: Hierarchical trees are binary and discrete; small changes in correlation estimates can restructure the dendrogram, causing weight jumps at rebalancing.
• Loss of within-cluster heterogeneity: Assets in the same cluster are treated as homogeneous for allocation, potentially ignoring idiosyncratic risk differences.
• No explicit return modeling: The method is purely risk-based; it does not incorporate expected returns, making it unsuitable for active return targeting without hybridization.
• Computational overhead for large universes: Although more efficient than full covariance inversion, building the dendrogram scales as roughly O(n²) or worse, which can be costly for thousands of assets without approximation.

In practice, these limitations are mitigated by combining hierarchical clustering with regularization (e.g., shrinkage of correlation estimates), rolling estimation windows, and stress-testing under regime shifts. The method is best suited for risk budgeting and diversification enhancement-not for forecasting alpha or timing markets.

• Using raw correlation without distance transformation: Clustering on ρ directly violates metric axioms and produces unstable trees; always convert to a proper distance (e.g., √(2(1 − ρ))).
• Ignoring clustering stability: Re-running clustering with perturbed correlation matrices can reveal fragile groupings; unstable clusters should be merged or excluded.
• Overlooking estimation error in cluster risk: Using sample covariance within clusters without shrinkage or Bayesian regularization leads to overfitting, especially with short histories.
• Applying to non-stationary data: Correlation structures change over time; applying static clustering to long lookback windows dilutes the method’s responsiveness.
• Confusing with factor-based clustering: Hierarchical clustering on assets ≠ clustering on factor exposures; the former targets correlation topology, the latter targets latent factor risk drivers.

Best practice is to validate the dendrogram structure visually and compare out-of-sample performance against benchmarks (e.g., equal-weight, risk parity) before live implementation.