Nested Clustered Optimization

Feed a standard mean-variance optimizer a covariance matrix estimated from noisy historical data and it will happily bet huge, offsetting long-short positions on the two assets whose estimated correlation happens to look most favorable, purely because a matrix inversion is extremely sensitive to small errors in near-collinear inputs. The optimizer is not wrong given its inputs; the inputs are the problem. Nested clustered optimization (NCO) restructures the problem so a bad correlation estimate between two similar assets can no longer wreck the whole portfolio.

The analogy before any symbols

Imagine allocating a company's budget by asking every one of two hundred employees, individually, to negotiate directly with each other over who gets what, all at once. Small misunderstandings between any two people can cascade unpredictably through the whole negotiation. A functioning company instead groups people into departments, lets each department sort out its own internal split, and then has department heads negotiate the budget split across departments. Errors inside one department's internal negotiation stay contained there; they do not corrupt the cross-department allocation. NCO applies exactly that two-level structure to a covariance matrix instead of a budget.

The mechanics

NCO runs in three steps. First, cluster. Convert the correlation matrix $\rho_{ij}$ into a distance, commonly $d_{ij} = \sqrt{\tfrac12(1-\rho_{ij})}$, so highly correlated assets sit close together, and run hierarchical or k-means clustering to partition $N$ assets into $K$ clusters of similar assets. Second, optimize within each cluster. Inside each cluster, solve the usual minimum-variance (or mean-variance) problem using only the assets in that cluster, producing intra-cluster weights. Third, optimize across clusters. Treat each cluster as if it were a single synthetic asset, with return and variance built from its intra-cluster weights, and solve a small $K$-asset optimization across clusters. The final weight on any individual asset is:

$w_i = w_i^{\text{(intra-cluster)}} \times w_{c(i)}^{\text{(inter-cluster)}},$

where $c(i)$ is the cluster asset $i$ belongs to. In words: how much weight an asset gets inside its own cluster, times how much weight that whole cluster gets in the overall portfolio.

The key benefit: the matrix that has to be inverted at each step is small, either the handful of assets inside one cluster, or the handful of clusters themselves, not the full $N \times N$ matrix at once. Inversion error scales badly with matrix size when the matrix is close to singular (which correlated financial assets usually make it), so two small, better-conditioned inversions beat one large, ill-conditioned one.

This scatter is what NCO's first step is built to exploit: assets with high $\rho$ cluster tightly together in return space, and clustering groups exactly those tight scatters before any optimization touches them, rather than letting the optimizer discover (and over-trust) that relationship on its own.

Worked example: twelve assets, three clusters

Twelve equity names split into three clusters of four by sector similarity: tech, energy, and staples. A naive full-matrix minimum-variance optimization on all twelve, with a covariance matrix estimated from only 100 days of data, produces a handful of weights above 40% and several negative weights below −20%, driven by spurious cross-sector correlation estimates. Under NCO: within tech, the four names split roughly 30/20/35/15%; within energy, 40/25/20/15%; within staples, 20/30/35/15%. The cross-cluster step then allocates, say, 45% tech, 20% energy, 35% staples based on each cluster's aggregate risk. No individual final weight exceeds 45% × 35% ≈ 16%, and nothing goes negative, because neither optimization step ever saw all twelve assets' noisy pairwise correlations simultaneously.

Worked example: out-of-sample stability

Backtest both methods on 250 days in-sample, then measure realized volatility on the next 250 days out-of-sample. The naive full-matrix optimizer's in-sample "optimal" variance might be 8% annualized, but realized out-of-sample variance often lands near 14 to 16%, because the extreme long-short weights were fit to noise that does not repeat. NCO's in-sample variance is typically higher, closer to 10 to 11% (it never chased the noise as hard), but its out-of-sample realized variance tends to land close to that same 10 to 11%, since the clustering step removed the main channel through which noise could distort the weights. The comparison de Prado popularized shows NCO giving up a little in-sample "optimality" for a lot of out-of-sample reliability.

What this means in practice

NCO is a practical middle ground between full mean-variance optimization, which is fast and elegant but fragile with noisy inputs, and equal-weighting, which is robust but ignores real risk differences entirely. It is closely related to Hierarchical Risk Parity (López de Prado), which uses a similar clustering idea but skips the cross-cluster optimization step in favor of a simpler risk-based split. Both are common defaults exactly when $N$ is large relative to the amount of clean history available, the regime where a full covariance matrix is least trustworthy.