Skip to content

Building Diversified Portfolios that Outperform Out-of-Sample (HRP): summary and key takeaways

The paper proposes Hierarchical Risk Parity (HRP), a portfolio method that builds weights from a cluster tree instead of inverting the covariance matrix. In Monte Carlo tests, HRP produced lower out-of-sample variance than Markowitz's CLA minimum-variance portfolio and than an inverse-variance (risk parity) portfolio.

Prado, M.L.d. · 2016 · The Journal of Portfolio ManagementSummary updated 20 November 2026
Published2016Journal article
Asset focusAsset-agnosticPortfolio construction
TypePortfolio MgmtPortfolio Construction & Risk
Reading time4 minthis summary

Building Diversified Portfolios that Outperform Out-of-Sample (HRP) at a glance

The question
Can a tree-based allocation that avoids covariance inversion give more stable and less risky out-of-sample portfolios than quadratic optimization and traditional risk parity?
Dataset & period
Simulated data only, no real market series. Monte Carlo: 10 series of Gaussian returns with zero mean, 520 observations each (about 2 years of daily history), repeated 10,000 times. Portfolios use a 260-observation lookback and are rebalanced every 22 observations.
Method
HRP has three stages. First, it turns the correlation matrix into a distance matrix and clusters the assets into a tree. Second, it reorders the covariance matrix so similar assets sit next to each other. Third, it splits weight top-down between adjacent clusters in inverse proportion to their variance, using inverse-variance weights inside each cluster. It is compared with CLA's minimum-variance portfolio and the inverse-variance portfolio (IVP).

Key findings of Building Diversified Portfolios that Outperform Out-of-Sample (HRP)

  • In a worked example, CLA put 92.66% of the portfolio in its top five holdings, while HRP put 62.57% there.
  • In that same example, in-sample risk was similar for the two methods, and CLA gave zero weight to several assets.
  • Out of sample over 10,000 simulations, variance of portfolio returns was 0.1157 for CLA, 0.0928 for IVP and 0.0671 for HRP.
  • CLA's out-of-sample variance was 72.47% greater than HRP's, even though minimum variance is CLA's optimization goal.
  • IVP's out-of-sample variance was 38.24% greater than HRP's.
  • Mean out-of-sample returns for all three portfolios were essentially zero, so the difference comes from risk.

Robustness, as reported in the paper

  • The out-of-sample test uses rolling re-estimation and rebalancing across 10,000 random runs, not one chosen dataset.
  • Random shocks and a random correlation structure were added to the simulated data to mimic jumps and cross-asset dependence.
  • The author states, without showing results in the text, that HRP's edge grows with larger universes, more shocks, stronger correlation or rebalancing costs.

Caveats and limits

  • All evidence comes from simulated data; no real-asset backtest is reported.
  • The paper says the mathematical proof of HRP's outperformance is beyond its scope.
  • Transaction costs are not modelled in the reported results (inferred).
  • Results depend on the chosen distance measure, linkage rule and simulation settings, which the paper says can be varied (inferred as a sensitivity risk).

Items marked (inferred) are our reading of the design, not statements by the authors.

Why it matters

Covariance inversion becomes unstable when assets are highly correlated or history is short. HRP offers a way to build a diversified portfolio without that step. Practitioners can use the paper's code as a starting point for testing it on their own universe.

Questions about this paper

Does HRP need to invert the covariance matrix?

No. The paper says HRP can compute a portfolio even on an ill-degenerated or singular covariance matrix, which quadratic optimizers cannot do.

How did HRP compare with CLA out of sample?

HRP had the lowest out-of-sample variance at 0.0671, versus 0.1157 for CLA and 0.0928 for IVP in the Monte Carlo test.

What are the three stages of HRP?

Tree clustering, quasi-diagonalization of the covariance matrix, and recursive bisection that splits weights by inverse cluster variance.

Where can I read the original paper?

The original is at its publisher: https://doi.org/10.3905/jpm.2016.42.4.059. We link to the paper and do not host the PDF.

Follow new quant research

Download MacroFinance market data as CSV or Excel, or pull it by API, and test the ideas in these papers yourself. New summaries are added as we review them.

Create a free account
Summaries are written by MacroFinance from the published papers, in our own words, for information only. They are not investment advice and may omit detail. Papers are not hosted here: each page links to the original at its publisher.