Hierarchical risk parity, or HRP, is a portfolio construction method proposed by Marcos Lopez de Prado in 2016. It groups similar assets using clustering and allocates risk down the resulting tree, which avoids inverting the covariance matrix and tends to give more stable weights than classic optimization.
The three steps
- Cluster. Group assets by how similar their returns are, using a distance based on correlation, producing a tree.
- Quasi-diagonalise. Reorder assets so that similar ones sit next to each other in the covariance matrix.
- Recursive bisection. Split the ordered list in two, give each half a share of capital in inverse proportion to its risk, then repeat inside each half.
An illustrative example
Take four assets: two large banks and two software companies. Clustering pairs the banks together and the software companies together. At the top split, if the bank group has twice the risk of the software group, it receives one-third of the capital and software two-thirds. Inside each group, the two stocks are split by inverse variance. The numbers are made up.
Why it is used
- It does not need expected returns or an inverted covariance matrix, which avoids the instability of mean-variance optimization when assets are highly correlated.
- Weights are diversified across clusters, and never extreme.
- In Lopez de Prado's simulations, it had lower out-of-sample variance than the quadratic optimizers tested.
Limits
- It is a heuristic with no guarantee of optimality for a given objective.
- Results depend on the clustering method and distance used.
- It ignores expected returns entirely, so it is a risk-allocation method, not a return-seeking one.
- Correlations still need estimating. See covariance matrix.
See also risk parity and portfolio optimization.
Frequently asked questions
What is hierarchical risk parity?
It is a portfolio method that clusters assets by similarity and splits risk down the tree, giving diversified weights without inverting the covariance matrix.
How is HRP different from risk parity?
Classic risk parity equalises each asset's risk contribution using the covariance matrix. HRP uses clustering and inverse-variance splitting down a hierarchy.
Does HRP use expected returns?
No. It only uses the risk structure of the assets.
Why can HRP be more stable than mean-variance optimization?
Because it avoids matrix inversion, which amplifies estimation errors when assets are highly correlated.