Modern Portfolio Theory vs Hierarchical Risk Parity
You have a candidate list. Say forty tickers you actually researched, screened for liquidity, and would be happy to own. The remaining question is not what to buy but how much of each — and that question has two very different answers depending on which allocation method you trust.
On one side sits modern portfolio theory (MPT), the Markowitz mean-variance framework that has been the default since the 1950s: estimate expected returns and a covariance matrix, pick an objective, solve for the weights. On the other sits hierarchical risk parity (HRP), which skips the matrix inversion entirely, clusters stocks by correlation distance, and allocates risk down the resulting tree.
Both aim at the same outcome — a portfolio that is less volatile because its holdings do not all fall at once. They get there with different assumptions, different inputs, and very different behaviour when the input data is noisy. This article compares them side by side and shows how RSJ Portfolio exposes both as selectable allocation methods, so the comparison is a dropdown, not a research project.
The Allocation Decision: MPT or HRP?
The decision matters because of estimation error. Every optimizer consumes a covariance matrix estimated from a finite window of historic end-of-day prices. That estimate is wrong in ways you cannot see, and the two methods respond to that error differently.
- MPT inverts the covariance matrix. Inversion is sensitive: small changes in correlations can move the solution a long way, sometimes into lopsided weights concentrated in a handful of names.
- HRP never inverts the matrix. It converts correlations into distances, builds a hierarchy, and splits risk recursively between branches. Noisy pair correlations get averaged inside clusters instead of being amplified by an inverse.
Same candidate list, same price history, two different personalities. One tries to be optimal given the estimate; the other tries to stay sensible when the estimate is shaky.
What follows is a comparison, not a tutorial. If you want the statistical background on why weak correlation lowers volatility in the first place, we cover that separately in Why Low Correlation Stocks Reduce Portfolio Volatility.
The Landscape: How MPT and HRP Approach Allocation
Modern portfolio theory: choose an objective, solve for weights
MPT treats allocation as an optimization problem. You supply a risk model (the covariance estimate), often a return model, and an objective. The solver then finds the weight vector that best satisfies that objective subject to your constraints and the budget condition that weights sum to one.
The unconstrained core of it is almost embarrassingly short:
# The heart of mean-variance optimization — and the reason estimation
# error matters so much. A small change in `cov` moves `w` a long way.
import numpy as np
w = np.linalg.solve(cov, mu) # tangency-style solution, unconstrained
w = w / w.sum() # scale to fully invested
That one solve call is a matrix inversion in disguise. When the covariance matrix is estimated from a few hundred overlapping daily observations of forty correlated stocks, it is close to singular, and the inverse amplifies exactly the noise you wanted to average out.
Hierarchical risk parity: cluster first, allocate second
HRP, introduced in 2016, was designed as an answer to that fragility. It takes the correlation matrix, converts each pairwise correlation into a distance, builds a hierarchical clustering tree, reorders the assets so similar ones sit next to each other, and then splits capital and risk recursively down the tree — each split allocating inversely to the variance of the two branches.
from scipy.cluster.hierarchy import linkage
from scipy.spatial.distance import squareform
dist = np.sqrt(0.5 * (1 - corr)) # correlation distance
tree = linkage(squareform(dist, checks=False), method="single")
For reference, that is the step HRP adds before any money is allocated. No inversion appears anywhere in the pipeline.
Where they actually diverge
- Inputs. MPT wants a covariance matrix and, for return-seeking objectives, expected returns. HRP wants a correlation matrix plus a linkage choice, and no return forecasts at all.
- Error handling. MPT propagates estimation error through an inverse. HRP dilutes it through cluster averages and recursive bisection.
- Output shape. MPT can produce concentrated weights when correlations are unstable. HRP typically spreads risk more evenly across the correlation structure.
- Cost of being wrong. A bad return forecast in MPT hurts directly. In HRP there is no return forecast to be wrong about.
Both still rely on the same underlying idea: allocate toward stocks whose returns — or whose downside — correlate weakly with the rest. The difference is how aggressively each method trusts the raw pairwise numbers.
Selection Criteria: Which Method Fits Your Situation?
Ease of use
MPT asks more of you: pick a risk model, often a return model, and an objective. HRP asks less on the return side but still needs a correlation window and a linkage method — the choices exist, they are just fewer. Neither is hard once the inputs are on screen; both are painful if you are assembling the matrices by hand.
Automation
This is where the tooling question usually gets decided. In RSJ Portfolio both approaches are implemented as allocation methods you select in the interface — you pick the method and its parameters, and the optimizer runs. There is no optimizer code to maintain and no matrix assembly step to get wrong.
Robustness
If your correlations are stable and your return views are well-founded, MPT’s precision is an asset. If your correlations are noisy — thin trading, regime shifts, a window that straddles a crisis — HRP’s clustering tends to hold up better out of sample. A useful rule of thumb: the more you distrust your inputs, the more HRP’s conservatism is worth paying for.
Integration
Both methods consume the same data in RSJ Portfolio: historic end-of-day prices that go through the same cleaning pipeline and land in the same local cache. Multi-currency candidate lists follow the same conversion path regardless of which method you pick — see Portfolio Optimization for Multi-Currency Portfolios: A Guide for that workflow.
Price
RSJ Portfolio is a Windows desktop application sold under a single-user license; the current terms are on the pricing page. There is no metered API and no per-optimization charge, so comparing MPT and HRP on the same candidate list costs you time, not credits.
Side-by-Side Comparison: MPT vs HRP
| Modern portfolio theory (MPT) | Hierarchical risk parity (HRP) | |
|---|---|---|
| Objective | Best risk-return trade-off for a chosen target: minimum volatility, maximum Sharpe ratio, minimum semi-variance, or conditional value at risk | Risk parity across a correlation hierarchy |
| Inputs | Covariance matrix from a risk model; expected returns for return-seeking objectives; weight constraints | Correlation matrix, a linkage choice, a correlation window |
| Estimation error | Propagated through matrix inversion; can move weights sharply | Absorbed by clustering and recursive bisection; more stable |
| Typical weight concentration | Can concentrate in a few names when correlations are unstable | Usually more evenly spread across the correlation structure |
| Computational complexity | Matrix inversion dominates; grows steeply with candidate count | Clustering plus recursive splits; no inversion |
| RSJ Portfolio support | Minimum volatility, maximum Sharpe ratio, minimum semi-variance, conditional value at risk, with eight selectable risk models | Hierarchical risk parity, selectable as an allocation method alongside the MPT objectives |
How RSJ Portfolio Implements Both Methods
The workflow is the same regardless of method, which is what makes a genuine side-by-side comparison practical.
- Build the candidate list. Search symbols and add them in the Selection step of the user guide.
- Configure the data layer once. Data provider, local cache, local currency, date range, and cleaning settings. This is the Configuration chapter, and it is worth doing before your first run — the data page explains the cleaning pipeline and the currency conversion path.
- Pick the allocation method. Minimum volatility, maximum Sharpe ratio, minimum semi-variance, conditional value at risk, or hierarchical risk parity.
- Set the risk model and constraints. Eight risk estimators are available, plus return models and weight constraints. The features page lists them all.
- Run the optimizer and read the results. Weights and portfolio statistics come back in the Recommendation view; exports are available for further analysis.
The installer is a Windows executable with a versioned filename, and the SHA-512 hash is published on the downloads page. Verify it before you run anything:
Get-FileHash .\RSJPortfolio-Setup-<version>.exe -Algorithm SHA512
# compare the output to the hash printed on the downloads page
Then run both methods on the same candidate list and export the weights to compare:
symbol,weight
AAA,0.0812
BBB,0.0734
CCC,0.0651
Exports give you the two weight vectors side by side without rerunning either optimization by hand. You could build both optimizers yourself — the math is public and a correlation-distance matrix plus a linkage call is a dozen lines — but that is the hard way, and it leaves you maintaining a solver instead of evaluating a portfolio. RSJ Portfolio runs the cleaning, caching, and optimization as one managed step from a desktop UI.
For a deeper look at the risk-estimator dropdown specifically, see Portfolio Optimizer with Selectable Risk Models: A Comparison.
Verdict: Who Should Pick Which?
Choose MPT if you have return forecasts you actually believe, you want to target a specific objective such as maximum Sharpe ratio or minimum volatility, and you are comfortable reviewing concentrated weights rather than accepting them.
Choose HRP if you want a diversification-first allocation that does not depend on expected returns, and you would rather have a robust answer than a precise one built on noisy correlations.
Run both if you are undecided — which, honestly, is most of the time. RSJ Portfolio keeps both methods one selection away on the same data, so the comparison is cheap. Whichever you pick, check the output before you trade it; How to Validate Portfolio Optimizer Output Before You Trade covers that step, and Portfolio Optimization for Multi-Currency Portfolios: A Guide covers the currency side.
FAQ
Which method is better for reducing volatility: MPT or HRP?
It depends on your inputs and goals. MPT can directly minimize volatility — or maximize the Sharpe ratio — but it is sensitive to estimation error in the covariance matrix. HRP tends to produce more diversified, robust allocations by using hierarchical clustering, which can lead to lower out-of-sample volatility when correlations are noisy. RSJ Portfolio lets you run both and compare results.
Can I use both MPT and HRP in RSJ Portfolio?
Yes. RSJ Portfolio includes both modern portfolio theory objectives (minimum volatility, maximum Sharpe ratio, minimum semi-variance, conditional value at risk) and hierarchical risk parity as selectable allocation methods. You can switch between them and compare outputs.
Does HRP require expected returns?
No. Hierarchical risk parity allocates risk based on the correlation structure and does not require expected return forecasts. That makes it less prone to errors from return estimation — unlike MPT, which typically uses expected returns.
How does RSJ Portfolio handle correlation estimation for these methods?
RSJ Portfolio computes correlations from historic end-of-day prices after cleaning. You can select the risk model (eight options) and the date range. For HRP, it builds a hierarchical tree from the correlation matrix; for MPT, it uses the covariance matrix produced by the chosen risk model.
Conclusion
MPT and HRP are not rivals so much as two answers to the same question with different tolerance for bad inputs. MPT optimizes harder and asks more of your estimates; HRP optimizes more cautiously and asks almost nothing of your return views. If you cannot articulate why your covariance matrix is trustworthy this quarter, HRP is usually the safer default — and if you can, MPT gives you a sharper target.
RSJ Portfolio puts both behind the same candidate list, the same cleaned price history, and the same one-click run, so the decision becomes empirical: run both, export the weights, and keep the one whose behaviour you can defend. It is a Windows desktop application under a single-user license — start with the getting started guide, verify the installer hash from the downloads page, and run your first comparison this week.
Related posts
- Portfolio Optimization for Multi-Currency Portfolios: A Guide
- Why Low Correlation Stocks Reduce Portfolio Volatility
- Portfolio Optimizer with Selectable Risk Models: A Comparison