RSJ Portfolio Portfolio Optimization

← All posts

2026-11-03

Minimum Variance vs Hierarchical Risk Parity: Which Fits?

Minimum Variance vs Hierarchical Risk Parity: Which Fits?
portfolio optimization modern portfolio theory portfolio correlation portfolio optimization hierachical risk parity minimize volatility minimize semivariance maximize sharpe

Minimum Variance vs Hierarchical Risk Parity: Which Fits?

Minimum variance vs hierarchical risk parity is a decision you make before you run a single optimization — and it quietly determines how much you trust your own covariance estimate. Both methods are correlation-aware, both target a calmer portfolio, and both take the same input: historic end-of-day prices. What separates them is what they do with the correlation structure once they have it. Minimum variance inverts a covariance matrix and lets the optimizer find the lowest-volatility combination. Hierarchical risk parity clusters assets by correlation and splits risk down the resulting tree, never inverting anything.

That difference matters if you want a statistically grounded allocation instead of a hand-picked mix of “things that don’t move together.” This article compares the two methods on the criteria that actually decide the outcome — estimation error, asset count, automation, data plumbing — and shows how RSJ Portfolio runs both on the same candidate list so you can compare them without rebuilding your workflow.

The Landscape of Correlation-Aware Allocation Methods

Minimum variance is a modern portfolio theory objective in the classic sense. You give the optimizer a covariance matrix estimated from historic returns, and it solves for the weights that minimize portfolio variance:

minimize    wᵀ Σ w
subject to  Σ wᵢ = 1
            wᵢ ≥ 0   (and any other weight constraints you set)

Everything the result depends on lives in Σ. If Σ is estimated from a long, clean history and the candidate list is modest, the solution is stable and genuinely low-volatility. If Σ is noisy or ill-conditioned — many assets, short history, missing prices — the inverse amplifies estimation error and the weights can concentrate in ways that look arbitrary.

Hierarchical risk parity (López de Prado, 2016) takes a different route. It converts pairwise correlations into a distance metric, builds a hierarchical clustering tree, reorders the covariance matrix so similar assets sit next to each other, and then allocates risk top-down through the tree by recursive bisection. No matrix inversion anywhere in the pipeline. The practical consequence: HRP degrades more gracefully when correlations are estimated badly, which is the normal condition in a large universe.

Both are correlation-aware in the sense the product positioning cares about — they reward assets that correlate weakly, or whose downside correlates weakly, and thereby reduce the volatility of the resulting portfolio. They just get there differently.

RSJ Portfolio implements both, alongside maximum Sharpe ratio, minimum semi variance and conditional value at risk. The full list, the eight risk models, the return models and the export options are on the methods and features page; the statistical background behind each objective is in the concept chapter. If you want the deeper treatments of the neighbouring objectives, they’re covered separately: minimum semi variance and conditional value at risk.

Selection Criteria: How to Choose Between the Two

Ease of use and estimation demands

Minimum variance asks more of your data. It needs a covariance estimate stable enough to invert, which in practice means enough observations relative to the number of assets and a price series without gaps. HRP asks for a correlation matrix and a clustering step — more moving parts in the algorithm, less fragility in the input.

As a rule of thumb: the smaller and cleaner your universe, the less the two methods disagree. The larger and noisier it gets, the more HRP’s robustness shows up in the weights.

Automation

Both are objectives you select in RSJ Portfolio’s optimizer. Data cleaning, risk model selection, return model handling and weight constraints are applied by the application rather than by you writing a solver loop — which is the difference between running the comparison once and running it every month.

# The math, sketched — this is what you'd maintain yourself without a tool.
import numpy as np

cov = np.array([
    [0.0400, 0.0090, 0.0045],
    [0.0090, 0.0625, 0.0060],
    [0.0045, 0.0060, 0.0324],
])

def portfolio_vol(weights, cov):
    return float(np.sqrt(weights @ cov @ weights))

# minimum variance, unconstrained and fully invested: w proportional to Sigma^-1 * 1
inv = np.linalg.inv(cov)
w_mv = inv @ np.ones(3)
w_mv /= w_mv.sum()

# HRP: no inversion, risk split through a cluster tree
w_hrp = np.array([0.42, 0.31, 0.27])

for name, w in (("min variance", w_mv), ("HRP", w_hrp)):
    print(f"{name:13s} vol={portfolio_vol(w, cov):.2%}  weights={np.round(w, 3)}")

The clustering half of HRP is the part people underestimate:

import numpy as np
from scipy.cluster.hierarchy import linkage

returns = np.array([...])                  # historic end-of-day returns, one column per stock
corr = np.corrcoef(returns, rowvar=False)
dist = np.sqrt(0.5 * (1.0 - corr))         # correlation distance
tree = linkage(dist[np.triu_indices_from(dist, k=1)], method="single")

# then: quasi-diagonalize, recursive bisection, inverse-variance split at each node

That’s a maintainable afternoon project. Keeping it correct across data providers, currencies, missing prices and rebalances is not.

Reliability and where your data lives

RSJ Portfolio is a Windows desktop application with a local cache. Prices you pull stay on your machine, and repeat runs read from the cache instead of hammering the provider — useful when you re-run the same candidate list under different objectives. The data and currencies page describes the providers, the price cleaning pipeline and the conversion path for portfolios quoted in several currencies.

Price and integration

A single user license covers the application; the pricing page has the current terms. Results export for further analysis, so neither method becomes a dead end — you can take the allocation into your own reporting or backtesting stack. Compare that to the cost of owning a custom optimizer: the solver is the cheap part, the data pipeline is the expensive one.

Data requirements

Both methods consume historic end-of-day prices. Neither needs intraday data, order books or a live feed. Multi-currency candidate lists are handled through the product’s conversion path, and cleaning settings are configuration, not code.

# A decision sheet worth filling in before you touch either optimizer
minimum_variance:
  input: covariance matrix, estimated
  failure_mode: ill-conditioned estimate -> concentrated, unstable weights
  best_when: assets are few and the price history is long
hierarchical_risk_parity:
  input: correlation matrix plus a cluster tree
  failure_mode: cluster structure drives the allocation
  best_when: assets are many and pairwise correlations are noisy

Side-by-Side Comparison: Minimum Variance vs Hierarchical Risk Parity

Minimum variance Hierarchical risk parity
Objective Minimize portfolio variance for the given candidate set Distribute risk evenly across a hierarchy of correlated assets
Risk estimator Covariance matrix — sample estimate or one of the selectable risk models Correlation matrix converted to a distance metric, plus cluster-level variances
Handling of correlation Correlation enters directly through the covariance matrix; low-correlation pairs get more weight Assets are clustered by correlation, then risk is allocated top-down through the tree
Sensitivity to estimation error Higher: inverting a noisy matrix can concentrate weights Lower: no matrix inversion, so noisy pairwise correlations matter less
Typical use case A focused universe with a long, clean history and high confidence in the risk estimate A broad universe, short or patchy histories, and an allocation you want to hold through rebalances
Support in RSJ Portfolio Available as the minimum volatility objective Available as the hierarchical risk parity objective

Because RSJ Portfolio exposes eight risk models and selectable weight constraints, the comparison isn’t a one-shot coin flip — you can run minimum variance against HRP under the same risk model, then re-run both under a different one, and see whether the ranking survives. If the answer flips with the risk model, you’ve learned something more useful than either individual result.

Verdict: Who Should Pick Which?

Pick minimum variance if your candidate list is concentrated enough that the covariance matrix is well-conditioned, your history is long and clean, and your goal is literally the lowest achievable volatility for that set of assets. It’s the direct answer to the direct question, and when the estimate is trustworthy it usually wins on realized volatility.

Pick hierarchical risk parity if you’re running dozens of names, your correlations are noisy or regime-dependent, or you’ve been burned by an optimizer that piled 60% into two assets because of a numerical accident. HRP’s allocation tends to be more diversified by construction, and it doesn’t require you to trust a matrix inverse.

In practice, run both. The two methods are selectable objectives in the same application working from the same cached price data, so the marginal cost of the second run is a few clicks. Compare the resulting allocations, compare the ex-ante volatility, and check whether the difference is inside the noise of your estimates. If the candidate list has been giving you concentrated weights, that’s your signal to prefer HRP; if you have a tight universe and want the minimum, minimum variance is doing exactly what it says. For the broader question of why low-correlation names lower volatility in the first place, see this piece.

Putting Both Methods to Work in RSJ Portfolio

  1. Install and license. The Windows installer is on the downloads page, published under a versioned filename with its SHA-512 hash on the same page. Verify the hash before running it; the walkthrough for that covers the command-line side. Then apply your license.
  2. Configure data and currency. Choose your data provider, set the local currency, pick the date range and review the cleaning settings. This is the step that decides whether your covariance estimate is usable — a five-year window with gaps behaves very differently from a clean ten-year one. Details in the configuration chapter.
  3. Build the candidate list. Search symbols and assemble the universe you actually want to allocate across, as described in the selection chapter. Everything downstream is only as good as this list.
  4. Pick the objective. Select minimum volatility or hierarchical risk parity. Then choose a risk model and set your weight constraints. Change one variable at a time if you want to learn anything from the comparison.
  5. Run and read the result. The optimizer returns the recommended allocation; the recommendation chapter explains what each parameter and output means. Export it if you’re taking it into reporting.
  6. Repeat under the other objective. Same list, same data, same risk model — one different setting. That’s your comparison.

If you’re running this on a schedule rather than ad hoc, the cached end-of-day workflow is worth reading next: it’s the difference between a monthly manual refresh and a repeatable process.

FAQ

What is the main difference between minimum variance and hierarchical risk parity?

Minimum variance minimizes portfolio volatility using a covariance matrix, while hierarchical risk parity allocates risk across a hierarchy of correlated assets, avoiding matrix inversion and often being more robust to estimation error.

Can I use both methods in RSJ Portfolio?

Yes. RSJ Portfolio lets you select minimum variance, hierarchical risk parity and other objectives, so you can compare allocations on the same data set without switching tools or re-preparing your prices.

Which method is better for portfolios with many assets?

Hierarchical risk parity is often preferred for large portfolios because it handles noisy correlations better, but minimum variance can work well if you have a stable covariance estimate. Testing both in RSJ Portfolio on your own candidate list is the fastest way to decide.

Do I need to provide my own correlation matrix?

No. RSJ Portfolio computes correlations from historic end-of-day prices using your chosen data provider and risk model — you supply the symbols and the configuration, not the statistics.

Conclusion

Minimum variance and hierarchical risk parity aren’t competing answers to the same question so much as two answers to a slightly different one: how much do you trust your correlation estimate? When you trust it, minimum variance extracts the lowest volatility available. When you don’t — many assets, patchy history, regime shifts — HRP’s cluster-based allocation holds up better and stays diversified.

The only way to know which applies to your universe is to run both on the same data and compare. RSJ Portfolio is built for exactly that: select symbols, configure the data and risk model, then switch objectives and read the difference. Start with the getting started guide for the configuration worth setting before your first run, and see the downloads page for the installer and the full user guide.

Related posts

EU label: AI-generated content