Hierarchical Risk Parity vs Minimum Variance: Which Allocation Fits?
You have a candidate list, a few years of end-of-day prices, and one decision in front of you: which objective should turn that list into weights? For most people building a portfolio by hand, it comes down to two: minimum variance, which pushes expected volatility as low as the data allows, and hierarchical risk parity (HRP), which clusters assets by how they move together and spreads risk across that structure.
Both reduce portfolio volatility. They get there differently, and the difference shows up in the weights you end up holding. This article compares them on the things that actually matter when you run an optimizer: how they treat correlation, how much they trust your estimates, how concentrated the output is, and which one fits your goals.
The good news: in RSJ Portfolio both methods are selectable objectives in the same optimizer, on the same candidate list, with the same cleaned price data. So the question is not which tool to buy — it is which objective fits your portfolio.
The Allocation Decision: HRP or Minimum Variance?
The choice appears the moment you stop picking stocks and start sizing them. You know what you want to hold. Now you need to decide how much of each — and that is a statistical question, not a conviction.
Two philosophies collide here:
- Minimize volatility directly. Treat the portfolio as an optimization problem: find the weight vector that produces the lowest expected portfolio variance given your risk model. Every asset competes for weight, and the lowest-volatility, least-redundant names win the most.
- Allocate risk across a hierarchy. Group assets by correlation, then distribute risk top-down through those groups instead of solving one big global problem. No asset gets a weight simply because its covariance with everything else happened to look good in-sample.
Both aim at the same outcome — a portfolio that moves less than its components would on their own — but they disagree about how much to trust a covariance matrix and how evenly risk should be spread.
The criteria this comparison uses are the ones you can actually evaluate before committing capital: ease of use, degree of automation, transparency of the result, how each behaves across market conditions, and how cleanly each integrates with the rest of your workflow. Cost enters only as a footnote: RSJ Portfolio is licensed per single user, with the details on the pricing and license page — the same license regardless of which objective you select.
How Minimum Variance and Hierarchical Risk Parity Work
Minimum variance: optimize volatility directly
Minimum variance is the classical modern-portfolio-theory answer with the return forecast removed. You supply a risk model (RSJ Portfolio ships eight of them, plus selectable return models), and the optimizer searches for the weight combination with the smallest expected portfolio volatility.
Mathematically, the core is a covariance matrix and its inverse. Strip it to the bone and it looks like this:
import numpy as np
# Illustrative only: the naive closed-form minimum-variance weights
cov = returns.cov().values # needs to be well-conditioned
inv = np.linalg.inv(cov) # the fragile step
ones = np.ones(len(cov))
w = inv @ ones / (ones @ inv @ ones)
That single inversion is where the method’s strength and its weakness live. When the risk model describes the world accurately, the result is genuinely the lowest-volatility portfolio available from your universe. When the matrix is noisy — many correlated names, a short history, a regime that has moved on — the inverse amplifies estimation error, and weights can land in odd places. The optimizer does exactly what you asked; it just asked a matrix that was guessing.
Because low-volatility assets have the smallest variance terms, minimum variance naturally tilts toward them. Concentration is a feature, not a bug: if three defensive names genuinely dampen portfolio swings, they earn large weights.
Hierarchical risk parity: cluster first, allocate second
HRP takes a different route. It never inverts the covariance matrix.
- Cluster. Assets are grouped by correlation distance, building a tree of which names behave alike.
- Order. The tree is rearranged so similar assets sit next to each other.
- Bisect and allocate. Risk is split recursively down the tree — first between the two halves of the portfolio, then within each half, and so on — with each split weighted by the inverse of that group’s variance.
The output spreads risk more evenly across correlated groups. Two near-identical banks will not both receive large weights simply because each looked good individually; they sit in the same cluster and share its allocation. HRP is more forgiving of a noisy covariance estimate, because clustering only needs the ordering of pairwise relationships to be roughly right, not their precise magnitudes.
Where they diverge
- Minimum variance optimizes correlation effects globally — every pairwise relationship is priced in one solve.
- HRP handles correlation structurally — through grouping and recursive allocation.
- Minimum variance can concentrate in a few low-volatility names. HRP tends to spread risk across clusters, which usually means more positions with smaller individual weights.
- Minimum variance is sensitive to the risk model choice; HRP is less so, because it leans on relative structure rather than absolute estimates.
This is not a two-horse race, either. RSJ Portfolio also exposes minimum semivariance, maximum Sharpe ratio, and conditional value at risk as objectives, and the methods and features page lists every one alongside the risk models, return models, weight constraints, and export options. If you are weighing a different pair of objectives, the same reasoning applies — see Maximum Sharpe Ratio vs Minimum Semivariance Allocation.
Selection Criteria: Which Method Fits Your Portfolio?
- Ease of use. Minimum variance asks more of your inputs: a risk model you trust and a history long enough for the covariance estimate to mean something. HRP asks less — it still needs the correlation structure, but degrades more gracefully when that structure is imperfect. If your candidate list is long and internally correlated, HRP is the lower-maintenance starting point.
- Automation. Both are one selection in RSJ Portfolio’s recommendation step. You change the objective, change the risk estimator, re-run, and read the allocation. No scripting, no matrix plumbing — the recommendation chapter of the user guide walks through the parameters and the resulting allocation.
- Transparency. Minimum variance gives you a clean headline: this is the lowest expected volatility from this universe under this risk model. HRP gives you a structure: this is how risk is distributed across clusters. Which is easier to explain depends on your audience. A committee usually understands “we spread risk across these groups” faster than “we inverted a covariance matrix.”
- Integration. Both run on the same cleaned price history, the same local cache, and the same multi-currency handling. If your candidates trade in several currencies, the data and currencies page documents the conversion path, so both methods see the same comparable return series. Exports work identically, too — useful when you want to compare weights offline (see Exporting Optimized Portfolio Weights to a Spreadsheet).
- Switching cost. Effectively zero. The objective is a parameter, not a commitment. That matters more than any single backtest.
Before you compare anything, make sure the inputs are sane. Data provider, cache, local currency, date range, and cleaning settings all shift the result — the first-run setup guide covers the settings worth fixing before your first optimization run.
Side-by-Side Comparison: HRP vs Minimum Variance
| Dimension | Minimum variance | Hierarchical risk parity |
|---|---|---|
| Objective | Find the weight vector with the lowest expected portfolio volatility | Allocate risk across a correlation-derived hierarchy |
| Handling of correlation | Direct optimization — all pairwise covariances priced in one solve | Clustering — assets grouped by correlation, then risk bisected down the tree |
| Covariance matrix inversion | Required | Not required |
| Required inputs | Risk model, return history, weight constraints | Correlation structure from the same return history |
| Sensitivity to estimation error | Higher — the inverse amplifies noise | Lower — relies on relative ordering, not precise magnitudes |
| Typical weight concentration | Can concentrate in a few low-volatility names | Spreads risk more evenly across clusters; usually more positions |
| Computational complexity | Grows quickly with the number of assets | Scales comfortably to larger universes |
| Calm markets | Tends to hold the genuinely low-volatility names; may look under-diversified | Spreads risk broadly; may hold more assets than strictly needed |
| Stress periods | Concentration can help or hurt depending on whether the low-vol names stay low-vol | Diversification across clusters tends to hold up better when correlations shift |
| Running it in RSJ Portfolio | Select it as the objective, pick a risk model, set constraints, run | Select it as the objective, run — same candidate list, same cleaned data |
A practical note on that last row: because both methods run against exactly the same dataset in RSJ Portfolio, the comparison is genuinely apples-to-apples. Run minimum variance, export or note the weights, switch the objective to hierarchical risk parity, run again, and compare. Then look at what the two portfolios actually hold:
import pandas as pd
# One column per optimization run, one row per symbol.
# Build this file from the weights you exported for each run.
weights = pd.read_csv("weights_comparison.csv", index_col="symbol")
def concentration(w: pd.Series) -> dict:
w = w / w.sum()
return {
"positions": int((w > 0).sum()),
"effective_n": round(1 / (w ** 2).sum(), 2), # 1 / Herfindahl index
"top3_share": round(w.nlargest(3).sum(), 4),
}
for column in ["min_variance", "hrp"]:
print(column, concentration(weights[column]))
Two numbers usually tell the story: how many positions carry meaningful weight, and how much of the portfolio sits in the top three. If minimum variance gives you six positions and 55% in the top three while HRP gives you fourteen positions and 30%, you are looking at the concentration trade-off in hard numbers.
Verdict: Who Should Pick Which?
Pick minimum variance if your priority is the lowest achievable expected volatility from your universe, you have a risk model you trust and enough clean history behind it, and you are comfortable holding a concentrated book of low-volatility names. It is the sharper tool when the estimates are good, and it gives you a single, defensible number to point at.
Pick hierarchical risk parity if you want risk spread across correlated groups rather than concentrated in whatever looked safest in-sample, your candidate list is large or heavily correlated, or you are skeptical of your covariance estimates. HRP is the more robust default — it gives up some theoretical optimality in exchange for not over-fitting the matrix.
Most likely, pick both — temporarily. Run each objective on your candidate list and compare allocations, risk metrics, and weight concentration before committing capital. The whole point of a selectable objective is that you can change your mind cheaply. And you should: as your views, your universe, or the market regime change, the objective that fits may change with it.
RSJ Portfolio is a Windows desktop application, so the workflow is install-and-run rather than a cloud pipeline. Verify the installer before you launch it — the downloads page publishes a SHA-512 hash alongside a versioned filename:
# Compare this output against the SHA-512 published on the downloads page
Get-FileHash .\<installer>.exe -Algorithm SHA512
Then set your data provider, currency, and date range once, and start comparing objectives. The getting started guide lists the requirements and the recommended starting parameters. If you would rather do the clustering and recursive bisection yourself in Python, you can — that is the hard way, and it is a week you will not get back.
FAQ
Is hierarchical risk parity always better than minimum variance? No. HRP is more robust to estimation errors and tends to spread risk more evenly, but minimum variance can achieve lower volatility if your risk model is accurate. The best choice depends on your data quality and goals. RSJ Portfolio lets you compare both.
How does RSJ Portfolio help me choose between HRP and minimum variance? RSJ Portfolio implements both methods as selectable objectives. You can run both on the same candidate list and compare the resulting allocations, risk metrics, and weight concentrations before deciding.
What data do I need to run these methods in RSJ Portfolio? You need historical end-of-day prices for your candidate stocks. RSJ Portfolio handles data cleaning, currency conversion, and caching automatically, so you can focus on comparing methods.
Can I use both methods together in one portfolio? RSJ Portfolio allows you to select one objective per optimization run. You can run separate optimizations and combine insights manually, or use the export feature to analyze results in a spreadsheet.
Which method is more computationally intensive? Minimum variance grows quickly with the size of your universe because it works with the full covariance matrix. HRP scales more comfortably, since it operates on a clustering tree. In practice, on a desktop machine, both finish fast enough that computation is not the constraint — data quality is.
How do I avoid overfitting either result? Keep the comparison honest: same candidate list, same date range, same cleaning settings, same risk model where applicable. Then check whether the resulting weights survive a change of window — a portfolio that only works on one specific history is a curve-fit, not an allocation. The concept chapter covers the statistical background, and How to Verify a Portfolio Optimizer’s Allocation Results walks through the checks.
Conclusion
Minimum variance and hierarchical risk parity both reduce portfolio volatility; they disagree about how much to trust a covariance matrix and how evenly risk should be spread. If your estimates are solid and you want the lowest possible volatility, minimize variance. If you want resilience to estimation error and a more even risk distribution, use HRP. If you are not sure — and you probably shouldn’t be until you have seen both — run them both and compare the weights.
RSJ Portfolio makes that comparison a two-click exercise: install it, point it at your candidate list, and switch the objective between runs. Start with the overview to see the full workflow, then download the installer and let the data decide.
Related posts
- How to Scan a Watchlist for Low-Correlation Stocks
- Exporting Optimized Portfolio Weights to a Spreadsheet
- How to Verify a Portfolio Optimizer’s Allocation Results