Why Acubic Uses Hierarchical Risk Parity

Portfolio construction has two distinct problems. The first is which assets to include. The second is how much of each to hold. Most attention goes to the first problem, but the second is where real diversification is won or lost. Acubic uses hierarchical risk parity as its default method for the second problem, and this article explains why.

If you want the full pipeline from data inputs through ranking and weighting, what data does Acubic use covers it end to end. This piece goes deeper on one step: why HRP, why not mean-variance optimization, and when you might want to change the default.

What the weighting step actually has to do

Once a shortlist of holdings has been selected and ranked, the construction method assigns a weight to each one. That weight determines how much risk the holding actually contributes to the full portfolio. Getting this wrong is straightforward: assign too much capital to a group of stocks that move together, and the stated diversification is cosmetic. You own ten names, but you have taken one bet.

The naive solution is equal weighting. It sidesteps estimation error almost entirely by ignoring correlation. Its weakness is that ten holdings with strong correlations behave like fewer independent positions regardless of how the capital is split. A proper construction method has to use the correlation structure deliberately, treating diversification as something to be achieved in risk space rather than just in the count of holdings.

Why mean-variance optimization is harder to use reliably

Mean-variance optimization is the standard framework in portfolio theory. It takes expected returns, volatilities, and correlations and finds the allocation that offers the highest expected return for a given level of portfolio risk. The output is a point on the efficient frontier: the best available risk-return trade-off under the inputs provided.

The difficulty is in those last three words. The optimizer accepts its inputs as given. When the expected return estimate for one asset is slightly too high relative to the others, the optimizer allocates heavily to it. When an estimate is slightly too low, the optimizer avoids or underweights that asset. Small errors in return estimates become large distortions in the final weights, because the optimizer is doing exactly what it is designed to do: concentrating capital where the numbers look best.

In practice this means unconstrained mean-variance portfolios tend to hold very few names at extreme weights, change dramatically between periods as estimates shift, and perform worse out of sample than the historical inputs imply. That behavior is not a failure of any particular implementation. It is a structural property of optimizing on noisy return forecasts. The mean-variance optimization article covers what the framework gets right and the implementation choices that constrain its output; the instability problem is the reason more robust alternatives were developed.

How hierarchical risk parity works

HRP was developed to preserve the goal of risk-spreading diversification while removing the dependence on return estimates that makes MVO unstable. It operates in three steps.

The first step is tree clustering. HRP analyzes the historical correlation matrix of the candidate holdings and groups them into a hierarchical tree. Assets whose return histories move together appear on adjacent branches. Stocks in the same sector, funds tracking similar themes, or any holdings that tend to rise and fall in tandem cluster together without the investor having to define those groups in advance. The structure comes from the data, not from a judgment about which sectors belong together.

The second step reorganizes the correlation matrix so that correlated assets appear near each other. This creates a structure where the weighting step can work along the branches of the tree rather than treating every pair of assets as equally independent from every other.

The third step is recursive bisection. Working down the tree from the top, the algorithm divides capital between branches based on their relative variance. A more volatile branch receives a smaller share of the capital allocated to that fork in the tree. A less volatile branch receives a larger share. This continues at every level until each individual holding has a weight. The result is a portfolio whose capital is distributed across the natural risk clusters the data identifies, rather than concentrated in the assets with the best historical return numbers.

This is the meaning of hierarchical in the method's name. The diversification happens at the level of groups, working down through layers of correlation structure, not just at the level of individual asset pairs. A set of highly correlated holdings contributes as one cluster, not as ten independent sources of return.

Why HRP handles estimation error better

The key difference between HRP and MVO is what each method uses as its primary input. Mean-variance optimization requires expected return estimates, which are noisy and sensitive to the period of history used to derive them. HRP uses only the covariance structure of returns, and it works with that structure hierarchically rather than inverting the full matrix as MVO does.

Inverting the covariance matrix amplifies small errors in the estimates. Assets whose correlations are slightly misestimated receive systematically too much or too little weight as a result. HRP avoids this because its bisection step compares variance at the branch level, not at the individual asset level with a full matrix inverse. The resulting weights are less sensitive to any single misestimated correlation.

The practical consequence is that HRP portfolios change less dramatically from one scheduled rebalance to the next. Broad risk relationships between sectors and asset types are more persistent over time than any single stock's expected return. A portfolio structured around those relationships stays coherent as prices move, which matters when the goal is long-term maintenance rather than reacting to the most recent data. This connects to the broader risk parity family of methods: the shared objective is a deliberate risk budget, not a concentrated return bet.

How HRP fits Acubic's construction process

In Acubic's pipeline, the construction method receives a shortlist of holdings that have already passed the screening step for your constraints and been ranked by the selected method: market capitalization, trailing Sharpe ratio, or trailing momentum. HRP assigns weights to that shortlist without using the ranking scores themselves as inputs.

Because HRP does not depend on return estimates, the ranking step and the weighting step operate separately. An asset that ranked well on trailing Sharpe ratio does not automatically receive a higher weight. Its weight depends on how its historical returns correlate with the other shortlisted holdings, and on how much variance the branch it belongs to contributes relative to the others. That separation means the final allocation is not dominated by whichever lookback window was used in the ranking step.

At each scheduled rebalance, the full construction process reruns from the current data. Correlations shift as market conditions change, and each run of HRP reflects the current structure rather than the one at the time the portfolio was first built. For the mechanics of how the rebalance cycle works, see how does Acubic work; for the full method including all three construction approaches and the data that feeds them, the methodology page is the authoritative reference.

When MVO or CVaR fits better than HRP

HRP is the default because it produces stable, diversified portfolios without requiring return estimates or calibration choices beyond the constraint settings you already provide. It fits the common case well. It is not the only reasonable choice, and understanding when to change it is part of understanding the method.

Mean-variance optimization is available for users who want to lean toward higher expected risk-adjusted returns and are prepared to accept greater sensitivity to the historical inputs. Acubic applies constraints that prevent the most extreme MVO corner solutions, so in practice it produces a focused portfolio within the asset set you specified rather than an unconstrained one. MVO fits best when you have a clear view on risk-adjusted historical performance and want the optimizer to express that view through more concentrated weights.

CVaR minimization sits at the other end of the spectrum. Rather than targeting the average relationship between return and volatility, it minimizes the expected loss in the worst outcomes defined by a percentile threshold. A CVaR-optimized portfolio will typically reduce the severity of tail losses at the cost of lower expected returns in normal market conditions. It fits when limiting downside is the primary objective and the trade-off of lower expected upside is one you have explicitly considered.

Choosing deliberately between the three methods is better than leaving the default unchanged without understanding what it does. HRP is the right starting point for most users precisely because it requires fewer assumptions and remains stable. But the construction choice should reflect the objective you are actually trying to achieve, not just the default setting.

Getting started

The construction method is one of the settings in the AI portfolio builder. HRP is the default and does not need to be changed to build a well-diversified portfolio. MVO and CVaR are available in the advanced settings for users who want them. All three methods receive the same screened and ranked shortlist; the only difference is how the weighting step runs.

Acubic does not promise a return and does not predict short-term market moves. It applies a stated, documented construction method to the inputs you provide and maintains the result on the schedule you control. The value of HRP is not that it outperforms in any particular period. It is that the risk allocation it produces is deliberate, grounded in the correlation structure of the holdings, and stable enough to maintain on a regular schedule without constant recalibration.