RSJ Portfolio Create a statistically optimized portfolio

← All posts

2026-11-20

Risk Models in Modern Portfolio Theory, Compared

Risk Models in Modern Portfolio Theory, Compared
portfolio optimization modern portfolio theory portfolio correlation portfolio optimization hierachical risk parity minimize volatility minimize semivariance maximize sharpe

Risk Models in Modern Portfolio Theory, Compared

Give two investors the same candidate list, the same expected returns, and the same optimization objective, and they can still end up with different portfolios. The reason is usually not the optimizer. It is the risk model.

Modern portfolio theory needs a number for risk before it can allocate anything: a covariance matrix estimated from historical prices. That estimate is where the judgment lives. The optimizer is arithmetic — feed it two different covariance matrices and it will hand you two different sets of weights, both “optimal” by its own lights. So when you compare modern portfolio theory risk models, you are really comparing estimators of the same quantity, and deciding which one’s blind spots you can live with.

RSJ Portfolio ships eight selectable risk models, which makes this a testable question rather than a theoretical one. This article is a comparison for investors and quants who have to pick one. It is not an MPT tutorial, and it is not a deep dive on any single estimator — for that, see What Is Semi Variance in Portfolio Optimization?.

The Risk Model Decision: Why Your Estimator Matters More Than Your Optimizer

The objective answers “what am I optimizing for?” — minimum volatility, maximum Sharpe ratio, minimum semi variance, hierarchical risk parity, conditional value at risk. The risk model answers “what does risk look like in this candidate set?”. Change the objective and you change the question. Change the risk model and you change the units the question is asked in.

That second change is easy to underestimate. Assume 30 candidates and five years of daily prices: roughly 465 parameters have to be estimated from about 1,260 observations. The sample covariance matrix will fit that data, including its noise. A shrinkage estimator will deliberately bias the same matrix toward a simpler structure, trading a little bias for a lot less variance. Both are defensible. They will not produce the same weights.

You can see the effect in a dozen lines of Python — this is the hard way, computing the estimators yourself:

import numpy as np
from sklearn.covariance import LedoitWolf

# returns: T x N array, one column per candidate, daily simple returns
S_sample = np.cov(returns, rowvar=False) * 252
S_shrunk = LedoitWolf().fit(returns).covariance_ * 252

def min_variance_weights(S):
    inv = np.linalg.pinv(S)
    ones = np.ones(S.shape[0])
    return inv @ ones / (ones @ inv @ ones)

print(min_variance_weights(S_sample).round(3))
print(min_variance_weights(S_shrunk).round(3))

Same return data. Same optimizer. Different portfolio. In RSJ Portfolio the risk estimator is a selectable parameter alongside the objective and the weight constraints, so the comparison is a re-run rather than a rewrite.

The Landscape: Eight Risk Models in Modern Portfolio Theory

RSJ Portfolio exposes eight risk models. The exact roster is documented on the methods and features page and in the user guide; what matters here is the vocabulary those eight options share. In practice, risk estimators fall into four families.

Sample-based — the raw covariance matrix, or a weighted version of it. They estimate co-movement directly from the price history with no structural assumptions. Typical use: small candidate lists with long, clean histories.

Factor-based — covariance derived from exposure to one or more common factors rather than from every pairwise relationship. They estimate a much smaller parameter set. Typical use: large candidate lists where the full matrix cannot be estimated reliably.

Shrinkage — the sample matrix pulled toward a structured target (constant correlation, a factor model, a diagonal). They estimate the same object as the sample model but with a deliberate bias to control estimation error. Typical use: the default “robust” middle ground.

Downside-focused — covariance computed only over negative or below-target returns. They estimate co-movement in losses rather than in all variation. Typical use: objectives that care about drawdowns rather than total dispersion.

The choice does not stand alone. It interacts with the objective: minimum semi variance pairs naturally with a downside estimator, hierarchical risk parity needs a correlation structure it can cluster, and conditional value at risk cares about tail co-movement in a way that a plain sample matrix only partially captures. If you are still deciding between the two most common objectives, read Minimum Variance vs Maximum Sharpe Portfolio: Which Is Better? first, then come back to the estimator.

Selection Criteria: How to Compare Risk Models for Your Portfolio

Five criteria do most of the work when you are choosing between estimators.

  • Ease of use and data requirements. Sample covariance needs nothing but prices. Factor models need factor returns or a defensible proxy. Downside estimators need enough negative observations to estimate anything at all. Every extra input is a place for the data to be wrong, and every tuning parameter is a choice you can get wrong.
  • Automation. A risk model is only useful if you can re-estimate it as new prices arrive and as your candidate list changes. RSJ Portfolio handles this through its data providers, cleaning pipeline and local cache, so re-running the optimizer against refreshed prices does not mean rebuilding a spreadsheet. The data and currencies page documents that pipeline.
  • Robustness. Ask how the estimator behaves when a single bad print, a stock split that slipped through, or a regime change lands in the window. Sample estimators are the most sensitive; shrinkage and factor estimators are the least; downside estimators are sensitive in a different way, because a short window of losses can dominate the estimate.
  • Price and licensing. Building your own estimators is free in licence fees and expensive in everything else: a cleaned price feed, currency handling, a cache, and the maintenance when a provider changes a field name. RSJ Portfolio is a Windows desktop application under a single-user license — see pricing and license for what that includes.
  • Interpretability. You should be able to explain the resulting volatility number to whoever owns the money. A sample estimator’s output is easy to describe and hard to trust; a factor estimator is the reverse. Pick the one whose failure mode you can articulate.

Side-by-Side Comparison of Risk Models

The table below covers the eight estimator archetypes that MPT practice is built from. Read it as the vocabulary for the features page, which lists exactly which eight models RSJ Portfolio offers.

Model What it estimates Data requirements Strengths Weaknesses Use when
Sample covariance Full pairwise covariance from history Prices, decent history length Transparent, no assumptions, fast Noisy when candidates outnumber observations; outlier-sensitive Small candidate list, long clean history
Exponentially weighted Covariance with recency weighting Prices plus a decay parameter Adapts to regime shifts Extra parameter; shorter effective sample Recent behaviour should count more
Single-factor Covariance via beta to one market factor Prices plus a market index Few parameters, very stable Misses sector and idiosyncratic structure Many candidates, short history
Multi-factor Covariance from several factor exposures Prices plus factor returns Captures common drivers; robust Factor-set choice is model risk of its own Equity books with sector or style structure
Constant correlation Average pairwise correlation plus individual volatilities Prices Very stable, easy to reason about Flattens real differences in correlation Early exploration, noisy samples
Shrinkage Sample covariance pulled toward a structured target Prices Better conditioned, less estimation error Target choice; slower to react to genuine change Medium candidate lists; sensible default
Downside / semi-covariance Co-movement of negative returns only Prices; enough loss observations Matches loss-averse objectives Fewer observations; threshold choice Downside-focused objective
Robust (winsorized) Covariance after downweighting outliers Prices, usually with cleaning Resistant to bad prints and one-day shocks Heavier computation, tuning Messy or long histories with gaps

Two effects sit under that table. First, the risk model changes the level of estimated portfolio volatility — shrinkage and factor models usually report a smoother, sometimes lower figure than the raw sample matrix. Second, it changes the structure of correlation, which is what actually moves weights: constant correlation compresses dispersion in co-movement, while downside estimators can reveal that two stocks that look independent in calm markets fall together in selloffs.

Verdict: Which Risk Model Should You Pick?

There is no universally correct estimator, but there are sensible starting points by profile.

  • Conservative investors should start with a robust estimator — shrinkage or a factor model — and run it against minimum volatility. The goal is not the lowest backtested volatility; it is a weight vector that will not swing wildly when one month of data changes.
  • Quants should compare a factor model against a shrinkage model on the same candidate list. If the two disagree materially, the disagreement tells you where your portfolio’s risk is coming from, which is more useful than either answer alone.
  • Advisors and multi-account work should favour estimators that are stable across rebalancing dates. See Portfolio Optimization for Financial Advisors for that workflow, and Automate Portfolio Rebalancing With an Optimizer for the repeatability side.

A practical sequence: start with the sample covariance so you have a baseline you fully understand, then move to a shrinkage estimator and compare. Then vary one thing at a time — the estimator, then the objective, then the constraints. The getting started guide publishes recommended starting parameters for exactly this reason, and a candidate list that has been checked once should be checked again after the model changes: How to Verify Portfolio Optimization Results covers that step.

Putting Risk Models to Work in RSJ Portfolio

The workflow is deliberately short, because the interesting part should be the model choice, not the plumbing.

  1. Configure once. Pick the data provider, set the local currency, choose the date range and the cleaning settings. Prices are cleaned and cached locally, and portfolios quoted in several currencies are converted along a defined path.
  2. Select candidates. Search symbols and build the candidate list you actually want to allocate across.
  3. Choose the objective, the risk model and the weight constraints. All three are selectable in the optimizer.
  4. Run and review. The resulting allocation is the output; re-run with a different risk model to see how much of the allocation was driven by the estimator.
  5. Export the result for your own records or downstream reporting.

RSJ Portfolio is a managed application for portfolio optimization — you install the Windows build, work through the guided interface, and let the product own the data pipeline. It is not a CLI you script or a protocol you wire up by hand. The installer is published on the downloads page with a SHA-512 checksum in a versioned filename, and there is a user guide plus a short explainer video if you prefer to watch the workflow before running it.

FAQ

What are the eight risk models in RSJ Portfolio? RSJ Portfolio includes eight risk models, selectable in the optimizer. The exact names are listed on the features page and in the user guide; the models fall into the sample-based, factor-based, shrinkage and downside-focused families described above, and you can compare them directly inside the application.

How does the choice of risk model affect portfolio volatility? Different risk models estimate volatility and correlation differently. Sample-based models are sensitive to estimation error, while shrinkage and factor models aim to be more robust by accepting a little bias. The resulting portfolio volatility can vary meaningfully, so it is worth testing several models against your own candidate list rather than trusting a general recommendation.

Can I use RSJ Portfolio to compare risk models side by side? Yes. The risk estimator is a selectable parameter, so you can re-run the optimization with a different model and compare allocations and volatility estimates. The software is built for this kind of what-if analysis.

Is RSJ Portfolio suitable for individual investors or only quants? Both. It provides a guided workflow with recommended starting parameters, while also exposing the objective, the risk estimator and the weight constraints for users who want to fine-tune.

Conclusion

The optimizer gets the credit and the risk model does the work. When two allocations disagree, the productive question is not “which optimizer is right?” but “which covariance estimate am I trusting, and what does it assume about the world?” Comparing estimators on your own candidate list — rather than arguing about them — is the only way to answer that.

RSJ Portfolio is built for that comparison: eight selectable risk models, five objectives, weight constraints you control, and a data pipeline that handles cleaning, caching and currency conversion for you. Install it, load your candidate list, and run the same optimization twice with two different estimators. The difference between those two portfolios is the risk model decision, made visible. Start at portfolio.staging.rsj.de.

Related posts

EU label: AI-generated content