Third-Party Research & Methodology Only

This section shares summaries of third-party academic research and descriptions of quantitative models. The content represents the findings of the original researchers, not the opinions or recommendations of Foxholm Financial. Foxholm Financial does not publish hypothetical or backtested performance metrics on its quantitative research pages. All content is restricted to methodology, signal construction, factor logic, and risk architecture. SEC rules require that investment advisers not present misleading performance data, and our methodology-only approach reflects that standard and the firm's fiduciary obligations.

Mean-Variance Optimization

Portfolio Construction Optimization Asset Allocation

Robert Stowe, AAMS® Investment Advisor

Mean-variance optimization (MVO) is the mathematical framework for building portfolios that offer the highest expected return for a given level of risk. Developed by Harry Markowitz in 1952, it is the foundation of modern portfolio theory and the starting point for nearly every quantitative approach to asset allocation.

The core insight is that an investor should not evaluate each investment in isolation. What matters is how investments interact within a portfolio. Two volatile assets can create a stable portfolio if their price movements offset each other. MVO formalizes this idea by finding the combinations of assets that produce the best trade-off between expected return and risk (measured as the variance, or spread, of returns).

Conceptual Framework

Markowitz published "Portfolio Selection" in The Journal of Finance in 1952, introducing the idea that diversification can be quantified mathematically. Before this work, diversification was understood intuitively ("don't put all your eggs in one basket") but there was no framework for measuring exactly how much diversification a portfolio achieved or how to improve it systematically.

The framework requires three inputs for each asset in the portfolio: its expected return, its risk (measured as the standard deviation of returns), and its correlation with every other asset. With these inputs, the optimizer calculates the portfolio weights that produce the highest expected return at each level of risk. The collection of these mathematically efficient portfolios forms the "efficient frontier," a curve that represents the best possible trade-offs available to the investor.

Core Assumptions

MVO is built on several assumptions about investor behavior and market returns. Understanding these assumptions, and where they break down, is essential to using the model effectively:

Investors care only about return and risk: The model assumes that the only things an investor needs to know about a portfolio are its expected return and its variance. This ignores other properties investors may care about, such as the chance of extreme losses (tail risk), how easy it is to sell the investments (liquidity), or the potential for very large gains (skewness).
Returns follow a bell curve: MVO assumes that investment returns are normally distributed (bell-shaped), which makes variance a complete description of risk. In reality, financial returns have fatter tails than a bell curve predicts, meaning extreme losses occur more often than the model expects.
Inputs are known with certainty: The optimizer treats expected returns, volatilities, and correlations as if they were exact numbers. In practice, these are estimates based on historical data or forecasts, and they come with substantial uncertainty. Small changes in the inputs can produce dramatically different portfolio weights.
Single-period framework: The standard model optimizes for a single time period. It does not account for the fact that investors rebalance, face taxes on realized gains, or have different needs at different life stages.
No transaction costs or taxes: The model assumes that buying and selling are free. Adding constraints for transaction costs and tax efficiency requires extensions to the basic framework.

Optimization Process

Building a mean-variance optimized portfolio follows a structured process from gathering inputs to selecting a portfolio on the efficient frontier.

Step 1

Estimate Inputs

Step 2

Build Covariance Matrix

Step 3

Run Optimizer

Step 4

Plot Efficient Frontier

Step 5

Select Portfolio

Input Estimation

The optimizer needs expected returns, volatilities, and correlations for every asset in the universe. Expected returns are the hardest to estimate and the most influential on the final result. Historical averages are commonly used, but past returns are a poor predictor of future returns. More sophisticated approaches use equilibrium models (like the Capital Asset Pricing Model), factor-based forecasts, or the implied returns from the Black-Litterman framework.

Volatilities and correlations are more stable than returns and can be estimated more reliably from historical data. However, correlations between assets tend to increase during market stress, exactly when diversification is needed most. Using calm-period correlations in the optimizer overstates the diversification benefit during crises.

The Covariance Matrix

The covariance matrix is a table that describes how every pair of assets in the portfolio moves in relation to each other. For a portfolio with 10 assets, the matrix contains 55 unique relationships (10 volatilities plus 45 pairwise correlations). For 100 assets, it grows to 5,050 relationships. As the number of assets increases, estimating the matrix reliably becomes increasingly difficult because each estimate introduces potential error.

Practitioners use several techniques to make the covariance matrix more reliable. Shrinkage methods blend the sample covariance matrix with a structured target (like the identity matrix or a single-factor model) to reduce estimation noise. Factor-based covariance matrices express asset correlations through a smaller number of common factors (like market, size, and value), which requires fewer parameters and produces more stable estimates.

The Efficient Frontier

The optimizer solves a mathematical problem: for each target level of risk, find the portfolio weights that maximize expected return (subject to constraints like "no short selling" or "at least 20% in bonds"). The set of solutions traces out the efficient frontier, a curve on a risk-return chart. Every portfolio on the frontier is optimal in the sense that no other portfolio offers higher expected return at the same risk level.

Portfolios below the frontier are inefficient: they take on more risk than necessary for the return they provide, or deliver less return than possible for the risk they bear. The investor then selects a point on the frontier that matches their risk tolerance, either by specifying a maximum acceptable risk level or by maximizing a risk-adjusted measure like the Sharpe ratio (the ratio of excess return to risk).

Risk Architecture

MVO is both a risk management tool and a source of model risk. Its outputs can look mathematically optimal but behave poorly in practice when the inputs are inaccurate.

Model Risk

The single biggest practical problem with MVO is its extreme sensitivity to the expected return inputs. Small changes in expected returns (even within the margin of estimation error) can cause the optimizer to swing from a 60% allocation in one asset to a 0% allocation. This behavior has been described as "an error maximizer": the optimizer concentrates the portfolio in exactly the assets whose expected returns happen to be most overestimated.

This sensitivity creates a paradox. The model is most aggressive in its allocations precisely where the inputs are least reliable. Without constraints or regularization techniques, unconstrained MVO frequently produces portfolios that are heavily concentrated in a small number of assets and perform poorly out of sample.

Known Limitations

Limitations to Consider

Input sensitivity: The optimizer amplifies errors in expected return estimates. A 1% change in a single asset's expected return can shift its portfolio weight by 20% or more.
Concentrated portfolios: Without constraints, MVO often recommends putting most of the portfolio in just a few assets. This looks optimal on paper but is fragile in practice because the concentration is driven by estimation errors.
Unstable weights over time: Re-running the optimization each month with updated data produces wildly different portfolios, leading to excessive turnover and high transaction costs.
Ignores tail risk: Because it uses variance as the sole risk measure, MVO treats upside volatility and downside volatility equally. An investor who fears large losses more than they value large gains needs a different risk measure.
Single-period limitation: The framework does not naturally accommodate multi-period goals like retirement drawdown plans, where the sequence of returns matters as much as the average.

Practical Considerations

Taming the Optimizer

Decades of practical experience have produced techniques to make MVO more useful. The most common approaches include:

Weight constraints: Setting minimum and maximum allocations for each asset class (e.g., "between 5% and 40% in any single asset"). This prevents extreme concentrations but reduces the optimizer's freedom and may move the portfolio away from the unconstrained mathematical solution.
Resampled optimization: Running the optimization thousands of times with slightly different inputs (drawn from the range of estimation uncertainty) and averaging the resulting weights. This produces more diversified, more stable portfolios.
Black-Litterman approach: Starting from market-implied expected returns rather than historical estimates, then adjusting for the investor's specific views. This produces a more stable starting point and avoids the extreme sensitivity to return estimates.
Shrinkage estimators: Using statistical techniques to pull extreme estimates toward more reasonable central values. The Ledoit-Wolf shrinkage method for covariance matrices is one widely used example.

MVO vs. Alternative Approaches

MVO is the original portfolio optimizer, but several alternatives have been developed to address its weaknesses. Risk parity ignores expected returns entirely and allocates so that each asset contributes equally to portfolio risk. The minimum-variance portfolio uses only the covariance matrix, removing the most error-prone input (expected returns). The 1/N portfolio (equal weighting) ignores all optimization inputs entirely. Research by DeMiguel, Garlappi, and Uppal (2009) showed that the simple 1/N approach often matched or outperformed optimized portfolios out of sample, highlighting how much damage estimation error can do.

When MVO Is Most Useful

Despite its limitations, MVO remains valuable as a framework for thinking about portfolio trade-offs. It is most useful when the number of assets is small (reducing estimation error), when expected return estimates are grounded in economic logic rather than historical extrapolation, and when constraints or regularization are applied to prevent extreme allocations. It is least useful when applied mechanically to a large universe of assets with historical average returns as inputs.