Understanding the Long-Only Minimum Variance Portfolio

This paper investigates the relationship between long-only global minimum variance portfolios and asset factor exposures, providing explicit analytical solutions for single-factor models and geometric descriptions for multi-factor scenarios, validated through empirical US stock data.

The Gist

Imagine you are the captain of a ship trying to navigate through a stormy sea. Your goal is simple: reach your destination with the least amount of shaking possible. In the world of finance, this "shaking" is called volatility or variance. The less the ship shakes, the safer the journey.

This paper is a guide for captains (investors) who want to build the smoothest possible portfolio of stocks, but with one very strict rule: You can only buy things; you cannot bet against them.

Here is the breakdown of the paper's big ideas, translated into everyday language.

1. The Two Types of Sailors: The "Short-Seller" vs. The "Long-Only"

In the financial world, there are two ways to build a portfolio:

• The Long-Short Sailor (The Unrestricted Captain): This sailor can buy stocks (bet they go up) and short stocks (bet they go down). If the market crashes, they can make money by betting against it. Mathematically, this is easy to solve. It's like having a magic compass that points exactly where the calm water is, even if you have to sail backward to get there.
• The Long-Only Sailor (The Real-World Captain): This is the paper's focus. Most real investors (like pension funds or regular people) are not allowed to short sell. They can only hold positive amounts of stocks. If a stock is dangerous, they just have to leave it out of the boat. They can't bet against it.

The Problem: Because the Long-Only sailor can't use the "bet against" trick, finding the smoothest ride is much harder. It's like trying to find the flattest path through a mountain range, but you are only allowed to walk up or across, never down.

2. The "Factor" Map: Why Stocks Move Together

The authors realize that stocks don't move randomly. They move because of hidden forces, which they call Factors.

• The 1-Factor Model (The Single Giant Wave): Imagine the ocean is dominated by one giant wave (the "Market"). When this wave rises, almost all boats rise. When it falls, they all fall. Some boats are light and bounce a lot (high sensitivity), while others are heavy and barely move (low sensitivity).
• The Multi-Factor Model (The Storm System): Sometimes, it's not just one wave. Maybe there's a wind factor, a current factor, and a tide factor all pushing at once.

The paper asks: If we know how each stock reacts to these waves (their "exposure" or "beta"), can we figure out exactly which stocks to keep in the portfolio to minimize the shaking?

3. The Magic Formula for One Wave (The 1-Factor Case)

When there is only one main force driving the market, the authors found a beautiful, simple rule.

The Analogy: The "Beta" Threshold
Imagine you have a list of 1,000 boats. Each boat has a "sensitivity score" (Beta) telling you how much it bounces with the giant wave.

• Some boats are very sensitive (High Beta).
• Some are calm (Low Beta).
• Some are weird and move opposite to the wave (Negative Beta).

The paper proves that for the smoothest ride, you should only keep the boats with the lowest sensitivity scores, up to a certain cutoff point.

• The Rule: You keep the calmest boats. You kick out the wild ones.
• The Twist: You don't just pick the top 10% randomly. There is a specific mathematical "cutoff line." If a boat's sensitivity is below the line, it's in. If it's above, it's out.
• The Result: You end up with a portfolio of only the "calmest" stocks. The wild ones are left on the dock.

The paper gives a step-by-step recipe to find exactly where that cutoff line is, so you don't have to guess.

4. The Complex Storm (The Multi-Factor Case)

What if there are two or more forces (like wind AND current)? The math gets messy. You can't just sort them by a single number anymore.

The Analogy: The Invisible Wall
In this scenario, the authors describe the solution using geometry. Imagine the stocks are dots floating in a 3D room (or a 2D room if there are 2 factors).

• The "Active" stocks (the ones you keep) are all clustered on one side of an invisible wall.
• The "Inactive" stocks (the ones you drop) are on the other side.

The paper proves that this "wall" (called a hyperplane) exists. It separates the good candidates from the bad ones based on how they react to the combination of factors. While you can't write a simple "if-then" formula like in the single-factor case, you know that the solution is defined by this geometric boundary.

5. The Real-World Test

The authors didn't just do math on paper; they tested it on the top 1,000 US stocks.

• They used data from 2022.
• They applied their formulas to find the "Long-Only Minimum Variance" portfolio.
• The Surprise: Out of 1,000 stocks, the optimal portfolio only kept about 50 to 65 stocks.
• The Lesson: To get the smoothest ride, you don't need a huge boat with 1,000 passengers. You need a small, elite crew of the calmest, most stable stocks. The rest are just noise that adds shaking.

Summary: What Should You Take Away?

1. Constraints Matter: Being forced to only "buy" (Long-Only) changes the math completely compared to being able to "short" (bet against).
2. Simplicity in Chaos: Even in a complex market, if you understand the main drivers (factors), you can find a simple rule to pick the best stocks.
3. Less is More: The safest portfolio isn't the one with the most stocks. It's the one with the right stocks—specifically, the ones that are least sensitive to the market's wild swings.
4. The "Cutoff": There is a specific threshold. Stocks below it are safe; stocks above it are too risky for a "minimum variance" strategy.

In a nutshell: This paper is a map that tells investors how to filter out the noisy, volatile stocks and build a tiny, ultra-stable portfolio by understanding how stocks react to the market's underlying waves.

Technical Summary

Here is a detailed technical summary of the paper "Understanding the Long-Only Minimum Variance Portfolio" by Gunther, Kercheval, and Sowunmi.

1. Problem Statement

The paper addresses the Long-Only Minimum Variance (LOMV) portfolio optimization problem. While the Global Minimum Variance (GMV) portfolio is a standard benchmark in finance, the unconstrained solution (allowing short sales) often yields portfolios with extreme weights and high turnover. Real-world investment constraints typically require long-only positions ($w_i \geq 0$).

The core mathematical challenge is that the LOMV problem is a quadratic programming problem with inequality constraints. Unlike the unconstrained GMV, which has a closed-form solution ($w_{LS} = \frac{\Sigma^{-1}\mathbf{1}}{\mathbf{1}^T\Sigma^{-1}\mathbf{1}}$), the long-only solution depends on identifying the set of active assets (those with strictly positive weights). Determining this set $K$ is computationally difficult for general covariance matrices $\Sigma$, as it requires solving a combinatorial problem to find which constraints are binding.

The authors investigate this problem specifically when the covariance matrix $\Sigma$ is derived from a factor model, aiming to provide explicit or geometric characterizations of the active set $K$.

2. Methodology and Theoretical Framework

The authors analyze the LOMV problem under two distinct factor model structures: a Single-Factor Model and a Multi-Factor Model ($q > 1$).

A. General Framework (Theorem 1)

For any positive definite covariance matrix $\Sigma$, the authors establish that once the set of active assets $K = \{i : w_i > 0\}$ is known, the solution is simply the unconstrained GMV solution restricted to the sub-matrix $\Sigma_K$ (the principal submatrix of $\Sigma$ corresponding to assets in $K$).
$$w_K = \frac{\Sigma_K^{-1}\mathbf{1}_k}{\mathbf{1}_k^T \Sigma_K^{-1} \mathbf{1}_k}$$
The primary difficulty remains determining the set $K$.

B. Single-Factor Model (Theorem 2)

Assuming a single-factor model where $\Sigma = \sigma^2 \beta \beta^T + \Delta$ (with $\beta$ as factor exposures and $\Delta$ as diagonal idiosyncratic variances), the authors derive an explicit, closed-form solution for the active set $K$.

• Key Insight: If assets are ordered such that their factor exposures $\beta_i$ are non-decreasing, the set of active assets $K$ in the optimal long-only portfolio consists of the first $k$ assets (an initial segment of the sorted list).
• The Sequence $R_i$: The authors define a sequence $R_i$ based on the model parameters:
  $$R_1 = \frac{1}{\sigma^2}, \quad R_i = \frac{1}{\sigma^2} + \sum_{j=1}^{i-1} \frac{\beta_j}{\delta_j^2}(\beta_j - \beta_i) \quad \text{for } i \geq 2$$
• Threshold Rule: The sequence $\{R_i\}$ is unimodal (increases then decreases). The number of active assets $k$ is the largest index such that $R_k > 0$.
  $$K = \{1, 2, \dots, k\} \quad \text{where } k = \max \{i : R_i > 0\}$$
• Condition: An asset $i$ is included if and only if its beta is below a specific threshold determined by the aggregate properties of the active set. This generalizes and provides an explicit calculation for the implicit solution found in prior literature (Clarke et al., 2011).

C. Multi-Factor Model (Theorem 3)

For $q > 1$ factors, where $\Sigma = B\Omega B^T + \Delta$, a simple sorting of assets is impossible because exposures are vectors in $\mathbb{R}^q$.

• Geometric Characterization: The authors prove that the active assets $K$ are those whose factor exposure vectors $B_i$ lie on the same side of a separating hyperplane $H$ as the origin.
• Hyperplane Definition: The hyperplane is defined by $H = \{x \in \mathbb{R}^q : x^T h_K = 1\}$, where $h_K = \Omega B_K^T (\Sigma_K)^{-1} \mathbf{1}_k$.
• Implication: An asset $i$ is active ($w_i > 0$) if and only if $B_i h_K < 1$. This provides a necessary condition that reduces the search space from a combinatorial problem to a geometric separation problem, though it does not yield a direct closed-form algorithm for $K$ like the single-factor case.

3. Empirical Results and Numerical Examples

The paper validates the theoretical findings using daily returns of the top 1,000 US stocks (Jan–July 2022).

A. Covariance Estimation

Since $p > n$ (1000 assets vs. ~126 days), the sample covariance matrix is singular. The authors compare three estimators:

1. $\Sigma_{JSE}$: A James-Stein Eigenvector shrinkage estimator (statistical factor model).
2. $\Sigma_{MJSE}$: A single-index market model derived from $\Sigma_{JSE}$.
3. $\Sigma_{MS}$: A single-index market model derived from the sample covariance matrix (standard approach).

B. Single-Factor Findings

• Sparsity: The LOMV portfolios are highly sparse. Out of 1,000 assets, only 51 to 66 were active (positive weights).
• Selection Criteria: Active assets generally have lower betas and lower idiosyncratic risk compared to the full universe.
• Consistency: The statistical factor model ($\Sigma_{JSE}$) and the market model derived from it ($\Sigma_{MJSE}$) produced nearly identical active sets and weights, whereas the standard market model ($\Sigma_{MS}$) differed slightly.
• Efficiency: The closed-form solution (Theorem 2) allows for rapid computation ($O(\log p)$ via bisection) compared to iterative quadratic programming.

C. Two-Factor Findings

• Hyperplane Separation: Using a 2-factor model, the authors visualized the factor exposures in 2D space.
• Result: The active assets (orange dots) were clearly separated from inactive assets (blue dots) by a linear hyperplane (red line), confirming Theorem 3.
• Comparison: The 2-factor model selected 41 assets, significantly fewer than the 65 selected by the 1-factor model. The 2-factor model included assets with negative exposure to the second factor that compensated for high exposure to the first, demonstrating how multi-factor models can alter the "efficient frontier" of active assets.

4. Key Contributions

1. Explicit Solution for 1-Factor LOMV: The paper provides the first rigorous, explicit description of the active set $K$ for a long-only minimum variance portfolio under a single-factor model, removing the need for iterative numerical optimization in this specific case.
2. Geometric Insight for Multi-Factor: It establishes a necessary geometric condition (hyperplane separation) for the active set in multi-factor models, offering a new perspective on how factor exposures interact to determine portfolio composition.
3. Algorithmic Efficiency: By reducing the problem to finding a threshold index $k$ (for 1-factor) or a separating hyperplane (for $q$-factor), the methodology offers a more computationally efficient alternative to general convex optimization solvers for large universes.
4. Estimator Consistency: The paper demonstrates that James-Stein eigenvector shrinkage estimators ($\Sigma_{JSE}$) are asymptotically consistent with market-based factor models, validating the use of statistical factor models for portfolio construction.

5. Significance

This work bridges the gap between theoretical portfolio optimization and practical implementation constraints.

• Practicality: Long-only constraints are ubiquitous in institutional investing. Understanding the structure of the solution helps practitioners anticipate which assets will be held and why (e.g., low beta, low specific risk).
• Interpretability: The results demystify the "black box" nature of constrained optimization. In a single-factor world, the portfolio is simply a selection of assets with the lowest betas (adjusted for idiosyncratic risk).
• Robustness: The empirical results suggest that using robust covariance estimators (like JSE) leads to stable and sparse portfolios, reducing turnover and transaction costs.
• Future Directions: The geometric characterization of the multi-factor case suggests that dimensionality reduction or specific factor selection could be used to approximate the active set without solving the full optimization problem.

In summary, the paper transforms the LOMV problem from a generic quadratic program into a structured selection problem driven by factor exposures, providing both theoretical clarity and practical computational advantages.