Markowitz Portfolio Construction at Seventy

Stephen Boyd, Kasper Johansson, Ronald N. Kahn, Philipp Schiele, Thomas M. SchmelzerSSRN 2024 · DOI

Abstract

Since its inception in 1952, Harry Markowitz’s Mean-Variance Optimization (MVO) has served as the bedrock of modern portfolio theory. However, practical implementation has long been hindered by the model's sensitivity to input errors and its inability to handle real-world frictions. This paper introduces Markowitz++, a modernized framework that leverages robust optimization, soft constraints, and modern convex solvers to transform MVO into a high-performance trading engine. We demonstrate that Markowitz++ significantly reduces drawdowns and scales to handle tens of thousands of assets, bridging the gap between theoretical elegance and institutional requirements.

1. Introduction

In 1952, Harry Markowitz revolutionized finance by framing portfolio construction as a mathematical optimization problem: maximizing expected return for a given level of risk. While theoretically profound, "Basic Markowitz" has often been dismissed by practitioners as a "noise maximizer." Small errors in return forecasts frequently lead to erratic portfolios characterized by extreme weights, high turnover, and unrealistic leverage. Furthermore, the original theory largely ignored the "friction" of the real world—transaction costs, market impact, and the potential for hard constraints to render problems unsolvable during periods of market stress.

2. The Markowitz++ Framework

To address these historical shortcomings, we propose Markowitz++. This framework remains within the computationally efficient realm of convex optimization but introduces several layers of sophistication designed for the modern quantitative trading environment.

2.1 Robust Optimization

Rather than assuming return forecasts are precise, Markowitz++ incorporates uncertainty intervals. By solving for the "worst-case" scenario within these intervals, the optimizer effectively "tames" the model's sensitivity to noise, preventing the extreme allocations typical of traditional MVO.

2.2 Soft Constraints

A common failure point in automated trading is the "infeasible" solution—where market movements make it impossible to satisfy hard constraints (e.g., strict leverage or turnover limits). Markowitz++ replaces these "hard walls" with soft penalties in the objective function. This ensures the solver always returns a feasible "least-bad" move, maintaining operational continuity during volatile regimes.

2.3 Factor Models and Scalability

To manage large-scale universes (10,000+ assets), the framework utilizes factor models to decompose risk. This reduction in dimensionality shifts the computational complexity from $O(n^3)$ to $O(nk^2)$, where $n$ is the number of assets and $k$ is the number of factors.

2.4 Implementation via CVXPY

The framework is implemented using CVXPY, a high-level domain-specific language for optimization. This allows complex financial logic—such as transaction cost modeling and multi-factor risk constraints—to be expressed in concise, readable Python code, facilitating rapid iteration and transparency.

3. Experimental Methodology

The performance of Markowitz++ was evaluated using a combination of historical and synthetic data:

  • S&P 100 Data: Daily adjusted close prices from 2000 to 2023, covering 74 assets that survived the full period.
  • Synthetic Alpha: To simulate realistic trading conditions without relying on proprietary signals, the authors generated synthetic signals with a realistic information coefficient (correlation to actual returns).
  • Market Frictions: The model accounted for borrowing costs (Federal funds rates) and transaction costs (bid-ask spreads).

4. Key Results

The transition from "Basic" to "Modern" optimization yields staggering improvements in portfolio stability and performance.

4.1 Risk Mitigation

In back-tests, the "Basic Markowitz" model suffered a catastrophic 78.9% maximum drawdown. In contrast, the Markowitz++ version, equipped with weight limits and robust optimization, achieved superior returns with a maximum drawdown of only 7.0%.

4.2 Computational Efficiency

Using the MOSEK solver, Markowitz++ demonstrates industrial-grade scalability. A 10,000-asset problem is solved in under one second. Even a massive 50,000-asset universe—exceeding the size of the global investable equity market—is processed in approximately 18 seconds.

4.3 Hyperparameter Optimization

The study highlights the importance of "tuning" the optimizer. By adjusting the weights assigned to turnover versus risk based on historical back-tests, the Sharpe ratio was improved from 4.32 to 4.65.

5. Discussion and Limitations

While Markowitz++ represents a significant leap forward, certain limitations remain. The reliance on convex optimization means that certain non-convex constraints (like minimum trade sizes) must be handled via heuristics. Additionally, while the framework is robust to noise, it remains dependent on the quality of the underlying "alpha" signal; it ensures a forecast is not ruined by poor construction, but it cannot generate returns from a signal with no predictive power. Finally, the use of a survivorship-biased S&P 100 dataset suggests that while relative performance gains are robust, absolute return figures should be interpreted with caution.

6. Conclusion

Seventy years after its introduction, Markowitz’s vision of mathematical portfolio construction is more relevant than ever. By integrating modern computational tools and robust optimization techniques, Markowitz++ transforms a historically brittle theory into a resilient, scalable, and practical engine for modern finance.

Summary updated

BibTeX
@misc{boyd2024markowitz,
  author = {Stephen Boyd and Kasper Johansson and Ronald N. Kahn and Philipp Schiele and Thomas M. Schmelzer},
  title = {{Markowitz Portfolio Construction at Seventy}},
  year = {2024},
  eprint = {2401.05080},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2401.05080},
  howpublished = {SSRN},
}
Figure