# Portfolio Optimization Rules beyond the Mean-Variance Approach

**Authors:** Maxime Markov, Vladimir Markov  
**arXiv:** [2305.08530](https://arxiv.org/abs/2305.08530)  
**Date:** 2023-05-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.08530v1/

## Key points

- Authors address limitations of traditional portfolio allocation rules
- Propose models to enhance asset allocation strategies
- Focus on incorporating uncertainty in expected returns and improving stability of covariance matrix estimation
- Analyze historical returns of major stock indexes and find asymmetric Laplace distribution (ALD) provides reasonable fit
- Develop allocation rules specifically tailored for ALD returns
- Compare ALD-based allocation rules with traditional mean-variance approach, revealing significant skewness term in situations where µ/σ is small
- Fat tails in return distributions impose cutoff on risk by saturating allocation weights for stocks with high expected returns
- Derive allocation rules for scenarios where expected return µ is a Gaussian random variable with uncertainty σ0 much larger than µ0
- In average scenario, allocation rule resembles mean-variance solution with regularization as σ0 approaches zero
- In worst-case scenario, optimal solution smoothly approximates between equal weights and minimum variance portfolio, providing convex alternative to risk parity portfolios
- Explore log-normal distributed returns as base model for long-term cross-sectional returns
- Investigate how model parameters influence portfolio construction using microscopic portfolio model with random drift and analytical expression for expected utility function
- Propose incorporating block structure constraints on precision matrix estimation to address issues related to singular covariance matrices
- Enhance allocation weight stability by introducing conditional independence between inter-block constituents
- Mitigate sampling errors of large covariance matrices
- Proposed framework and methods anticipated to be valuable for practitioners in portfolio construction

## Summary

In this paper, the authors address the limitations of traditional portfolio allocation rules and propose a set of models to enhance asset allocation strategies. They focus on incorporating uncertainty in expected returns and improving the stability of covariance matrix estimation. To begin, the authors analyze historical returns of major stock indexes on different time scales and find that the asymmetric Laplace distribution (ALD) provides a reasonable fit. They derive empirical parameters for location (µ), scale (σ), and skewness (κ) as well as their scaling laws. This allows them to develop allocation rules specifically tailored for ALD returns. Comparing these ALD-based allocation rules with the traditional mean-variance approach, the authors reveal that in situations where µ/σ is small, the Markowitz contribution is accompanied by a significant skewness term. Additionally, they find that fat tails in return distributions impose a cutoff on risk by saturating allocation weights for stocks with high expected returns. Next, the authors consider scenarios where the expected return µ is a Gaussian random variable with uncertainty σ0 much larger than µ0. They derive allocation rules for both average and worst-case scenarios. In the average scenario, they demonstrate that as σ0 approaches zero, the allocation rule resembles a mean-variance solution with regularization, where σ0 acts as a regularizer or shrinkage factor. In contrast, in the worst-case scenario, they find an optimal solution that smoothly approximates between equal weights and a minimum variance portfolio, providing an attractive convex alternative to risk parity portfolios. The authors also explore log-normal distributed returns as a base model for long-term cross-sectional returns. They investigate how model parameters influence portfolio construction using a microscopic portfolio model with random drift and an analytical expression for expected utility function. Finally, addressing issues related to singular covariance matrices, the authors propose incorporating block structure constraints on precision matrix estimation. This approach enhances allocation weight stability by introducing conditional independence between inter-block constituents and mitigates sampling errors of large covariance matrices. Overall, the authors anticipate that their proposed framework and methods will be valuable for practitioners in portfolio construction; potentially improving or superseding traditional mean-variance and risk parity approaches by providing more accurate estimates of expected returns while accounting for uncertainty in market conditions such as fat tails or singular covariance matrices which can lead to unstable allocations when not properly addressed.