Covariance Estimation Methods for Portfolio Risk
Summary
This document surveys methods for estimating and adjusting covariance matrices used in portfolio risk analysis. It covers the empirical estimator, robust Minimum Covariance Determinant, basic and data-driven shrinkage methods, semi-covariance, exponentially weighted covariance, and spectral denoising and detoning. The methods address different weaknesses: outlier sensitivity, unstable eigenvalues or matrix inversion, downside-focused risk measurement, changing relevance of observations, and noise in estimated correlations. It also introduces conversion between covariance and correlation matrices.
An example applies several estimators to a small set of ETF price histories and compares the resulting matrices. The discussion explains that empirical covariance is most suitable when there are enough observations relative to the number of assets, while robust estimation is useful when outliers matter and shrinkage can improve conditioning. Semi-covariance focuses on returns below a chosen threshold. These methods serve different purposes, so the article does not identify one estimator as universally best. Its comparisons are illustrative, based on a limited dataset, and do not establish which estimator leads to better portfolio outcomes.
Key ideas
- Empirical covariance can be sensitive to outliers and unstable when observations are limited relative to assets.
- Minimum Covariance Determinant offers a robust alternative when outliers are a concern.
- Shrinkage can improve covariance matrix conditioning and support more stable inversion.
- Semi-covariance focuses on returns below a chosen threshold to represent downside risk.
- Exponential weighting and spectral denoising address recency and noise, respectively.
- Estimator choice should reflect the risk objective and data characteristics.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.