Viewing Mean-Variance Portfolio Weights as Asset Selection

Summary

The document asks whether portfolio optimization can be understood as a form of feature selection. It compares assets’ return series, represented as columns in a data matrix, with features in statistical learning. Under that analogy, mean-variance optimization chooses weights for the assets, while methods such as principal component analysis, ridge regression, LASSO, or support vector machines might supply alternative ways to select or weight them.

This is posed as a conceptual question rather than a worked method. It offers no portfolio construction procedure, empirical comparison, or evidence that substituting a machine-learning algorithm for mean-variance optimization yields useful allocations. The analogy also leaves open important distinctions: asset returns are investment outcomes, and portfolio weights must account for constraints, risk, estimation error, and trading costs. The document is useful as a prompt for connecting statistical learning and allocation, but does not resolve when the proposed framing is valid.

Key ideas

• The document frames each asset’s return history as a feature in a data matrix.
• Mean-variance optimization can be viewed as assigning weights to those asset features.
• It raises the possibility of using statistical learning methods to select or weight assets.
• The analogy is exploratory and gives no evidence that these methods can replace portfolio optimization directly.

Tags

# Isn't portfolio optimization basically just feature selection?


# Isn't portfolio optimization basically just feature selection?












Statistical learning has a large assortment of tools for conducting feature selection such as PCA analysis, ridge regression, LASSO, SVM and almost every other machine learning algorithm.

In portfolio optimization, portfolio weights are optimized based on data consisting only of asset return time series, $x$, which line the matrix $X$.

If each asset (each column of $X$) can be thought of as features, then isn't the mean-variance model just a supervised learning model that weights the individual assets based on feature selection philosophy? meaning that pretty much any learning algorithm can be replaced into the traditional model in order to conduct feature selection (the fractional choosing of assets according to each algorithm's own unique criteria)?

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.