Graduation Year

2026

Date of Submission

12-2026

Document Type

Open Access Senior Thesis

Degree Name

Bachelor of Arts

Department

Economics

Reader 1

Benjamin Gillen

Terms of Use & License Information

Terms of Use for work posted in Scholarship@Claremont.

Abstract

This thesis evaluates whether dynamically tuned weight penalties can improve high-dimensional mean-variance portfolio choice. Classical Markowitz portfolios perform poorly when the number of assets is large relative to the data, because small estimation errors in means and covariances can generate extreme, fragile allocations. A common remedy is to regularize the weights using L1 (lasso), L2 (ridge), or elastic-net penalties, but these are typically chosen once and held fixed. Using monthly excess returns on the 48 Fama-French industry portfolios from January 2010 to June 2025, this thesis compares penalized mean-variance and minimum-variance strategies under long-only and long-short constraints, with and without Ledoit-Wolf covariance shrinkage. At each rebalance, a parametric bootstrap treats the estimated mean vector and covariance matrix as the data-generating process, simulates returns, and selects the penalty level that maximizes simulated utility. The results show that regularization is essential whenever shorting is allowed: unpenalized or weakly penalized long-short portfolios exhibit extreme leverage, turnover, and poor performance, while dynamic penalization with minimum penalties and gross-exposure caps avoids such failures. In long-only settings, by contrast, dynamic tuning offers only modest gains over well-chosen static penalties. For mean-variance long-only portfolios, lasso performs best by concentrating on a small set of high-mean industries, whereas in minimum-variance mandates ridge and elastic-net penalties undo over-concentration in low-volatility industries and raise Sharpe ratios, especially when combined with covariance shrinkage.

Download

Share

COinS