Portfolio optimization has been a fundamental problem in financial economics since Harry Markowitz laid the foundations of modern portfolio theory with his pioneering work on portfolio selection. Despite its theoretical relevance, the practical implementation of the mean-variance framework is affected by estimation uncertainty, which can lead to unstable allocations, extreme positions, and poor performance. This thesis provides a comparative empirical analysis of portfolio optimization methodolo- gies aimed at addressing these limitation. The analysis begins with the classical Markowitz framework, considering Global Minimum Variance and Maximum Trade-Off portfolios. Unconstrained portfolios are compared with long-only formulations to assess the effect of short-selling restrictions. The analysis then introduces regularization through the Least Absolute Shrinkage and Selection Operator. By reformulating the Markowitz problem within a linear regression framework, Lasso is employed to construct portfolios under different gross-exposure constraints. Finally, the framework is extended to Conditional Value-at-Risk (CVaR) optimiza- tion to explicitly account for downside tail risk. The empirical analysis considers a universe of 100 stocks selected from the constituents of the S&P 500 Index over the period from September 2004 to December 2025. The implemented strategies are evaluated through a comprehensive back-test framework using multiple performance and risk indicators to assess their empirical performance. The results show that unconstrained Markowitz portfolios are highly sensitive to estimation error and can generate extreme exposures and severe losses. Long-only constraints improve stability, although not necessarily risk-adjusted performance. Lasso-based portfolios provide a more effective balance between sparsity, diversification, and performance, while CVaR optimization improves downside-risk control without consistently producing superior overall performance. Overall, the findings highlight the limitations of unconstrained mean-variance optimization and demonstrate the superior out-of-sample performance of Lasso-based portfolios, suggesting that regularization provides a more effective approach to portfolio construction.

Portfolio optimization has been a fundamental problem in financial economics since Harry Markowitz laid the foundations of modern portfolio theory with his pioneering work on portfolio selection. Despite its theoretical relevance, the practical implementation of the mean-variance framework is affected by estimation uncertainty, which can lead to unstable allocations, extreme positions, and poor performance. This thesis provides a comparative empirical analysis of portfolio optimization methodolo- gies aimed at addressing these limitation. The analysis begins with the classical Markowitz framework, considering Global Minimum Variance and Maximum Trade-Off portfolios. Unconstrained portfolios are compared with long-only formulations to assess the effect of short-selling restrictions. The analysis then introduces regularization through the Least Absolute Shrinkage and Selection Operator. By reformulating the Markowitz problem within a linear regression framework, Lasso is employed to construct portfolios under different gross-exposure constraints. Finally, the framework is extended to Conditional Value-at-Risk (CVaR) optimization to explicitly account for downside tail risk. The empirical analysis considers a universe of 100 stocks selected from the constituents of the S&P 500 Index over the period from September 2004 to December 2025. The implemented strategies are evaluated through a comprehensive back-test framework using multiple performance and risk indicators to assess their empirical performance. The results show that unconstrained Markowitz portfolios are highly sensitive to estimation error and can generate extreme exposures and severe losses. Long-only constraints improve stability, although not necessarily risk-adjusted performance. Lasso-based portfolios provide a more effective balance between sparsity, diversification, and performance, while CVaR optimization improves downside-risk control without consistently producing superior overall performance. Overall, the findings highlight the limitations of unconstrained mean-variance optimization and demonstrate the superior out-of-sample performance of Lasso-based portfolios, suggesting that regularization provides a more effective approach to portfolio construction.

Sparse regression for portfolio optimization : implementation and evaluation

GNOATO, FILIPPO
2025/2026

Abstract

Portfolio optimization has been a fundamental problem in financial economics since Harry Markowitz laid the foundations of modern portfolio theory with his pioneering work on portfolio selection. Despite its theoretical relevance, the practical implementation of the mean-variance framework is affected by estimation uncertainty, which can lead to unstable allocations, extreme positions, and poor performance. This thesis provides a comparative empirical analysis of portfolio optimization methodolo- gies aimed at addressing these limitation. The analysis begins with the classical Markowitz framework, considering Global Minimum Variance and Maximum Trade-Off portfolios. Unconstrained portfolios are compared with long-only formulations to assess the effect of short-selling restrictions. The analysis then introduces regularization through the Least Absolute Shrinkage and Selection Operator. By reformulating the Markowitz problem within a linear regression framework, Lasso is employed to construct portfolios under different gross-exposure constraints. Finally, the framework is extended to Conditional Value-at-Risk (CVaR) optimiza- tion to explicitly account for downside tail risk. The empirical analysis considers a universe of 100 stocks selected from the constituents of the S&P 500 Index over the period from September 2004 to December 2025. The implemented strategies are evaluated through a comprehensive back-test framework using multiple performance and risk indicators to assess their empirical performance. The results show that unconstrained Markowitz portfolios are highly sensitive to estimation error and can generate extreme exposures and severe losses. Long-only constraints improve stability, although not necessarily risk-adjusted performance. Lasso-based portfolios provide a more effective balance between sparsity, diversification, and performance, while CVaR optimization improves downside-risk control without consistently producing superior overall performance. Overall, the findings highlight the limitations of unconstrained mean-variance optimization and demonstrate the superior out-of-sample performance of Lasso-based portfolios, suggesting that regularization provides a more effective approach to portfolio construction.
2025
Sparse regression for portfolio optimization : implementation and evaluation
Portfolio optimization has been a fundamental problem in financial economics since Harry Markowitz laid the foundations of modern portfolio theory with his pioneering work on portfolio selection. Despite its theoretical relevance, the practical implementation of the mean-variance framework is affected by estimation uncertainty, which can lead to unstable allocations, extreme positions, and poor performance. This thesis provides a comparative empirical analysis of portfolio optimization methodolo- gies aimed at addressing these limitation. The analysis begins with the classical Markowitz framework, considering Global Minimum Variance and Maximum Trade-Off portfolios. Unconstrained portfolios are compared with long-only formulations to assess the effect of short-selling restrictions. The analysis then introduces regularization through the Least Absolute Shrinkage and Selection Operator. By reformulating the Markowitz problem within a linear regression framework, Lasso is employed to construct portfolios under different gross-exposure constraints. Finally, the framework is extended to Conditional Value-at-Risk (CVaR) optimization to explicitly account for downside tail risk. The empirical analysis considers a universe of 100 stocks selected from the constituents of the S&P 500 Index over the period from September 2004 to December 2025. The implemented strategies are evaluated through a comprehensive back-test framework using multiple performance and risk indicators to assess their empirical performance. The results show that unconstrained Markowitz portfolios are highly sensitive to estimation error and can generate extreme exposures and severe losses. Long-only constraints improve stability, although not necessarily risk-adjusted performance. Lasso-based portfolios provide a more effective balance between sparsity, diversification, and performance, while CVaR optimization improves downside-risk control without consistently producing superior overall performance. Overall, the findings highlight the limitations of unconstrained mean-variance optimization and demonstrate the superior out-of-sample performance of Lasso-based portfolios, suggesting that regularization provides a more effective approach to portfolio construction.
Econometrics
Machine learning
Asset allocation
Risk management
File in questo prodotto:
File Dimensione Formato  
Gnoato_Filippo.pdf

accesso aperto

Dimensione 9.95 MB
Formato Adobe PDF
9.95 MB Adobe PDF Visualizza/Apri

The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12608/115046