Thesis: Bayesian Black-Litterman under Nonlinear Dependence: A Hybrid Framework with GARCH-EVT Marginals, R-Vine Copulas, and Reinforcement Learning for Tactical Portfolio Allocation
| datacite.subject.fos | Natural sciences::Computer and information sciences::Computer sciences | |
| dc.contributor.correferente | Escudero Barros, Felipe Andres | |
| dc.contributor.department | Departamento de Industrias | |
| dc.contributor.guia | Kristjanpoller Rodriguez, Werner David | |
| dc.coverage.spatial | Campus Casa Central Valparaíso | |
| dc.creator | Mora Rojas, Felipe Ignacio | |
| dc.date.accessioned | 2025-10-09T12:32:52Z | |
| dc.date.available | 2025-10-09T12:32:52Z | |
| dc.date.issued | 2025-10-07 | |
| dc.description.abstract | Los marcos tradicionales de optimización de portafolios basados en media-varianza enfrentan dificultades para capturar dependencias no lineales y riesgos de cola, particularmente durante turbulencia del mercado. Este estudio propone un marco híbrido que integra cinco metodologías avanzadas: modelamiento marginal GARCH-EVT, cópulas R-vine, optimización bayesiana Black-Litterman, ensambles de aprendizaje profundo, y aprendizaje por refuerzo. Los modelos GARCH(1,1) con Teoría de Valores Extremos capturan clustering de volatilidad y colas pesadas, mientras que las cópulas R-vine descomponen la estructura de dependencia de 27 dimensiones en 351 pair-copulas optimizadas. Las views de inversionistas se generan mediante un ensamble de modelos LSTM, Transformer, y XGBoost, integrados en un marco bayesiano Black-Litterman, y desplegados a través de Optimización de Política Proximal para rebalanceo adaptativo. Utilizando 15.5 años de datos diarios (2010-2025) para 27 acciones principales de EE.UU. bajo un diseño walk-forward, la estrategia BL-RL propuesta logra un ratio de Sharpe de 0.910, superando significativamente los benchmarks tradicionales con ratios de Sharpe negativos (Reality Check p = 0.023). También demuestra protección mejorada a la baja (drawdown máximo -7.32\% vs. -16.59\% a -19.58\%) y resiliencia durante la caída del COVID-19 y el mercado bajista de 2022. Estos resultados subrayan el potencial de integrar técnicas estadísticas avanzadas, aprendizaje automático, y aprendizaje por refuerzo dentro de un marco bayesiano para gestión institucional robusta de portafolios. | es |
| dc.description.abstract | Traditional mean-variance portfolio frameworks struggle to capture nonlinear dependencies and tail risks, particularly during market turmoil. This study proposes a hybrid framework that integrates five advanced methodologies: GARCH-EVT marginal modeling, R-vine copulas, Bayesian Black-Litterman optimization, deep learning ensembles, and reinforcement learning. GARCH(1,1) models with Extreme Value Theory capture volatility clustering and fat tails, while R-vine copulas decompose the 27-dimensional dependence structure into 351 optimized pair-copulas. Investor views are generated via an ensemble of LSTM, Transformer, and XGBoost models, integrated into a Bayesian Black-Litterman framework, and deployed through Proximal Policy Optimization for adaptive rebalancing. Using 15.5 years of daily data (2010-2025) for 27 major U.S. equities under a walk-forward design, the proposed BL-RL strategy achieves a Sharpe ratio of 0.910, significantly outperforming traditional benchmarks with negative Sharpe ratios (Reality Check p = 0.023). It also demonstrates enhanced downside protection (maximum drawdown -7.32\% vs. -16.59\% to -19.58\%) and resilience during the COVID-19 crash and the 2022 bear market. These results underscore the potential of integrating advanced statistical, machine learning, and reinforcement learning techniques within a Bayesian framework for robust institutional portfolio management. | en_US |
| dc.description.degree | Magíster en Ciencias de la Ingeniería Industrial | |
| dc.driver | info:eu-repo/semantics/masterThesis | |
| dc.format.extent | 150 páginas | |
| dc.identifier.doi | 10.71959/cqfg-js51 | |
| dc.identifier.uri | https://cris.usm.cl/handle/123456789/4149 | |
| dc.identifier.uri | https://doi.org/10.71959/cqfg-js51 | |
| dc.language.iso | es | |
| dc.publisher | Universidad Técnica Federico Santa María | |
| dc.rights | Attribution-ShareAlike 4.0 International | en |
| dc.rights.uri | http://creativecommons.org/licenses/by-sa/4.0/ | |
| dc.subject | Bayesian Black-Litterman | |
| dc.subject | R-vine copulas | |
| dc.subject | Reinforcement learning | |
| dc.subject | Portfolio Allocation | |
| dc.subject | Tactical asset allocation | |
| dc.subject | Nonlinear dependence | |
| dc.subject | Extreme value theory | |
| dc.subject.ods | 9 Industria, innovación e infraestructura | |
| dc.title | Bayesian Black-Litterman under Nonlinear Dependence: A Hybrid Framework with GARCH-EVT Marginals, R-Vine Copulas, and Reinforcement Learning for Tactical Portfolio Allocation | |
| dspace.entity.type | Tesis |
Política de Privacidad 
