This paper introduces a deep portfolio theory that builds upon Markowitz's classic risk-return trade-off framework. The authors develop a comprehensive four-step routine consisting of encode, calibrate, validate, and verify steps to create an automated and general portfolio selection process. At the core of this methodology are deep hierarchical compositions of portfolios constructed during the encoding phase, which are then calibrated to target various objective functions through multivariate payouts. The validation step optimizes regularization parameters, while the verification step employs cross-validation to trace an ex post deep portfolio efficient frontier, ensuring robustness and performance evaluation.

The significance of this work lies in its systematic approach to modernizing portfolio theory with deep learning techniques, offering a structured and automated alternative to traditional methods. By integrating regularization and cross-validation, the framework addresses overfitting and enhances out-of-sample performance. The numerical demonstration provides empirical validation of the theory's practical utility, showcasing its potential for applications in quantitative finance where complex, data-driven portfolio optimization is required.