Comparative Evaluation of Modern Deep Learning Methodologies for Portfolio Optimization
Comparative Evaluation of Modern Deep Learning Methodologies for Portfolio Optimization
Authors: Samuel Ozechi, Banjo Francis, Wisdom Yakanu, Joe Wayne Byers Date: 2026-04-27 Paper ID: openalex:2604.24486
Summary
This paper evaluates various deep learning architectures and hybrid models for portfolio optimization using historical market data from 2015 to 2023. By comparing models like GNNs, Transformers, and DRL against classical baselines such as MVO and 60/40, the study highlights the efficacy of combining deep learning’s predictive power with established financial theory. The findings indicate that hybrid models, particularly Transformer+GNN, offer the best risk-adjusted performance, while cautioning that standalone deep learning models may struggle with structural market complexities. The research underscores the enduring value of traditional optimization frameworks when supported by modern, representation-learning-based inputs.
Key Contributions
- Introduces a hybrid portfolio optimization framework combining deep learning architectures (Transformers, GNNs, DRL, Autoencoders) with classical financial models.
- Demonstrates that Transformer+GNN hybrid architectures provide superior stability and risk control by simultaneously capturing relational and temporal market dynamics.
- Provides empirical evidence that Mean-Variance Optimization (MVO) remains highly competitive when augmented with deep-learning-based feature inputs.
- Reveals that standalone DRL models often lack the structural awareness required for robust asset allocation compared to integrated hybrid approaches.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.24486
- paper_id
- 2604.24486
- paper_source
- openalex
- domain
- time-series
- tags
- time-seriesdeep-learningportfolio-optimizationfinancial-forecasting
- architectures
-
- datasets
-
- concept_slugs
-
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-30T06:05:47Z