I am excited to announce that Dr. Joel Ong has successfully defended his thesis on Modern Portfolio Construction with Advanced Deep Learning Models as part of the AIFi Lab at SUTD. It has been a pleasure to supervise Joel and guide him through his exploration of various multitask architectures and mixture of expert models, all aimed at leveraging deep learning for effective portfolio construction.

You can read the full thesis here or below the abstract.

Abstract:
This thesis explores the modern application of deep learning techniques in portfo- lio construction, presenting innovative methodologies that significantly enhance traditional investment strategies. Central to this research are three advanced frameworks that leverage deep learning to optimize financial portfolios.
The first framework introduces a diversified risk-adjusted TSMOM strategy utilizing multi-task learning. This approach simultaneously optimizes portfolio construction and volatility forecasting, resulting in improved portfolio performance by learn- ing both momentum signals and volatility estimators. Experimental results involving a diversified portfolio of continuous futures contracts demonstrate that this method outperforms existing TSMOM strategies.
The second framework employs a multi-task learning model with a multi-gate mixture of experts to optimize momentum portfolios across multiple timeframes. This model excels over benchmarks across various asset classes, effectively capturing com- plex momentum dynamics in equity indexes, fixed income, foreign exchange, and commodities. Extensive backtesting highlights its capacity to enhance risk-adjusted returns, underscoring its practical utility for portfolio management.
The third framework presents an adaptive sparse Transformer model designed for index tracking. By combining sparse modeling with deep learning, this framework optimizes passive investment strategies. Backtesting spanning from 2005 to 2024 re- veals that it delivers higher excess returns and lower tracking errors compared to existing models, showcasing the effectiveness of this approach in refining index tracking methodologies.
Lastly, we introduce the CurveMMOE model, a deep-learning framework for trad- ing commodity futures curves. This model integrates multi-task learning with a Mixture of Expert architecture, outperforming traditional methods on risk-adjusted re- turns. These frameworks contribute significantly to portfolio construction by harness- ing the power of deep learning techniques. They provide investment practitioners with innovative approaches to improve financial performance, particularly in challeng- ing market environments. This research advances our understanding of deep learning in finance and offers practical strategies for real-world investment scenarios.