Doctoral Thesis: Next-Generation Intelligent Portfolio Management

By: Zijie Zhao

Thesis Supervisor(s) Roy Welsch (Management), Yoon Kim (EECS), Peter Szovolits (EECS)

Details

• Date: Thursday, May 9
• Time: 12:00 pm - 1:30 pm
• Category: Thesis Defense
• Location: Zoom link will be provided on request to rwelsch@mit.edu

Additional Location Details:

Abstract: In the fast-paced world of financial technology, the confluence of advanced Natural Language Processing (NLP) and Deep Reinforcement Learning (DRL) is revolutionizing portfolio management. This thesis presents a groundbreaking framework that utilizes Transformer-based models and Large Language Models (LLMs) to enhance return predictions and sentiment extraction from extensive financial texts, complemented by robust DRL trading agents for optimizing portfolio performance.

Our research introduces an adaptive retrieval-augmented framework for LLMs, which is finely tuned through Instruction Tuning to align with human instructions and incorporates real-time market feedback. This approach allows for dynamic weight adjustments within the Retrieval-Augmented Generation (RAG) module, demonstrating the synergy between extracting more accurate underlying sentiment and better capturing stock movement with improved portfolios from our financial LLMs. Additionally, we address the prevalent challenges in applying DRL to stock trading by developing the Hierarchical Reinforced Trader (HRT). This novel strategy utilizes a bi-level DRL framework that combines strategic stock selection via a High-Level Controller with effective trade executions managed by a Low-Level Controller. Our results demonstrate significant enhancements in portfolio management, achieving higher Sharpe ratios than the S&P 500 benchmark in bullish markets and substantially reducing losses and drawdowns in bearish and volatile market scenarios. We will showcase how these sophisticated techniques can significantly elevate investment strategies amidst the complexities of modern financial markets.