| Summary: | In the fast-paced world of financial technology, the integration of advanced Natural Language Processing (NLP) and Deep Reinforcement Learning (DRL) is transforming portfolio management. This thesis presents a pioneering portfolio management framework that leverages Transformer-based models and Large Language Models (LLMs) to enhance return predictions and sentiment extraction from extensive financial texts coupled with robust DRL trading agents to optimize portfolio performance. We introduce an adaptive retrieval-augmented framework for LLMs, finely tuned through instruction tuning to align with human instructions and incorporate market feedback. This approach enables dynamic weight adjustments within the Retrieval-Augmented Generation (RAG) module, showcasing the synergy between extracting more accurate underlying sentiment and better capturing stock movements, resulting in more profitable and robust portfolios. Additionally, we address the challenges of applying DRL to stock trading by developing the Hierarchical Reinforced Trader (HRT). This innovative strategy employs 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, while also substantially reducing losses and drawdowns in bearish and volatile market scenarios. Moreover, model interpretability is crucial given the black-box nature of both LLMs and DRL models. Practitioners without a strong machine learning background require clear interpretations of model outputs. To address this, one idea is to consider features univariately, omitting feature interactions to maintain interpretability. The Univariate Flagging Algorithm (UFA) identifies optimal cut points for each feature, flags them, and summarizes them to lower dimensions for each sample. We further enhance the UFA framework within the Generalized Additive Model (GAM), extending it to a broader framework capable of modeling any data generated by exponential family distributions. Our comparative analysis on various public benchmark datasets demonstrates that our extended framework not only achieves better predictive results than the original UFA but also retains its robustness against missing and imbalanced datasets. In conclusion, this thesis underscores the significant potential of integrating advanced NLP and DRL techniques into portfolio management, setting a new standard for intelligent financial decision-making.
|
|---|