Abstract
This thesis explores how artificial intelligence (AI), specifically Large Language Models (LLMs), Deep Reinforcement Learning (DRL), and agentic AI control layers, can enhance traditional portfolio optimisation frameworks and overcome limitations such as static risk modelling, exclusion of qualitative information, and backward-looking bias. First, a hybrid approach integrates sentiment derived views from multiple finance-specific LLMs using an LSTM aggregator within the Black Litterman model. This Meta-LLM framework delivers strong risk-adjusted performance, achieving annualised returns of 31.22%, a Sharpe ratio of 3.02, and Jensen’s alpha of 1.95%, outperforming conventional benchmarks.
Second, LLMs are tested as direct portfolio allocators through few-shot prompting and temperature tuning. While most allocations underperform optimised benchmarks, low-turnover variants such as Quant Factory LLaMA-3-8B-Instruct-Finance remain competitive post-costs and outperform naive diversification, highlighting the importance of turnover control. Third, a sentiment aware DRL framework combines Conditional Value-at-Risk (CVaR) diversification with FinBERT derived sentiment signals using the Deep Deterministic Policy Gradient algorithm. The resulting portfolio achieves 34.57% annualised returns, a Sharpe ratio of 1.05, and Jensen’s alpha of 0.38, demonstrating the value of integrating qualitative signals with dynamic, tail-risk-aware objectives.
Finally, an agentic architecture introduces a regime-aware state-action-controller that fuses regime inference, sentiment and uncertainty features, and friction-aware optimisation with deterministic control. Agentic portfolios deliver consistent improvements across different optimisation themes with forecast-centric and exploratory strategies showing the strongest gains, highlighting that the agentic layer enhances responsiveness to forward-looking signals while enforcing cost-aware control. In general, AI-enhanced methods outperform traditional approaches, sentiment integration improves responsiveness, and agentic control adds interpretability and execution discipline. For policy and governance, the results support the adoption of transparent, audit‑ready AI workflows...