Design of a Generative AI-Driven Intelligent Investment Advisory System

Abstract

To address the need for dynamic strategy generation and semantic adaptation in intelligent investment advisory systems, this study proposes a generative architecture that integrates multi-source knowledge while supporting semantic reasoning, interpretability, and real-time user interaction. The system comprises modular components, including task scheduling, multi-source data fusion, a generation engine, semantic understanding, and strategy explanation, all enhanced by context-aware mechanisms and multi-dimensional security protection. The architecture leverages a multi-layered Transformer-based model and a tensor-level knowledge fusion framework, enabling real-time asset allocation and policy explanation. Empirical validation using heterogeneous financial datasets demonstrates the system's superiority in generative quality and robustness. Evaluation metrics indicate a BLEU-4 score of 44.89, a BERTScore of 91.31, semantic consistency of 0.89, strategy accuracy of 93.4%, and a recognition success rate exceeding 94.7% under adversarial perturbations. As shown in comparative experiments, the proposed system outperforms existing models such as GPT-2 and FinBERT in interpretability and interaction latency. The results confirm that the proposed system achieves high-quality generation, strong semantic alignment, and user trustworthiness in complex financial advisory scenarios.