Investment Strategies Ideation Using LLMs and Structured Multi-Modal Data

Presented to 200+ industry professionals at MathWorks Global Conference on building AI pipelines for quantitative finance research.

What We Presented

A five-stage AI pipeline that transforms 4,000+ unstructured academic papers into a queryable knowledge graph, enabling researchers to surface contrarian investment strategies through natural language queries.

Pipeline Architecture:

1. Sourcing & Acquisition — Automated collection from SSRN, arXiv, and NBER
2. Cleaning & Structuring — NLP-driven extraction of entities, methodologies, and findings
3. Analyzing Structures — Knowledge graph construction with 45,000+ entities using NetworkX and SpaCy
4. Ingestion & Analysis — LightRAG architecture combining graph-aware and vector retrieval
5. Prompting & Evaluation — Systematic evaluation across naive RAG, graph RAG, and LightRAG approaches

Key Result

LightRAG delivered 2.4x more comprehensive responses than naive RAG by preserving hierarchical relationships between financial concepts, asset classes, and research methodologies.