This paper analyzes a potential solution to a key bottleneck in quantitative investment research: the reliance on highly trained researchers to manually interpret academic literature, generate signals, and validate strategies. A fine-tuned LLaMA-based agent is evaluated as a mechanism for automating repeatable research workflows, with paper replication used as a representative task. Creating a scalable research framework of this kind enables systematic investment firms to materially increase research throughput while reducing marginal human effort. The framework integrates transformer-based large language models, agentic orchestration with external tooling, fine-tuning, retrieval-augmented generation, and domain-specific chain-of-thought instruction to achieve high accuracy over large context windows. Results indicate that agentic LLM systems can reliably replicate core elements of quantitative research pipelines, demonstrating meaningful reductions in analyst time requirements while maintaining methodological fidelity.
Watch the team present this project at 1:47:20 in the session recording here.
Keywords: quantitative finance, large language models, LLaMA, fine-tuning, agentic AI,RAG, paper replication, portfolio simulation
Faculty Advisor
Dr. Justin Kurland is Vice President, Tech Fellow in the Engineering Division at Goldman Sachs. Justin’s 20+ years in data and analytics include a senior lectureship in the Department of Computer Science at the University of Waikato, 50+ peer-reviewed publications, and numerous open-source contributions in Machine Learning and Times Series forecasting. With degrees from Rutgers University, Boston University, and University College London, Justin also has extensive data science consulting experience working with organizations like Microsoft and the UK Home Office.