Large Language Models (LLMs) such as GPT-4 have reshaped numerous fields by demonstrating unprecedented capabilities in tasks traditionally reserved for specialized models. Recent research has highlighted the potential of LLMs in generating high-quality synthetic datasets, such as consumer interviews, in what is termed "Silicon Sampling" (Kraft et al., 2024). This raises an intriguing question: Could LLMs also excel in generating realistic synthetic financial transaction data, an area dominated by methods such as Generative Adversarial Networks (GANs) and Diffusion Models (Xu et al., 2019; Sattarov et al., 2023)?

Synthetic data has become essential in financial services to overcome challenges of data scarcity, regulatory constraints, and privacy protection (Assefa, 2020; Hilal et al., 2022). Exploring LLM-based approaches for synthetic transaction data could lead to transformative outcomes by enabling faster, more flexible, and potentially more privacy-preserving methods of synthetic data generation.

• Assessing the Potential of LLMs: Investigate methods for employing LLMs to generate realistic synthetic financial transaction data, inspired by recent advancements in silicon sampling techniques.
• Benchmarking Against Existing Methods: Evaluate if LLM-generated financial transaction data can match or exceed the quality, diversity, and utility of current state-of-the-art synthetic data generation methods (e.g., GANs, Diffusion Models).
• Impact on Financial Applications: Quantitatively compare how effectively models trained on LLM-generated data perform on critical financial tasks such as fraud detection, credit scoring, and risk assessment.
• Evaluating Efficiency and Practicality: Study the computational efficiency, cost-effectiveness, and scalability of LLM-based synthetic data methods compared to established alternatives.
   

• Strong analytical skills, curiosity, and willingness to think outside the box and apply existing LLMs in new contexts.
• Commitment to a structured and result-driven research approach, aiming for completion within ±6 months.
   

• Close supervision, regular feedback sessions, and direct access to expert guidance.
• Access to computational resources, relevant datasets, and collaboration opportunities with industry experts.
• Opportunity to contribute to groundbreaking research at the forefront of synthetic data generation with potential for publication in high-impact journals.

If you are interested, please reach out via email to mahei.li@unisg.ch. I look forward to discussing your thesis ideas and supporting you in achieving exceptional results.