Research Assistant - Financial Language Models

ABOUT THE RISK AND INSURANCE STUDIES CENTRE (RISC)

Our vision is for risk mastery to advance universal progress through interdisciplinary knowledge, innovative education, and dialogue. We aim for a Risk Management & Insurance (RMI) field grounded in translational, rigorous scientific research, attracting and nurturing top emerging talent, and collaborating with diverse stakeholders for a safer, more sustainable, and prosperous future. Our mission is to create and mobilize interdisciplinary knowledge to advance the field of Risk Management & Insurance and to develop and establish innovative education and training programs that empower individuals to navigate and master the complexities of risk.

More specifically, with the world’s finest research experts on board and a unique place in the RMI industry, our overarching mission involves the creation of transformative scientific research and its timely translation into tangible, real-world applications within the RMI industry.

This entails:

• Conducting comprehensive and rigorous scientific research that addresses industry challenges and paves the way for better systemic risk management solutions.
• Cultivating emerging talent to be ready to step into rewarding RMI jobs and become the risk leaders of tomorrow.
• Working collaboratively with diverse stakeholders to bridge the gap between Academia, Industry, and Governments. This collaboration ensures that our initiatives contribute to a safer, more sustainable, and prosperous future for all Canadians. As we continue to expand, we are looking for a motivated candidate to join our team.

This is a four-month paid research position with the possibility of developing into a publishable article. You will work directly with Dr. Ed Furman of RISC Foundation and York University and Dr. Walid Mnif and will have access to a collaborative network spanning industry and academia.

Financial language models hallucinate. They generate confident-sounding answers that are factually wrong. Most existing work catches these errors after the fact. This project asks a harder question: can we predict that a model is about to hallucinate before it generates a single token?

You will work on a three-part system applied to Canadian bank filings and financial disclosure analysis, where a missed hallucination has real regulatory and liability consequences.

• Risk scoring: build a scorer that examines the incoming question, retrieved documents, and internal model signals to output a probability that the response will contain an error
• Error structure: investigate whether model mistakes follow patterns. Consistent confabulation of regulatory ratios or hallucinated dates under sparse context are not random failures -- they are structured, and can be characterized
• Statistical guarantees: wrap the system in a formal guarantee, that is for any flagging threshold, bound the fraction of hallucinations missed, with a proof that holds regardless of query distribution shift

A secondary thread distinguishes between model failure and unanswerable questions, separating "the model got it wrong" from "the evidence was never there."

We are looking for a strong researcher who sits at the intersection of statistics, mathematics, and computer science.

• Graduate training -- Ph.D. level preferred -- in statistics, mathematics, computer science, or a closely related field
• Proficiency in Python and deep learning frameworks, e.g. Py Torch
• Solid understanding of neural network architectures including transformers
• Working knowledge of linear algebra and statistical testing

• Experience with model calibration
• Familiarity with conformal prediction or distribution-free uncertainty quantification
• Interest in financial applications or NLP