Scaling Tabular-Timeseries Foundation Models for Large-Scale Financial Data

TD Bank, as a leader in financial services, relies on predictive modeling to improve customer insights, risk management, and fraud detection. However, current machine learning approaches struggle to scale effectively across TD’s vast transactional datasets, leading to challenges in handling heterogeneous financial products, long-term forecasting, and multi-task learning. This project aims to address these challenges by developing scalable tabular-timeseries transformer architectures capable of learning from hundreds of millions of transactions across diverse financial services. By integrating self-supervised learning and multi-task optimization strategies, TD Bank will benefit from:
(1) Improved predictive performance in account acquisition, customer retention, fraud detection, and delinquency
prediction.
(2) Reduced model fragmentation, allowing a single scalable model to handle multiple financial objectives.
(3) Operational efficiency, decreasing computational costs by consolidating several models into a unified framework.
(4) Improved explainability, ensuring compliance with financial regulations and risk assessments.
This research aligns with TD’s commitment to AI innovation and financial technology advancements, providing direct business value through AI-driven decision-making and enhanced risk management.

Faculty Supervisor:

Scott Sanner

Student:

Partner:

Layer 6 AI

Discipline:

Computer science

Sector:

Finance and Insurance; Professional, scientific and technical services

University:

University of Toronto

Program: