While LLMs are trained on sequential data, most of the data in the financial world (or corporates in general) is in form of tables or relational tables. Thus, they face the distributional mismatch, and lack ability to understand and perform predictive or forecasting tasks on such datasets. In recent times, there is a growing interest in Relational, Graph and Tabular foundation models. These have some similarities and some differences, our work is to understand and develop general purpose foundation models for such predictive/forecasting tasks with zero-shot or few shot learning on new datasets with in-context learning and without parameter update.
