About This Event
This webinar investigates whether the neural scaling laws driving Large Language Model advancements apply to actuarial ratemaking. Utilizing a 4.5-million-record motor insurance dataset, we demonstrate that power-law performance improvements do exist for tabular data, though they are highly dependent on model architecture.
Speakers will present empirical evidence showing that while standard Transformers exhibit weak parameter scaling, "TabM" architectures and Transformers augmented with self-supervised learning achieve superior results in large-data regimes. The presentation covers the identification of distinct data regimes—from GLM-efficient small datasets to deep learning-dominated large datasets—and introduces novel architectural enhancements like multi-CLS pooling and TokenMoE.
