Abstract
Data science is increasingly adopted in actuarial research and insurance practice, yet its reliability depends on every stage of the workflow, from data preparation to model building to operational decisions, in a regulated, data-sensitive industry. This talk follows that pipeline through three connected contributions. We begin with data preparation, showing that conventional practices like random train-test partitioning produce unstable results on highly imbalanced insurance loss data. We propose a statistically principled alternative: support points for representative splitting, Chatterjee's correlation coefficient for non-parametric feature screening, and integrated missing-data handling. Evaluations on simulated and benchmark datasets demonstrate improved robustness and interpretability at lower computational cost. We then embed these methods in InsurAutoML, an insurance-tailored automated machine learning framework whose data balancing, ensemble pipelines, and customized loss functions let non-expert users deploy robust models with a few lines of code. Finally, moving from prediction to decision support, we use millions of historical warranty claims to fine-tune locally deployed large language models via Low-Rank Adaptation, generating structured corrective-action recommendations from unstructured claim narratives. Domain-adaptive fine-tuning substantially outperforms general-purpose LLMs, with about 98% of cases closely matching ground truth. Together, these results point toward trustworthy, governable automation across the insurance analytics pipeline.
