Maximizing the Interpretation of Black-Box Models & From Black Box to Glass Box

Maximizing the Interpretation of Black-Box Models 

We propose a novel global interpretable machine learning (IML) method for interpreting black-box models. One of the challenges actuaries face when applying machine learning in practice is the interpretability of the models, and our method contributes to solving it. Our method, called Maximum Interpretation Decomposition (MID), is designed to inherently maximize interpretability. MID addresses the limitations of existing global IML methods directly.

In the first part, we will discuss the theoretical background of this method. In the second part, we demonstrate how MID can be used to interpret black-box models using the open-source tools {midr} and {midlearn}. The demonstration highlights the practical utility of MID rather than focusing on implementation details.

Actuarial Digital Twin with Ontology-Driven Graph Database and Agentic AI

Actuaries face the “AI Paradox”: needing the speed of generative AI while requiring deterministic transparency for regulatory compliance. This session introduces an ontology-driven GraphRAG approach that moves beyond traditional document-based RAG by enabling structured reasoning across interconnected actuarial data. The Glass Box framework separates AI reasoning (“the Brain”) from actuarial calculation engines (“the Muscle”), enabling structured reasoning and automation of complex workflows such as product specification and reconciliation with full transparency. By applying a human-in-the-loop “Sandwich Workflow,” insurers can safely scale AI adoption while ensuring auditability and regulatory confidence, allowing actuaries to focus on high-value strategic decision-making.