AI 450 · Year 4 · Semester 2 · 3 credits · AI & Machine Learning

Generative Models & Synthetic Insurance Data

Engineered imagination: generating mathematically faithful, privacy-preserving synthetic insurance universes from first principles.

Full practice exam · 30 Q · 120 min

Season 1 · 8 episodes

The Digital Twin Portfolio

To secure a two hundred million dollar reinsurance treaty without violating privacy statutes, a lead actuary must build a mathematically faithful, privacy-preserving synthetic twin of his company's entire historical portfolio.

Protagonist · Marcus Vance, FSA, Lead Modeling Actuary at NorthStar Mutual.

Setting · NorthStar Mutual Headquarters in Minneapolis, Minnesota, under concurrent scrutiny from the State Insurance Commissioner and lead reinsurer Alpenglow Re.

Stakes · If Marcus fails to deliver a synthetic commercial lines portfolio that matches real-world tail dependence while proving differential privacy under ASOP 23, NorthStar faces an immediate downgrade, treaty cancellation, and solvency insolvency sanctions.

Pick an episode

Episodes

Season finale: Marcus presents a fully calibrated, diffusion-guided tabular twin to the joint regulatory-reinsurance panel, successfully defending extreme climate stress scenarios while demonstrating zero privacy leakage.

From the lecturer: Welcome to AI 450. In this course, we transition from predictive modeling to generative architectures tailored to actuarial science. Synthetic data solves the industry's most critical bottlenecks: stringent privacy constraints under GDPR and HIPAA, extreme scarcity in catastrophe tails, and safe data sharing for collaborative pricing research. To get the maximum value from these lecture films, approach each session with paper, pencil, and an active Python terminal. For every generative family—from Variational Autoencoders and Copula simulators to Score-Based Diffusion models—derive the variational bounds and loss gradients by hand before implementing the code. Keep a dedicated dual ledger comparing statistical fidelity metrics (such as Wasserstein distance and copula rank correlations) against privacy leakage metrics (such as Membership Inference Risk and Differential Privacy epsilon). While this course does not directly feed an SOA or CAS preliminary exam, it operates at the cutting edge of modern actuarial practice, directly addressing the professional standards set out in ASOP No. 23 (Data Quality), ASOP No. 56 (Modeling), and emerging international guidance on artificial intelligence governance in insurance.

Cold open

Marcus stares at a cease-and-desist draft from the state commissioner on his monitor while Alpenglow Re gives him forty-eight hours to hand over forty years of commercial liability microdata or lose treaty backing.

preparing lecture…

Transcript

Contrast parametric density fitting with nonparametric Kernel Density Estimation (KDE) on skewed, heavy-tailed insurance claims.
Derive the asymptotic mean integrated squared error (MISE) of univariate KDE and compute Silverman's rule-of-thumb bandwidth.
Implement boundary corrections and log-transforms to prevent probability leakage below zero for strictly positive actuarial loss data.
Formulate multivariate density estimation as the prerequisite foundation for generative modelling and synthetic portfolio creation.