ATAbhay Tewari
Course · Advanced

Catastrophe Pricing with Generative AI

Build a working reinsurance pricing model from public data: from event tables to a conditional WGAN-GP, an EVT tail, compound simulation and an RBC-aligned premium.

This course teaches the method behind the Gift Re Pricing Lab by having you rebuild it. Each week pairs two or three lessons with a notebook that reproduces one stage of the pipeline on the same public data, so that by the capstone you have a model you built, understand, and can defend to an actuarial reviewer.

What you will be able to do. Assemble and trend a catastrophe loss dataset from public Indian sources; train a conditional Wasserstein GAN on tabular loss data and know when it has converged; fit and diagnose a Generalized Pareto tail; simulate compound losses to an excess-of-loss treaty; compute an RBC-aligned technical premium; and present the whole chain with the diagnostics a regulator expects.

What you need. Python at the level of pandas and a little PyTorch, and a working knowledge of insurance or reinsurance. The statistics are taught from the ground up; the code is provided.

Syllabus

Module 1 · The problem and the data

  1. Why rare risks break traditional pricing
  2. Public data for Indian catastrophe modelling
  3. Trending losses to today's exposure

Module 2 · Generative models for tabular losses

  1. GANs, Wasserstein distance and the gradient penalty
  2. Conditioning on physical hazard
  3. Training on fifty data points without fooling yourself

Module 3 · Extreme Value Theory

  1. Peaks over threshold and the Generalized Pareto
  2. Threshold diagnostics and bootstrap uncertainty
  3. Splicing an EVT tail into synthetic scenarios

Module 4 · From scenarios to a price

  1. Frequency, severity and the compound distribution
  2. Excess-of-loss layers, reinstatements and aggregates
  3. VaR, TVaR and the RBC capital charge

Module 5 · Validation and governance

  1. Benchmarks, KS and Wasserstein distances
  2. Backtesting and hold-outs
  3. What a regulator will ask

Module 6 · Capstone

  1. Price a layer of your choice and defend it