Fast catastrophe bond valuation with neural-network surrogates

Abstract

Catastrophe bonds are increasingly important risk-transfer securities, but structural pricing is too slow for real-time valuation, screening, and sensitivity analysis. We develop a neural-network surrogate for the pricing operator of a compound-Poisson catastrophe bond model. Training labels are generated by Monte Carlo simulation with importance sampling, so the network learns variance-reduced structural prices rather than sparse market quotes which are difficult to extrapolate reliably. Across Gamma and Lognormal severity specifications, the selected networks attain very small absolute approximation error on the stated training domain. After training, 1000 contracts are priced in about 0.03--0.04 seconds, compared with tens to hundreds of seconds for Monte Carlo with importance sampling and many hours for a partial integro-differential equation benchmark. The result is a fast in-domain structural valuation engine that also produces economically interpretable sensitivities to catastrophe intensity, attachment threshold, and interest rates.