Pricing Catastrophe Bonds -- A Probabilistic Machine Learning Approach

Catastrophe (CAT) bonds have become increasingly vital in managing and transferring catastrophic risk. These bonds offer a source of capital to cover losses arising from natural disasters, allowing investors to diversify their portfolios while helping issuers mitigate potentially devastating financial consequences. Understanding the pricing dynamics of CAT bonds is essential, both for investors seeking informed decisions and for issuers optimizing their risk management strategies. This paper introduces a probabilistic machine-learning-based predictive framework for the pricing of CAT bonds, aiming to enhance empirical pricing accuracy and discover previously undetected nonlinear dependence between the key risk factors and CAT bond spreads. Early research by Lane (2000) laid the groundwork for CAT bond pricing literature, proposing a log-linear regression model employing conditional expected loss and probability of first loss as predictors. Subsequent studies expanded on this linear framework, incorporating additional predictors and examining pricing under diverse conditions. Gürtler et al. (2016) incorporated bond characteristics like trigger type and bond rating, while Braun (2016) integrated market condition indices, such as the Lane Synthetic Rate on Line index and the BB corporate bond spread. Götze and Gürtler (2020a) explored sponsor-related pricing inefficiencies across different market conditions, and Morana and Sbrana (2019) focused on the impact of climate change on CAT bond returns. Further extending the research scope, Zhao and Yu (2020) utilized actual catastrophe data to forecast CAT bond prices using market-based methods, Braun et al. (2022) developed factor pricing models for cross-sectional CAT bond returns, and Herrmann and Hibbeln (2023) investigated liquidity premiums in the secondary market.