Deep learning CAT bond valuation

Julian Sester 1 , Huansang Xu 1 · 2025

Read Deep learning CAT bond valuation by Julian Sester 1 , Huansang Xu 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-25899.txt.

Introduction

Julian Sester and Huansang Xu wrote a 2025 paper on pricing catastrophe bonds with neural networks. These bonds transfer disaster risk to investors, but full structural pricing is slow. Slow pricing blocks real time valuation, screening, and what if analysis for traders.

Their network acts as a fast stand in for the pricing engine of a standard catastrophe bond model. Training prices came from careful simulations that reduce noise, not from sparse market quotes. After training, the network priced a thousand contracts in a tiny fraction of a second. Full simulation took tens to hundreds of seconds for the same set, and an equation based benchmark took hours. The surrogate also yields clear sensitivities to disaster intensity and thresholds.

Catalog record

CREATORJulian Sester 1 , Huansang Xu 1
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE93 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-25899

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Deep learning CAT bond valuation Julian Sester 1 , Huansang Xu 1 Abstract . In this paper, we propose an alternative valuation approach for CAT bonds where a pricing formula is learned by deep neural networks. Once trained, these networks can be used to price CAT bonds as a function of inputs that reflect both the current market conditions and the specific features of the contract. This approach offers two main advantages. First, due to the expressive power of neural networks, the trained model enables fast and accurate evaluation of CAT bond prices. Second, because of its fast execution the trained neural network can be easily analyzed to study its sensitivities w.r.t. changes of the underlying market conditions offering valuable insights for risk management. Keywords: CAT-bonds, deep learning, importance sampling, sensitivities August 10, 2026 1 National University of Singapore, Department of Mathematics, 21 Lower Kent Ridge Road, 119077. 1. Introduction Hurricane Andrew struck the Gulf Coast of the United States in August 1992, resulting in insurance-related damages of $15.5B and the failure of at least 8 insurance companies. Several insurance companies reassessed their risk exposure to the coastal regions across the US, see also [ 1 ] , and subsequently, risk exposure to catastrophic events in American coastal regions substantially decreased, while demand for catastrophe insurance increased. This shortage paved the way for the creation of catastrophe (CAT) bonds in the mid 1990s. In 1995, the American insurance industry had an estimated capital holding of $240B and the possibility of single-event losses of over $100B, see [ 11 ] . Insurance firms consequently started issuing CAT bonds to transfer risk off its balance sheet to the capital markets where institutional investors sought high yields uncorrelated with equity and fixed income returns, provided they indemnified the issuer if a catastrophe-related or trigger event occurred. Providing evidence for the resultant success of CAT bonds, according to [ 7 ] , the total outstanding market has grown from $1B to $49.1B in June 2024. When issuing a CAT bond, a trigger event and corresponding payout mechanism are defined in the bond contract. The catastrophe is defined using peril (catastrophe type) and territory (catastrophe location). The trigger event type is usually either indemnity (loss of the sponsor exceeds a predetermined threshold), industry index (the aggregate estimated industry loss, measured by an independent third-party index, exceeds a predetermined threshold) or parametric (measurable characteristic of the defined catastrophe exceeds a predetermined threshold), compare also Table 1 for an overview of the market shares of the different loss types. Trigger Event Type Market Share (%) Indemnity 70 Industry Index 20 Parametric 3 Modeled 1 Other 6 Table 1 . Market share of CAT bond trigger event types in 2021. Source: [ 7 ] . Consequently, a special purpose vehicle (SPV) is established. This is a legally separate entity that issues the CAT bond to investors. The SPV isolates the bond proceeds from the sponsor’s credit risk and satisfies regulatory requirements. Investors (e.g., hedge funds, pension funds) purchase the bond and pay the principal to the SPV. This capital is held in a collateral account, typically invested in low-risk securities. The sponsor then pays periodic coupon payments (usually above-market rates) to the investors, which are funded by the sponsor and the collateral returns. If no trigger event occurs during the bond’s term, the investors receive full repayment of their principal at maturity. If in contrast a trigger event occurs, some or all of the principal is transferred to the sponsor to cover the specified catastrophic losses, resulting in a loss for the investors. The pricing of CAT bonds has been a subject of increasing interest in both financial and actuarial literature, particularly due to the complexities introduced by low-probability, high-impact trigger events and the structural features of the bonds themselves. The literature on CAT bond pricing frequently employs stochastic modeling frameworks. Notably, Baryshnikov et al. introduced in [ 40 ] a framework, in analogy to the pricing of defaultable bonds (compare, e.g., [ 4 ] ), where a compound Poisson process is employed to capture the the cumulated random losses. While their model effectively incorporates the stochastic nature of catastrophic events, it does not yield a closed-form pricing formula. Nevertheless, the study provides a thorough numerical analysis based on simulation techniques. A simplified valuation framework is presented in Jarrow ( [ 24 ] ), where a closed-form solution for CAT bond pricing is derived. The model relies on the initial LIBOR zero-coupon bond price curve along with two key inputs: the hazard rate of the catastrophic event and the expected loss severity. A foundational mathematical treatment was presented by Burnecki and Kukla ( [ 10 ] ), who subsequently refine the methodology from [ 40 ] , addressing key limitations in the original formulation. This work is particularly influential in bridging actuarial modeling with applied probability theory, offering a robust and tractable approach for CAT bond pricing. Their enhanced framework is further applied by Härdle and López ( [ 20 ] ) to examine the calibration of CAT bonds for Mexican earthquakes, demonstrating its practical applicability in real-world catastrophe-linked securities. Further mathematical refinement was proposed by Nowak and Romaniuk ( [ 32 ] ), who incorporated interest rate risk and dependence structures into CAT bond valuation. Their model provide a generalized catastrophe bond pricing formula, assuming the one-factor stochastic diffusion form of the risk-free interest rate process, where the interest rate behavior is described by the multi-factor Cox–Ingersoll–Ross (CIR) model. Additional contributions such as Ma and Ma ( [ 29 ] ) addressed the impact of multiple risk drivers - loss severity distributions, claim arrival intensities, trigger thresholds, and stochastic interest rates - on CAT bond valuation. Recent research has focused on enhancing CAT bond pricing models by incorporating more realistic features of catastrophe risk and financial market dynamics. Hofer et al. ( [ 21 ] ) developed a risk-neutral pricing framework for CAT bonds, extending the basic structure of compound Poisson models to account for time-varying intensities. Burnecki ( [ 9 ] ) formalizes the structure of contingent convertible catastrophe bonds (CocoCats), modeling catastrophe losses with a time-inhomogeneous compound Poisson process and interest rates using the Longstaff model. Several numerical methods to pricing CAT bonds have been developed over the years with Monte Carlo methods and numerical PDE solvers remaining the main tools, especially when analytical tractability is lost due to model complexity. For example, Reshetar ( [ 36 ] ) models payoff structures contingent on both catastrophic property losses and catastrophic mortality events, using simulation-based techniques. Moreover, Siyamah et al. ( [ 37 ] ) utilized Monte Carlo and quasi-Monte Carlo simulation techniques for CAT bond valuation, highlighting efficiency gains in estimating expected losses under stochastic triggers. Our contribution to the literature on the pricing of CAT bonds is to establish a novel deep learning-based valuation approach where the relationship between model and contract features and the bond price is learned by a deep feedforward neural network under a compound Poisson framework, akin to the model introduced in [ 40 ] . The input features include both model-specific parameters, such as the intensity and frequency of catastrophic events, and contract specifications, including maturity, trigger threshold, and

Read the complete text