Paper: arXiv 2508.10208

Authors: Dixon Domfeh, Saeid Safarveisi

Abstract

Traditional models for pricing catastrophe (CAT) bonds struggle to capture the complex, relational data inherent in these instruments. This paper introduces CATNet, a novel framework that applies a geometric deep learning architecture, the Relational Graph Convolutional Network (R-GCN), to model the CAT bond primary market as a graph, leveraging its underlying network structure for spread prediction. Our analysis reveals that the CAT bond market exhibits the characteristics of a scale-free network, a structure dominated by a few highly connected and influential hubs. CATNet demonstrates high predictive performance, significantly outperforming a strong Random Forest benchmark. The inclusion of topological centrality measures as features provides a further, significant boost in accuracy. Interpretability analysis confirms that these network features are not mere statistical artifacts; they are quantitative proxies for long-held industry intuition regarding issuer reputation, underwriter influence, and peril concentration. This research provides evidence that network connectivity is a key determinant of price, offering a new paradigm for risk assessment and proving that graph-based models can deliver both state-of-the-art accuracy and deeper, quantifiable market insights.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced geometric deep learning (R-GCN) with graph theory concepts like scale-free networks, indicating high mathematical complexity. It also demonstrates high empirical rigor through backtesting on a real-world dataset of 803 CAT bond contracts, outperforming a strong benchmark and providing interpretability analysis.

Research Flowchart

  flowchart TD
  A["Research Goal: Predict<br>CAT bond spreads using<br>network topology"] --> B["Data Collection & Graph Construction"]
  B --> C["Feature Engineering:<br>Topological Centrality Measures"]
  C --> D["Model Architecture:<br>Relational Graph Convolutional Network R-GCN"]
  D --> E["Computational Process:<br>Training & Learning<br>Network Embeddings"]
  E --> F["Key Findings/Outcomes"]
  
  B --> B1(("Primary Market Data"))
  B --> B2(("Bond Characteristics"))
  
  F --> F1["CATNet outperforms<br>Random Forest benchmark"]
  F --> F2["Market is Scale-Free<br>Network dominated by hubs"]
  F --> F3["Network features act as<br>quantitative proxies for<br>industry intuition"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes CATNet, a novel framework that applies a Relational Graph Convolutional Network (R‑GCN) to model the CAT bond primary market as a graph for spread prediction.

Source sentence
This paper introduces CATNet, a novel framework that applies a geometric deep learning architecture, the Relational Graph Convolutional Network (R‑GCN), to model the CAT bond primary market as a graph, leveraging its underlying network structure for spread prediction.

Assumes 2

  • External macroeconomic factors are already embedded in CAT bond contract pricing and therefore can be omitted as separate features. model
    Source sentence
    operating on the premise that such external factors are already incorporated into the contract pricing.
  • Traditional ML models assume IID data, which is violated in CAT bond datasets due to temporal, spatial, and peril‑based dependencies. data
    Source sentence
    Many machine learning models assume data points are independent and identically distributed (IID), a condition often violated in CAT bond datasets due to temporal, spatial, and peril‑based dependencies.

Data

  • CAT bond contract dataset · Lane Financial reports and Artemis Deal Directory alternative proprietary
    1999-2021 · 803 catastrophe bond contracts
    Source sentence
    Our dataset comprises 803 catastrophe bond contracts issued in the primary market between 1999 and 2021, of which approximately 64% cover multiple perils and 34% are issued across multiple countries.

Costs

Not applicable The paper reports no trading or execution result.

Finds 4

  • The R‑GCN model achieves the highest average out‑of‑sample R² of 79.05%, outperforming XGBoost (77.63%) and Random Forest (73.09%). performance
    Source sentence
    The R‑GCN model consistently outperformed established tabular benchmarks, achieving the highest average R2 of 79.05%, compared to 77.63% for XGBoost and 73.09% for Random Forest.
  • The R‑GCN reduces mean‑squared error by 22.0% relative to Random Forest and 7.7% relative to XGBoost. predictive-accuracy
    Source sentence
    the R‑GCN recorded the lowest average error (MSE: 0.2109), representing a significant 22.0% reduction in error relative to the Random Forest baseline and a 7.7% reduction compared to XGBoost.
  • The CAT bond market exhibits a scale‑free network structure dominated by a few highly connected hubs. statistical
    Source sentence
    Our analysis reveals that the CAT bond market exhibits the characteristics of a scale‑free network, a structure dominated by a few highly connected and influential hubs.
  • Interpretability analysis shows that network centrality measures serve as quantitative proxies for issuer reputation, underwriter influence, and peril concentration. other
    Source sentence
    interpretability analysis showed that these centrality measures are not statistical artifacts; they are quantitative proxies for the influence of these network hubs, effectively capturing long‑held industry intuition about issuer reputation, underwriter influence, and peril concentration.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 43 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.