A Wasserstein GAN-based climate scenario generator for risk management and insurance: the case of soil subsidence

By Antoine Heranval, Olivier Lopez, Didier Ngatcha, Daniel Nkameni

Rating

1738
Battle Count: 67

Relevance

3/10
While not directly related to quantitative trading, the paper has indirect relevance through: (1) economic scenario generation methodology applicable to financial risk modeling, (2) Value-at-Risk estimation techniques relevant to portfolio risk management, (3) GAN-based generative approaches transferable to financial time series simulation, (4) climate risk quantification increasingly relevant for ESG investing and catastrophe bond pricing. The methodology is more directly applicable to insurance risk management than to trading strategies.

Implementation Complexity

8/10
High complexity due to: (1) WGAN-GP training with multiple loss components (adversarial, reconstruction, feature matching), (2) UNet generator with 5 encoding/decoding stages, residual blocks, scSE attention, and stochastic depth, (3) Dual discriminator architecture (frame + patch), (4) Differentiable augmentation (DiffAugment), (5) Spectral normalization, (6) Complex data preprocessing with spatial interpolation, (7) 1500 epochs with careful hyperparameter tuning, (8) Iterative forecasting procedure requiring sequential generation, (9) Integration with CatNat eligibility methodology and financial cost estimation.

Reproducibility

4/5
All code and results are freely available on GitHub (https://github.com/dnkameni/SwiGAN). Detailed hyperparameters, architecture specifications, and training procedures are provided in the appendix. Data sources (DRIAS, Météo-France) are publicly accessible. However, the specific training infrastructure and computational resources are not fully detailed.

About this paper

Methodology: SwiGAN - Conditional Wasserstein GAN with Gradient Penalty (WGAN-GP). Problem types: Generative Modeling, Time Series Forecasting, Risk Management, Computer Vision, Density Estimation.

The interactive Everscope explorer (charts, battles, favorites) loads below.