An Extreme Value Perspective on Learning Stress Laws

Battle Count: 67

Highly relevant to quantitative trading and risk management. The paper directly addresses VaR and CVaR estimation at extreme confidence levels (up to 0.999), which are critical for portfolio risk assessment. The ability to extrapolate tail probabilities beyond observed data (to probability order n^{-κ} for κ>1) is particularly valuable for stress testing and scenario generation. The conditional tail distribution analysis (Experiment 2) directly applies to understanding how stress in one portfolio propagates to overlapping portfolios. The method works with linear portfolio losses (w'ξ), making it directly applicable to standard portfolio risk frameworks. However, it is a risk measurement tool rather than a trading signal generator.

Moderate complexity. The core idea (radial-angular decomposition) is mathematically elegant, but practical implementation requires: (1) careful threshold selection balancing angular learning data vs. tail extrapolation accuracy, (2) GAN training with checkpoint selection based on projected Wasserstein diagnostics, (3) tail index estimation via Hill estimator, (4) proper normalization of generated angular vectors to unit sphere, (5) Monte Carlo integration for VaR/CVaR estimation. The GAN architecture is standard (fully connected MLPs), but training stability and model selection add practical complexity. The light-tailed variant requires additional estimation of Λ⋆.

The paper provides detailed algorithms (Algorithm 1, 2, 3), GAN architecture specifications (hidden widths, activations, optimizer settings), and experimental parameters (n=1000, M=10000, 1000 replications). However, no code repository is provided, and the light-tailed experiments are deferred to a full version. The synthetic data generation process is described but not made available. GAN training instability and checkpoint selection strategy are documented.