Rethinking Synthetic Scenario Realism: Compatibility, Not Fidelity, Drives Hedging Performance

Battle Count: 73

Highly relevant to quantitative trading, particularly derivatives hedging and risk management. The paper directly addresses the critical question of how to design synthetic training data for deep hedging systems, which is central to modern quantitative finance. The compatibility framework provides actionable guidance for practitioners choosing scenario generators for hedging strategy development. The empirical findings about generator-hedger-task interactions are directly applicable to building production hedging systems.

The theoretical framework is mathematically sophisticated (IPM-based decomposition, constructive proofs). Empirical implementation requires: (1) calibrating multiple stochastic models (GBM, Merton, Heston) with stylized facts distance optimization, (2) training a VAE with specific architecture, (3) implementing three hedger architectures (Linear, NTBNet, MLP), (4) computing eight realism metrics, (5) training across multiple risk measures and task configurations. The hedger training itself is standard deep learning, but the full experimental pipeline is moderately complex.

The paper provides detailed specifications for all four scenario generators (GBM, Merton, Heston, VAE), three hedger architectures, four task configurations, and eight realism metrics. Calibration procedures (moment matching + L-BFGS-B) and training details (epochs, risk measures, parameters) are specified. However, no code repository is mentioned, and the VAE training data augmentation uses multiple external tickers. The Nikkei 225 data source is specified but exact data access details are limited.