DYSANOS: Generative Dynamic Smooth Arbitrage-free Non-parametric Option Surfaces

Battle Count: 75

Highly relevant to quantitative trading. The paper directly addresses the problem of generating realistic, arbitrage-free option market paths for training hedging algorithms, backtesting strategies, and risk management. The ability to price any option at any strike/expiry on simulated paths is critical for practical trading systems. The detection of dynamic arbitrage in period returns is particularly important for practitioners who trade options between dates. The framework enables agents to choose which options to trade at each time step, unlike prior discrete-grid simulators.

High complexity. The implementation requires: (1) SANOS calibration via linear programming, (2) ML-SANOS neural network training with constrained reparameterization (sigmoid transforms, tri-band matrix inversion via Thomas algorithm), (3) PCA decomposition and AR(1) fitting, (4) exact Gaussian transition computation for large time steps, (5) conditional spot importance sampling with Gaussian bridges, (6) comprehensive data pipeline with multiple filtering and normalization steps, (7) GPU-accelerated Black-Scholes pricing, and (8) extensive arbitrage testing via linear programs. The mathematical sophistication spans stochastic calculus, convex optimization, and deep learning.

The paper provides a detailed data pipeline (Appendix A), explicit model specifications including hyperparameters (n_h=20, M=7 expiries, N=20 strikes, mu=0.5), and uses publicly accessible data (Option Metrics IvyDB via WRDS). The ML-SANOS architecture (3-layer feed-forward, 100 hidden nodes, SELU activation) is fully specified. However, no code repository is explicitly linked, and some implementation details (e.g., exact training hyperparameters, random seeds) are not fully disclosed.