Inverse Learning of Latent Risk-Neutral Densities from Irregular Option Quotes

Battle Count: 50

The paper is highly relevant to quantitative finance practitioners working on option pricing, implied volatility surface construction, and tail-risk estimation. Accurate RND recovery supports VaR/CVaR computation, exotic option pricing, and risk management. However, it does not directly propose trading strategies or execution algorithms. The finding that price accuracy does not guarantee density recovery is important for practitioners who use option-implied distributions for risk decisions. The NIFTY evaluation is specific to Indian equity index options.

High complexity: requires implementing four distinct neural operator architectures (DeepONet, FNO, Transformer, Set Decoder), multiple classical baselines (mixture lognormal with soft-L1 fitting, SVI, regularized discrete density, BL extraction), a synthetic simulator with three process families and six sampling regimes, numerical conditioning analysis (SVD of pricing operator), test-time adaptation with proximal penalties, and a comprehensive evaluation pipeline with bootstrap uncertainty quantification. The matched representation experiments and family-stratified analyses add further engineering overhead.

The paper specifies 10 seeds, 95% bootstrap intervals, prespecified confirmatory comparisons with Holm adjustment, exact sign-flip tests, and detailed hyperparameters (128-node grids, 3000 examples, 70/15/15 split, 50 Adam steps for adaptation). NIFTY data is CC0 from Zenodo. However, no GitHub repository URL is provided, and the synthetic simulator code is not explicitly linked. The auxiliary stress suite uses different objectives/baselines from the primary benchmark, which could cause confusion.