Fast, Reliable, and Error-Bounded Option Pricing with Pretrained Neural Networks: A GJR–GARCH Study

Battle Count: 75

Highly relevant to quantitative trading. Provides a 4x10^5 speedup over Monte Carlo for GJR-GARCH option pricing, enabling real-time market-making, intraday risk management, and large-scale calibration previously impractical with simulation. The error-bounded, verifiable nature of the surrogate is critical for production trading systems. Microsecond-scale pricing on GPU (tens of millions of prices/second) makes GARCH-class models viable for latency-sensitive applications. The general recipe transfers to any simulation-based density model.

Moderate to high complexity. Requires: (1) Monte Carlo simulation infrastructure for GJR-GARCH with skewed-t innovations (GPU-accelerated, several days for full dataset), (2) Mixture Density Network training with CDF-matching loss over 512-quantile grid, (3) Dimensionless reduced-form parameterization to remove scale/shift symmetries, (4) Sobol low-discrepancy sampling for parameter space coverage, (5) Careful learning-rate scheduling (hold-then-decay), (6) Closed-form Black pricing from mixture components with drift correction. The architecture itself is simple (small fully-connected network), but the full pipeline with error analysis and design rules requires significant financial and ML expertise.

Training dataset publicly available on Hugging Face Hub (simu-ai/garch_densities). All experiment code (training, error analysis, figure generation) released alongside manuscript. Detailed architecture specifications, hyperparameters, and training procedures provided in appendices. However, GPU simulation generation took several days on rented servers, and some implementation details (exact Sobol streams, specific random seeds) may require additional effort to reproduce exactly.