RIDGE: An Autonomous Framework for Validation and Method Discovery in LLM-Generated Option Pricing

Battle Count: 63

Highly relevant to quantitative trading and derivatives pricing. The framework directly addresses the reliability of LLM-generated option pricing implementations, which is critical for derivatives desks, risk management, and algorithmic trading systems that depend on accurate option valuations. The stochastic volatility models (Heston, Bates, rough Heston) are industry-standard for equity and FX derivatives pricing. The methodological discoveries (CCF-PDE for HHW and G-SVJD) provide practical pricing algorithms. The validation framework could be integrated into production pipelines to ensure correctness of auto-generated pricing code. However, the paper focuses on validation methodology rather than trading strategies or market microstructure.

High complexity due to: (1) multi-block validation operator with deterministic numerical measurements (parity, monotonicity, convexity, density checks); (2) repository-driven knowledge management with structured entries and inheritance across model classes; (3) iterative LLM-guided generation and diagnosis loop; (4) stress-test design based on local sensitivity analysis; (5) benchmark hierarchy with multiple reference sources; (6) CCF-PDE solver development requiring Crank-Nicolson discretization, Lamperti coordinates, and adaptive grid refinement; (7) COS method with adaptive truncation domains and mode counts. The framework requires expertise in both numerical PDE methods and stochastic volatility modeling.

GitHub repository provided (https://github.com/ShQiangLiu/ridge) with supplementary material including per-iteration revision logs, accuracy results, benchmark results, computational costs, and stress breakdowns. The framework uses deterministic numerical measurements that are reproducible. However, LLM generation introduces stochasticity, and the paper acknowledges that finite-time guarantees and stability across independently initialized runs remain open questions. Ablation and reproducibility experiments with external generators (DeepSeek, ChatGPT) demonstrate framework robustness.