A Co-evolutionary Approach for Heston Calibration

Battle Count: 70

Highly relevant to quantitative trading and derivatives desk operations. The paper directly addresses the practical challenge of fast, robust Heston calibration needed for daily/intraday options pricing, hedging, and risk management. The amortized neural surrogate approach (offline training, instant inference) aligns with production latency constraints. The finding that dataset diversity (LHS) outperforms target-focused sampling (GA-history) for generalization has direct implications for how trading desks construct calibration pipelines. However, the ~264-generation convergence time and the focus on a single model (Heston) somewhat limit immediate production deployment.

High implementation complexity due to the coupling of multiple components: (1) a genetic algorithm with elitist selection, arithmetic crossover, and Gaussian mutation over constrained parameter space; (2) a neuro-evolutionary population with architecture search (layer addition/removal, width modification, activation switching) and weight evolution; (3) a semi-analytical Heston pricing engine via Fourier inversion with numerically stable characteristic functions; (4) a closed feedback loop between GA elites and neural network training/injection; (5) constraint enforcement (Feller condition, parameter bounds). Requires careful handling of stochastic components, numerical integration stability, and multi-population synchronization.

The paper provides detailed hyperparameters (Tables 4), parameter ranges (Table 3), GA and NN operator specifications (Tables 1-2), and a clear algorithmic description. However, no code repository or implementation link is provided. The synthetic data generation procedure is well-described, and SPX data is sourced from Yahoo Finance (publicly available). The pricing engine follows Zhang et al. (2024) with a numerically stable characteristic function. Reproducibility is moderate: one could reconstruct the method from the description, but exact numerical results may vary due to stochastic components (GA randomness, neural network initialization).