Fatigue-Modulated Heston Pricing with Physics-Informed Neural Networks
Ethan Li
26/10/2026
Reliable option-pricing research must separate improvements in numerical learning from evidence that a proposed mechanism describes markets. This study tests whether a cumulative, fatigue-inspired shock index can usefully modify the variance target in a synthetic Heston model. European calls on a non-dividend-paying spot asset are priced with neural networks trained using both simulated prices and the governing pricing equation. An independent validation sample selects the equation-loss weight before two fresh test samples are generated. A conventional Fourier solution for the same time-dependent model supplies the evaluation reference and is checked against separate Monte Carlo calculations. Across eight training seeds, the selected network reduces pricing error relative to the data-only network by 10.3% on a fixed-variance test and 12.7% on a variance-spanning test. Enforcing the generating fatigue strength gives a small advantage over plain Heston physics on the variance-spanning sample, but the fixed-variance sample favors plain Heston physics. Conventional numerical pricing remains more accurate than the neural approximations. The findings demonstrate conditional sensitivity to a prescribed synthetic correction; they do not establish a fatigue mechanism in real markets or predictive validity for energy futures.
