Pricing of financial derivatives can generally be reduced to solving partial differential equations (PDEs). Starting from the classical Black-Scholes framework, pricing models that better reflect market characteristics, such as stochastic volatility and jump-diffusion models, have been developed, and classical numerical methods, such as the finite difference method, have long dominated their solution. However, as the dimensionality of pricing problems rises, grid-based methods face a sharp growth in computational cost, the so-called curse of dimensionality, while deep learning methods such as physics-informed neural networks (PINNs) have recently offered new approaches to high-dimensional PDEs. This paper systematically reviews and compares classical numerical methods (finite difference, finite element and spectral methods) with deep learning methods in terms of convergence, computational efficiency and applicability, and discusses their suitability for free-boundary problems, such as the early exercise of American options, and for high-dimensional multi-asset pricing. A literature review combined with comparative analysis was adopted, and the methods were compared across numerical stability, computational complexity and dimensional adaptability. The analysis shows that classical methods rest on relatively mature convergence theory and produce stable results for low-dimensional problems but carry a heavy computational burden in high dimensions, whereas deep learning methods are more efficient in high-dimensional settings, although their training stability and theoretical convergence guarantees remain to be improved. On this basis, the paper offers a reference for selecting numerical methods in different pricing scenarios and identifies hybrid frameworks combining the strengths of both categories as a promising direction for future research.
Show more