Paper: arXiv 2505.05121
Authors: Jasper Rou
Abstract
Option pricing often requires solving partial differential equations (PDEs). Although deep learning-based PDE solvers have recently emerged as quick solutions to this problem, their empirical and quantitative accuracy remain not well understood, hindering their real-world applicability. In this research, our aim is to offer actionable insights into the utility of deep PDE solvers for practical option pricing implementation. Through comparative experiments in both the Black–Scholes and the Heston model, we assess the empirical performance of two neural network algorithms to solve PDEs: the Deep Galerkin Method and the Time Deep Gradient Flow method (TDGF). We determine their empirical convergence rates and training time as functions of (i) the number of sampling stages, (ii) the number of samples, (iii) the number of layers, and (iv) the number of nodes per layer. For the TDGF, we also consider the order of the discretization scheme and the number of time steps.
Complexity vs Empirical Score
- Math Complexity: 7.0/10
- Empirical Rigor: 6.5/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced mathematical concepts such as PDE discretization, Sobolev spaces, and deep learning optimizers, but lacks code/datasets for immediate backtesting; the experimental design is systematic and data-driven, focusing on empirical error analysis for practical implementation.
Research Flowchart
flowchart TD
A["Research Goal<br>Assess accuracy & utility of<br>deep PDE solvers for option pricing"] --> B{"Methodology"}
B --> C["Comparative Experiments<br>Black-Scholes & Heston Models"]
B --> D["Algorithms Evaluated<br>Deep Galerkin Method<br>Time Deep Gradient Flow TDGF"]
B --> E["Input Parameters Varied<br>Sampling stages, Samples, Layers, Nodes<br>TDGF: Discretization order, Time steps"]
C --> F["Computational Process<br>Simulation & Training<br>Convergence Rate & Time Analysis"]
D --> F
E --> F
F --> G["Key Findings & Outcomes<br>Empirical Convergence Rates<br>Training Time vs. Complexity<br>Actionable Insights for Implementation"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper empirically evaluates two neural network PDE solvers, the Deep Galerkin Method and the Time Deep Gradient Flow method, for option pricing in Black‑Scholes and Heston models.
Source sentence
Through comparative experiments in both the Black--Scholes and the Heston model, we assess the empirical performance of two neural network algorithms to solve PDEs: the Deep Galerkin Method and the Time Deep Gradient Flow method (TDGF).
Assumes 2
- Option prices follow the Black‑Scholes PDE with known analytical solution. model
Source sentence
Indeed, let u : [0, T] × Ω→R denote the price of this derivative, with Ω⊆R and t the time to maturity. Then, u solves the Black–Scholes PDE
- The Heston model does not have a closed‑form price, but a reference price can be obtained via the COS method using its characteristic function. model
Source sentence
This PDE does not have an exact solution. The characteristic function of the Heston model does have an analytical representation [13], from which a reference price can be determined using the COS method [8].
Data
- Black‑Scholes synthetic option data synthetic unknown
Source sentence
This PDE has an exact solution: u(t, S) = SΦ(d1) −Ke−rtΦ(d2), with Φ the standard normal cumulative distribution function,
- Heston synthetic option data (COS reference prices) synthetic unknown
Source sentence
The characteristic function of the Heston model does have an analytical representation [13], from which a reference price can be determined using the COS method [8].
Costs
Not applicable
Finds 5
- The L2‑error decreases almost linearly with the number of sampling stages for both methods, but the benefit stops after a certain point. performance
Source sentence
Since the training time grows linearly with the the number of sampling stages, it would be optimal to stop at this point.
- Increasing the number of layers tends to reduce error, though without a clear rate; two or three layers are recommended. performance
Source sentence
Since the number of layers has a big influence on the training time, we recommend using two or three layers.
- For TDGF, more time steps lower the L2‑error, with second‑order stepping reducing error faster than first‑order. performance
Source sentence
For the TDGF, the number of time steps decreases the L2-error. Using a second-order time‑stepping method, the error decreases quicker than using a first‑order method.
- For TDGF, larger numbers of nodes per layer improve accuracy; for DGM, error does not decrease with nodes per layer, so a smaller size is advised. performance
Source sentence
For the DGM, the L2‑error does not decrease with the number of nodes per layer. We recommend choosing a smaller number of nodes per layer like thirty.
- TDGF benefits slightly from more training samples, while DGM shows no error reduction with additional samples. performance
Source sentence
For the TDGF the L2‑error decreases slightly with the number of samples
Checked by others
Contradicted by
- Application of Tensor Neural Networks to Pricing Bermudan Swaptions
Compare the findings
This paper: Increasing the number of layers tends to reduce error, though without a clear rate; two or three layers are recommended.
That paper: Dense Neural Networks suffer from over‑fitting when depth is increased to 20 layers, leading to worse price accuracy and stability.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (21)
- Multifactor approximation of rough volatility models (2019)
- Deep optimal stopping (2019)
- The pricing of options and corporate liabilities (1973)
- Deep hedging (2019)
- Discrete gradient flow approximations of high dimensional evolution partial differential equations via deep neural networks (2023)
- An Overview on Machine Learning Methods for Partial Differential Equations: from Physics Informed Neural Networks to Deep Operator Learning (2024) — link
- Solving high-dimensional partial differential equations using deep learning (2018)
- Global Convergence of Deep Galerkin and PINNs Methods for Solving Partial Differential Equations (2023) — link
and 13 more
- Adam: A method for stochastic optimization (2014) — link
- Pricing options and computing implied volatilities using neural networks (2019)
- Convergence of Time-stepping Deep Gradient Flow Methods (2025)
- Machine learning for option pricing: an empirical investigation of network architectures — → 2307.07657
- A time-stepping deep gradient flow method for option pricing in (rough) diffusion models — → 2403.00746
- Deep learning for portfolio optimization (2020) — link
- D elft B lue S upercomputer ( P hase 1) (2022)
- Implicit–explicit BDF methods for the K uramoto– S ivashinsky equation (2004)
- A deep implicit-explicit minimizing movement method for option pricing in jump-diffusion models — → 2401.06740
- Error Analysis of Option Pricing via Deep PDE Solvers: Empirical Study — → 2311.07231
- A closed-form solution for options with stochastic volatility with applications to bond and currency options (1993)
- A novel pricing method for E uropean options based on F ourier-cosine series expansions (2009)
- DGM : A deep learning algorithm for solving partial differential equations (2018)
4 of 21 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.