Paper: arXiv 2406.00459

Authors: Lei Fan, Justin Sirignano

Abstract

Stochastic differential equation (SDE) models are the foundation for pricing and hedging financial derivatives. The drift and volatility functions in SDE models are typically chosen to be algebraic functions with a small number (less than 5) parameters which can be calibrated to market data. A more flexible approach is to use neural networks to model the drift and volatility functions, which provides more degrees-of-freedom to match observed market data. Training of models requires optimizing over an SDE, which is computationally challenging. For European options, we develop a fast stochastic gradient descent (SGD) algorithm for training the neural network-SDE model. Our SGD algorithm uses two independent SDE paths to obtain an unbiased estimate of the direction of steepest descent. For American options, we optimize over the corresponding Kolmogorov partial differential equation (PDE). The neural network appears as coefficient functions in the PDE. Models are trained on large datasets (many contracts), requiring either large simulations (many Monte Carlo samples for the stock price paths) or large numbers of PDEs (a PDE must be solved for each contract). Numerical results are presented for real market data including S&P 500 index options, S&P 100 index options, and single-stock American options. The neural-network-based SDE models are compared against the Black-Scholes model, the Dupire’s local volatility model, and the Heston model. Models are evaluated in terms of how accurate they are at pricing out-of-sample financial derivatives, which is a core task in derivative pricing at financial institutions.

Complexity vs Empirical Score

  • Math Complexity: 9.0/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced mathematics including stochastic differential equations, partial differential equations (Kolmogorov), and stochastic gradient descent theory. Empirically, it demonstrates out-of-sample pricing and hedging performance on real market data (S&P 500/100 options, American options) and compares results against established models, indicating a backtest-ready framework.

Research Flowchart

  flowchart TD
  A["Research Goal: Develop & evaluate neural network<br>models for pricing financial derivatives<br>(European & American options)"] --> B["Key Methodology: Use Neural Networks to model<br>drift & volatility in SDEs (vs. classical models)<br>Adaptive SDE Calibration"]
  B --> C["Inputs: Real Market Data<br>S&P 500, S&P 100, Single-Stock Options"]
  C --> D["Computational Processes:"]
  D --> E["European Options:<br>Stochastic Gradient Descent (SGD) on SDE<br>via unbiased path estimates"]
  D --> F["American Options:<br>Optimization via Kolmogorov PDE<br>with Neural Network coefficients"]
  E --> G["Key Findings/Outcomes:"]
  F --> G
  G --> H["NN-SDE models outperform<br>classical benchmarks (Black-Scholes,<br>Dupire, Heston) in out-of-sample pricing"]
  G --> I["Demonstrated feasibility of<br>scalable training on large<br>derivative datasets"]

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 proposes a fast stochastic gradient descent algorithm for training neural network SDE models for European options and a PDE optimization approach for American options.

Source sentence
For European options, we develop a fast stochastic gradient descent (SGD) algorithm for training the neural network-SDE model. Our SGD algorithm uses two independent SDE paths to obtain an unbiased estimate of the direction of steepest descent. For American options, we optimize over the corresponding Kolmogorov partial

Assumes 5

  • Drift and volatility functions are modeled by neural networks with many degrees of freedom. model
    Source sentence
    A more flexible approach is to use neural networks to model the drift and volatility functions, which provides more degrees-of-freedom to match observed market data.
  • Two independent SDE paths provide an unbiased gradient estimate. statistical
    Source sentence
    Our SGD algorithm uses two independent SDE paths to obtain an unbiased estimate of the direction of steepest descent.
  • Monte‑Carlo paths are i.i.d. and the risk‑free rate r is constant. model
    Source sentence
    where (Sℓ t, Y ℓ t ) are i.i.d. Monte Carlo paths of the SDE (St, Yt) and (without loss of generality) we have set Ti = T for notational convenience.
  • Training and testing splits are random with 80%/20% allocation each day. data
    Source sentence
    On each day, there are approximately 1100 contract pairs (call and put options). For each day, we randomly split the contracts into training and testing sets (80% training and 20% test).
Show 1 more
  • Automatic differentiation and GPUs provide fast parallel optimization over many PDEs. execution
    Source sentence
    We will use automatic differentiation (AD) and GPUs for fast, parallelized optimization over the PDEs for large numbers of contracts.

Data

  • S&P 500 index options (American) options unknown
    100 stocks from the S&P 500
    Source sentence
    The first dataset consists of American options on 100 different stocks from the S&P 500.
  • S&P 100 index options (European and American) options unknown
    S&P 100 index
    Source sentence
    The second dataset consists of European and American options on the S&P 100 Index.
  • Two‑month recalibration dataset options unknown
    September–October 2017 (42 trading days) · approximately 1100 contract pairs per day
    Source sentence
    The dataset consists of a 2 month time period (September and October of 2017), which includes 42 trading days. On each day, there are approximately 1100 contract pairs (call and put options).

Costs

Not applicable The paper reports no trading or execution result.

Finds 3

  • The recalibrated neural network‑SDE model outperforms the non‑recalibrated model on out‑of‑sample test data. performance
    Source sentence
    The recalibrated method outperforms the model with no recalibration, indicating that there is a benefit to recalibrating the neural network-SDE model.
  • Neural network SDE models achieve better pricing performance than traditional SDE models such as Black‑Scholes, Dupire local volatility, and Heston. pricing-calibration
    Source sentence
    The neural-network-based SDE models are compared against the Black-Scholes model, the Dupire's local volatility model, and the Heston model.
  • For American options, the 2D‑NN neural network model shows superior performance over traditional SDE models. performance
    Source sentence
    The neural network models, especially the 2D‑NN model, have better performance than the traditional SDE models.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 40 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 5 papers in the archive (5 by independent authors): Feynman-Kac Derivatives Pricing on the Full Forward Curve (uses its method or data)

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 (44)

  1. Some studies in machine learning using the game of checkers (1959)
  2. Imagenet: A large-scale hierarchical image database (2009)
  3. Learning word vectors for sentiment analysis (2011)
  4. Cifar-10 (canadian institute for advanced research) (2010)
  5. Option pricing: A simplified approach (1979)
  6. Options: A monte carlo approach (1977)
  7. The valuation of warrants: Implementing a new approach (1977)
  8. Finite difference methods and jump processes arising in the pricing of contingent claims: A synthesis (1978)
and 36 more
  1. The pricing of options and corporate liabilities (1973)
  2. Theory of rational option pricing (1973)
  3. An engine, not a camera: How financial models shape markets (2008)
  4. Option pricing when underlying stock returns are discontinuous (1976)
  5. A closed-form solution for options with stochastic volatility with applications to bond and currency options (1993)
  6. Pricing with a smile (1994)
  7. A jump-diffusion model for option pricing (2002)
  8. Stock price distributions with stochastic volatility: an analytic approach (1991)
  9. Option pricing with stochastic volatility models (2000)
  10. Stochastic implied trees: Arbitrage pricing with stochastic term and strike structure of volatility (1998)
  11. Beating the best: A neural network challenges the Black-Scholes formula (1993)
  12. A neural network model for estimating option prices (1993)
  13. A nonparametric approach to pricing and hedging derivative securities via learning networks (1994)
  14. Approximation by superpositions of a sigmoidal function (1989)
  15. Multilayer feedforward networks are universal approximators (1989)
  16. Valuing and hedging American put options using neural networks (1994)
  17. Data-Driven Option Pricing using Single and Multi-Asset Supervised Learning (2020) — link
  18. A data-centric approach to understanding the pricing of financial options (2002)
  19. Alternative neural network approach for option pricing and hedging (2003)
  20. Neural networks for option pricing and hedging: a literature review (2020)
  21. Option pricing using artificial neural networks: an Australian perspective (2013)
  22. Artificial Neural Network For Option Pricing With And Without Asymptotic Correction (2019)
  23. Robust pricing and hedging via neural SDEs (2020)
  24. An empirical analysis of the limit order book and the order flow in the Paris Bourse (1995)
  25. Limit order books (2013)
  26. High-frequency trading in a limit order book (2008)
  27. Deep learning (2015)
  28. Deep learning for limit order books (2019)
  29. Arbitrage-free neural-SDE market models (2021)
    1. stochastic integral (1944)
  30. The Dupire Formula
  31. Deeplob: Deep convolutional neural networks for limit order books (2019)
  32. Deep learning-based feature engineering for stock price movement prediction (2019)
  33. The Heston model and its extensions in Matlab and C (2013)
  34. "U ber die analytischen Methoden in der Wahrscheinlichkeitsrechnung (1931)
  35. Very deep convolutional networks for text classification (2016) — link

0 of 44 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.