A general framework for building machine learning models for pricing american index options with no-arbitrage and its limitation

Huisu Jang, Jaewook LeeMIDAS@PKDD/ECML 2016 · OpenReview

Abstract

Recent advancements in machine learning (ML) have significantly improved the accuracy of European option pricing. However, applying these techniques to American options—which can be exercised at any point before maturity—presents unique challenges, primarily the risk of generating prices that violate fundamental no-arbitrage principles. This paper introduces a novel ML framework designed to enforce no-arbitrage constraints during the training process. Evaluated on S&P 100 daily American put options, the framework demonstrates superior predictive accuracy compared to standard ML models. Nevertheless, the study also identifies a critical limitation: ML models exhibit reduced robustness compared to traditional parametric models when encountering out-of-sample data from different market regimes.

Introduction

In quantitative finance, the evolution of option pricing has moved from the foundational Black-Scholes model to sophisticated parametric models, such as Lévy processes, which better account for market phenomena like volatility smiles and fat-tailed distributions. While nonparametric machine learning models have recently emerged as powerful tools for price prediction, their application has been largely confined to European options.

American options, the most frequently traded variety, introduce a layer of complexity because they allow for early exercise. When standard ML models are applied to these instruments, they typically optimize for statistical fit alone. This often results in pricing outputs that allow for "arbitrage"—risk-free profit opportunities that violate the core laws of financial economics. There is a pressing need for a methodology that captures the high-dimensional accuracy of ML while remaining grounded in financial theory.

Methodology

The authors propose a general machine learning framework that structurally incorporates financial constraints into the learning objective. Unlike standard "black-box" models that map inputs to outputs without regard for underlying theory, this framework forces the nonparametric model to satisfy no-arbitrage conditions during the training phase.

By embedding these constraints directly into the model architecture or loss function, the framework ensures that the resulting price surfaces are theoretically sound and stable across time-varying daily data. This approach prevents the model from outputting financially impossible prices, making it a viable candidate for real-world trading and risk management environments.

Empirical Evaluation

The framework was tested using a dataset of S&P 100 daily American put options spanning one year. The research focused on two primary objectives:

  1. Predictive Accuracy: Comparing the constrained ML framework against state-of-the-art unconstrained ML models.
  2. Robustness: Benchmarking the ML approach against traditional parametric jump models, particularly under conditions of shifting data domains.

Key Findings

1. Superior Predictive Accuracy

The proposed no-arbitrage ML framework significantly outperformed existing machine learning models in terms of predictive accuracy on the S&P 100 dataset. By adhering to financial constraints, the model not only provides theoretically consistent prices but also achieves a better fit for the market realities of American options.

2. The Vulnerability of Domain Mismatch

A critical insight from this study is the exposure of ML limitations when training and testing data domains differ. While the ML model fits the training data with high precision, its performance can degrade significantly when the market regime changes (e.g., during a sudden volatility spike or a market crash).

In contrast, traditional parametric jump models—which are grounded in underlying diffusion processes—tend to generalize better in these "out-of-distribution" scenarios. This suggests that while ML is superior at interpolation within a known domain, it struggles with the extrapolation required during regime shifts.

Conclusion and Future Directions

Enforcing no-arbitrage constraints is an absolute requirement for making machine learning models safe and viable for American option pricing. The framework presented here successfully bridges the gap between statistical accuracy and financial theory. However, the sensitivity of these models to domain shifts remains a significant hurdle.

Future research should focus on developing hybrid models. By combining the robust, extrapolative power of parametric models with the high-accuracy, interpolative capabilities of machine learning, researchers may be able to create pricing tools that are both highly accurate and resilient to changing market conditions.

Summary updated

BibTeX
@inproceedings{jang2016general,
  author = {Huisu Jang and Jaewook Lee},
  title = {{A general framework for building machine learning models for pricing american index options with no-arbitrage and its limitation}},
  booktitle = {MIDAS@PKDD/ECML},
  year = {2016},
  url = {https://openreview.net/forum?id=EiZRHxo8i_a},
}
Figure