Exploring Nonparametric Strategies for Pricing American Index Options

Hyeongmin Byun, Hyunwoong Ji, Jaewook Lee, Youngdoo Son2015 · DOI

TL;DR

This paper explores the use of machine learning (ML) techniques, specifically Bayesian Neural Networks (BNN), for pricing American index options, a notoriously complex financial instrument. The authors found that BNN outperforms traditional ad-hoc pricing methods and proposed an ensemble model combining BNN and ad-hoc approaches, which achieved the best predictive accuracy.

What Problem Does This Paper Address?

Pricing American options is challenging because, unlike European options, they can be exercised at any time before expiration. This early exercise feature means there's no simple closed-form solution, necessitating approximations or heuristic methods. Traditional numerical methods like tree models and finite difference methods struggle with a high number of state variables, while Monte Carlo methods, though independent of state variables, are computationally intensive. The paper aims to leverage the power of machine learning to find more accurate and efficient pricing models for American index options, comparing them against a conventional ad-hoc Black-Scholes approach.

Methods

The study investigates several machine learning methods for option pricing. The core methods explored are:

  • Support Vector Regression (SVR): A standard ML technique for regression tasks.
  • Neural Network (NN): A general-purpose neural network architecture.
  • Bayesian Neural Network (BNN): A probabilistic approach to neural networks that provides uncertainty estimates, which can be beneficial in financial modeling.
  • Ad-hoc Black-Scholes Model: A conventional, parametric approach adapted for American options using a closed-form approximation. This method involves generating a volatility surface and then using it to substitute the volatility component in the closed-form valuation.

A novel ensemble method is also proposed. This method combines the predictions of the BNN and the ad-hoc model. The idea is to leverage the ad-hoc model's consistent performance and use the BNN's predictions when they are deemed reliable (i.e., when the discrepancy between the two models is below a certain threshold, suggesting the BNN is not over-fitting). If the discrepancy is large, the ad-hoc model's prediction is favored.

Datasets

The study uses a dataset of S&P 100 index American put options from the Optionmetrics IVY database, spanning from January 2, 2003, to August 30, 2013. The initial dataset contained 584,755 instances with 23 attributes.

For the analysis, the authors applied several data screening filters:

  • Only traded options were considered.
  • Time to maturity was restricted to between 7 and 90 days.
  • Data from the first three years were removed due to unstable patterns.
  • Options with prices less than $1 were excluded to mitigate the effect of price discreteness.

After filtering, 83,700 instances remained. The primary attributes used for prediction were "time to maturity" and "moneyness" (underlying price / strike price).

Key Insights and Ideas

  • BNN's Superiority: Bayesian Neural Networks demonstrated a significant advantage over traditional ad-hoc pricing models and other machine learning methods in terms of in-sample error. This suggests that the probabilistic nature of BNNs is well-suited for capturing the complexities and potential uncertainties in option pricing.
  • Ensemble Power: The proposed ensemble method, which intelligently combines the strengths of BNN and the ad-hoc model, consistently yielded the best results, outperforming individual models in out-of-sample forecasting. This highlights the benefit of leveraging multiple models and mitigating individual weaknesses.
  • Time-to-Maturity Impact: The ensemble method's effectiveness varied with time to maturity. It showed particular promise in reducing extrapolation errors for short and long maturities, a phenomenon attributed to the "bagging" effect.

Key Results

  • In-Sample Performance: Bayesian Neural Networks consistently achieved the lowest in-sample errors across various time-to-maturity and moneyness categories, significantly outperforming the ad-hoc model.
  • Out-of-Sample Performance (1-day and 7-day ahead forecasts):
    • For short and long times to maturity, the ensemble method generally produced the lowest errors, outperforming both BNN and the ad-hoc model.
  • For medium ranges of time to maturity, BNN performed well, sometimes even better than the ensemble, suggesting it can interpolate effectively without the need for an ensemble.
  • The ensemble method's advantage in forecasting error was more pronounced for shorter prediction horizons (1-day ahead) and less so for longer horizons (7-day ahead), indicating its strength in reducing extrapolation effects.

Limitations and Future Work

  • Limitations: The paper notes that while BNN's variance of error is higher than the ad-hoc model, a significant portion of its error is due to over-fitting. The ensemble method aims to address this. The authors also acknowledge that the ad-hoc model's performance, while good, is not always the best.
  • Future Work: The authors plan to explore other machine learning methods, including multi-layer perceptrons and neural networks with 1-norm regularization. They also aim to develop models that strictly adhere to the no-arbitrage condition.

Summary updated

Abstract

Nonparametric models such as machine learning methods that are counterparts of parametric financial models were applied to pricing American index options. 10 year S&P 100 Index American options were adopted as experimental dataset and the both training (in-sample) and test (out-of-sample) errors of machine learning and ad-hoc pricing which is a conventional financial pricing model were calculated and compared to each other. We found out that Bayesian neural network outperforms the other pricing methods. Furthermore, we suggested an ensemble method which takes advantage of both machine learning method and ad-hoc pricing method and as a consequence, it shows the best performance.
BibTeX
@misc{byun2015exploring,
  author = {Hyeongmin Byun and Hyunwoong Ji and Jaewook Lee and Youngdoo Son},
  title = {{Exploring Nonparametric Strategies for Pricing American Index Options}},
  year = {2015},
  doi = {10.15224/978-1-63248-041-5-85},
}
Figure