Mispricing Detection in the S&P 500 ETF Options Market Based on an XGBoost Model

Abstract

Based on high-frequency data for S&P 500 index options (SPX) obtained from the WRDS database, this paper builds an empirical framework for identifying option pricing mismatches. First, taking the Black–Scholes model as the theoretical pricing benchmark and combining it with actual market quotes, we introduce a relative mispricing measure and adopt a quantile standardization strategy to construct mispricing labels, which are divided into three categories: overpriced, underpriced, and fairly priced. Then, we construct a mispricing identification model using an XGBoost multiclass classifier, trained and predicted with multidimensional features including contract attributes, trading indicators, and volatility. In data preprocessing, we clean both option and underlying information and strictly align them using a dual primary key of trading date and contract identifier to ensure the accuracy and consistency of label construction. The empirical results show that the model achieves good identification performance in the out-of-sample test period (2023), with an overall accuracy of 77.6%. The precision and recall for the underpriced category are particularly strong, indicating that the model can effectively identify pricing deviations with potential arbitrage value. At the same time, the results reflect the existence of certain structural inefficiency in the options market. Although the Black–Scholes framework relies on idealized assumptions, it still has important reference value in mispricing identification. This study not only verifies the feasibility of using machine learning to identify option mispricing, but also provides a methodological foundation for subsequent quantitative trading strategy design.