LSTM-Based Option Pricing: Evaluating Model Performance through Input Variation and SHAP Interpretation

The rapid growth in options trading volume highlights the need for accurate pricing models in financial markets. Traditional models like Black-Scholes provide a foundational approach but often have limitation due to their assumptions of constant volatility and normally distributed returns, which do not align with real world data. These limitations result in pricing inaccuracies, creating a demand for more flexible and robust models. Machine learning models, particularly LSTMs, present a strong alternative by capturing complex financial patterns without relying on strict assumptions. How ever, their predictive power can make them black boxes, making it difficult to un derstand their decision-making processes. This lack of interpretability limits their use in finance, where transparency is critical. To address these challenges, this study applies an LSTM model along with SHAP to create a transparent and accurate approach for option pricing. Four models were developed for call and put options, progressively adding features: traditional Black Scholes variables, GARCH volatility, Black-Scholes prices and a range of market indicators like technical indicators, trading volume, the S&P 500 index and the VIX. The models’ performance was evaluated in different market conditions and SHAP analysis provided insights into feature importance. The results show that historical volatility is consistently the most significant fac tor in predicting option prices. The inclusion of GARCH volatility and technical indicators improves model performance, particularly for high-strike prices. SHAP analysis highlights the relative importance of features, confirming that while techni cal indicators are relevant under specific conditions, volatility remains the dominant factor. Model 4, which includes all features, demonstrates the best overall per formance, making it the preferred approach for accurate and interpretable option pricing. This work contributes to Explainable AI in finance, offering a transparent, data-driven solution for option valuation.