Optimizing Option Price Prediction with Attention-Augmented Temporal Convolutions

Accurate option pricing is essential for risk management and financial decision-making. However, conventional parametric models, including Black-Scholes and binomial trees, depend on rigid assumptions and do not adjust to fluctuating market conditions. Deep learning methods like LSTMs and CNNs show promise but often lack strong feature selection mechanisms. To mitigate these limitations, we propose an Attention-Based Temporal Convolutional Network (TCN) for option valuation. Our model incorporates a feature-wise attention mechanism to improve feature selection, along with dilated causal convolutions to capture long-term dependencies in financial time series. The proposed architecture is assessed utilizing the S&P 500 options dataset, incorporating several volatility indicators. Experimental findings indicate that the Attention-TCN model surpasses conventional TCN, CNN, and LSTM models, attaining reduced MAE of 0.0142, RMSE of 0.0195, and a higher R² score of 0.9857. The t-test validates the significance of the performance enhancement (p-values < 0.05). Moreover, 5-fold cross-validation underscores the resilience of the proposed model. The results indicate that attention-based feature selection enhances model interpretability and improves predictive accuracy in financial forecasting.