Revolutionizing agricultural stock volatility forecasting: a comparative study of machine learning and HAR-RV models

Abstract

This study investigates the realized volatility of the Shanghai Agricultural Stock Index (March 2017–May 2021), focusing on predictive accuracy. By incorporating three primary influencing factors, it evaluates the performance of traditional HAR-RV and LSTM models, demonstrating improved forecasting accuracy. The PCA-LSTM model, which integrates these factors through Principal Component Analysis (PCA), outperforms other models, showcasing its effectiveness. Empirical results indicate that: (1) incorporating influencing factors enhances the predictive performance of the HAR-RV and LSTM models; (2) the comprehensive models HAR-RV-ALL and LSTM-ALL, which integrate all three factors, demonstrate superior predictive accuracy; (3) the LSTM model consistently outperforms the HAR-RV model under comparable conditions; and (4) the PCA-LSTM model emerges as the most effective approach, proving its suitability for forecasting the realized volatility of agricultural stocks. This research contributes to the volatility forecasting literature by advancing the understanding of how machine learning enhances traditional models’ predictive performance.