Heterogeneous autoregression inspired neural network framework for predicting groundnut price volatility in India

Abstract

Groundnut, predominantly cultivated as a rainfed crop, is highly susceptible to significant price volatility. This study aimed to investigate and enhance the performance of traditional models for forecasting the realized volatility of groundnut price returns across five Indian states (Tamil Nadu, Telangana, Karnataka, Maharashtra, and Gujarat) by evaluating traditional models and neural network-based frameworks. Using groundnut price returns data spanning fourteen years and six months (01 January 2010 to 30 June 2024), weekly realized volatility was computed. The predictive behaviour of Heterogeneous Autoregression (HAR)-based neural network frameworks was evaluated. Neural networks were assessed using time series crossvalidation, and model metrics were employed to generate Model Confidence Sets (MCS). These sets were ranked based on model inclusion. The Extended Cochran–Armitage test was applied to identify and compare the best-performing models. Subsequently, model forecasts were tested and compared using the two-sided Diebold-Mariano test, and Model Confidence Sets were generated to evaluate predictive performance. For this unconventional weekly realized volatility forecast, the HAR (1,6,12) framework emerged as the most effective. Notably, the implementation of Convolutional Neural Network (CNNs) combined with RNNs, such as Conv1D -GRU and Conv1D-LSTM, demonstrated superior and consistent predictive performance across all states. Among standalone neural networks, GRU performed on par with CNN-based RNNs. These findings highlight the potential of CNN and GRU models as effective and accurate methods for forecasting agricultural price volatility.