Can Deep Learning Models Enhance the Accuracy of Agricultural Price Forecasting? Insights From India

Abstract

Forecasting agricultural commodity prices has been a long-standing challenge for researchers and policymakers. The diverse behaviors exhibited by price of different commodities, ranging from the high volatility, nonlinearity, and complexity of vegetables to the lower volatility and linear patterns of cereals. This different pattern necessitates the use of data-driven models to more precisely capture this complex behavior. This study aims to examine the efficiency of deep learning models in handling various types of price datasets. Three deep learning models, namely, gated recurrent unit (GRU), long short-term memory (LSTM), and recurrent neural network (RNN), are employed and compared against benchmark models including random walk with drift, autoregressive integrated moving average (ARIMA), artificial neural networks (ANN), and support vector regression (SVR). The monthly wholesale price data during January 2010 to December 2022 for 19 agricultural commodities across 143 markets in India have been utilized to illustrate the performance of models. Empirical comparison has been carried out by using different accuracy measures. The predictive accuracy is the highest for less-volatile crops such as cereals and pulses, while it is comparatively lower for crops with high-volatility like vegetables. The significant difference in prediction accuracy of different models has also been investigated with the help of Diebold Mariano test and its multivariate version. The study concluded that deep learning techniques outperformed machine learning and stochastic models across a wide range of commodities.