A Comparative Study of Linear and Machine Learning Models for Wheat Price Forecasting in Gujarat, India

Abstract

Time series modelling and forecasting is a vibrant research field that has attracted the interest of the scientific community in recent decades. Forecasts of agricultural prices are proposed to be useful for farmers, governments, policy makers and agribusiness industries. In this study, an effort is made to compare the forecasting capabilities of well-known linear ARIMA models, ANN models, RNN models and Hybrid (ARIMA-ANN) models using data on weekly prices of wheat crop of Gujarat. In Dahod (2002 to 2023), Gondal (2003 to 2023) and Rajkot (2001 to 2023). Data were collected from the Agmarknet portal of the Government of India and converted into weekly averages. The dataset was split into 80 per cent for training and 20 per cent for testing. The Augmented Dickey-Fuller (ADF) test was used to assess stationarity and the Brock-Dechert-Scheinkman (BDS) test was employed to detect nonlinearity in the price series. Results of the ADF test confirmed non-stationarity in original series, which became stationary after first-order differencing, while the BDS test confirmed significant nonlinear patterns (p < 0.001) across all markets and embedding dimensions. The forecasting performance of these models was evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and Mean Square Error (MSE). Results showed that the RNN model outperformed all other models across all three markets of wheat crop, achieving the lowest error values. For wheat, the RNN recorded MAPE values of 2.98 per cent (Dahod), 2.89 per cent (Gondal) and 2.60 per cent (Rajkot). The ANN ranked second, followed by the Hybrid ARIMA-ANN and ARIMA models. Key findings revealed that the RNN model outperformed each individual ARIMA, ANN and Hybrid model for forecasting weekly prices of wheat crops all three market in Gujarat.