A Comparative Study of Time Series, Machine Learning, and Deep Learning Models for Forecasting Global Price of Wheat

Wheat is a crucial grain in the global food supply and a key staple for many countries. Accurate wheat price forecasting is vital due to the economic impact of price volatility and supply–demand shocks. This study evaluated the forecasting performance of various time series, machine learning, and deep learning models in predicting global wheat prices from January 1990 to February 2024, using data from the Federal Reserve Economic Data on wheat and other commodity prices, and macroeconomic variables. Models tested were linear regression, partial least squares regression, autoregressive integrated moving average (ARIMA) and its variants, convolutional neural network (CNN), and long short-term memory (LSTM) were employed. Forecasting accuracy was assessed using mean absolute error and root mean squared error. The impact of the 2008 recession and COVID-19 pandemic was also considered. Overall, no single model excelled but ARIMA with exogenous variables (ARIMAX), partial least squares regression, and linear regression models performed best in capturing the influence of external variables on wheat prices. Seasonal ARIMA (SARIMA) performed the best among the univariate time series models. SARIMA was used to extrapolate variables for 6 months post-February 2024, which were fed into ARIMAX, linear, and partial least squares models for out-of-sample forecasting. The findings suggest that integrated models that combine historical data with external variables are needed for more accurate wheat price forecasting.