Integrating recursive feature selection with automated machine learning framework for global wheat price prediction

Wheat price forecasting is challenging due to the interplay between economic trends, environmental variability and unpredictable market conditions, as well as the scarcity of reliable data. This research presents an innovative method for predicting global wheat prices by combining Recursive Feature Elimination with Cross Validation (RFECV) and Bayesian Ridge Regression in an Automated Machine Learning (AutoML) framework. The inclusion of key input variables - Lag 1, Lag 2 and an outlier indicator, marks a significant improvement in capturing regular and extreme market events which have a strong impact on prices. Using a 32 years monthly wheat price dataset (1990 to 2022) from Federal Reserve Economic Database, 35 machine learning models were evaluated. Based on test set performance, five top performing models i.e. Bayesian Ridge, Linear Regression, Least Angle Regression (Lars), Least Angle Regression with Cross-Validation (LarsCV) and Least Absolute Shrinkage and Selection Operator Least Angle Regression with Cross-Validation (Lasso Lars CV), were selected and further assessed across four train-test splits (80:20, 75:25, 70:30 and 65:35). The 80:20 split provided the most consistent performance across all models. Bayesian Ridge attained the lowest RMSE of 12.26, surpassing the other leading models with relative enhancements between 1.39% to 2.61%. These findings illustrate the model's resilience and its utility in informing policy decisions, market interventions, and strategic planning in international wheat trade. • An automated and optimized modeling approach was developed by combining Recursive Feature Elimination with Bayesian Ridge Regression to forecast global wheat prices from 1990 to 2022. • The approach identified Lag 1, Lag 2, and an outlier indicator as the most effective predictors for capturing both regular and extreme price variations. • Bayesian Ridge Regression delivered the most accurate results, achieving an RMSE of 12.26 (80:20 split), outperforming 34 other machine learning models tested.