Agricultural price shocks strongly affect farmers' income and food security. It is therefore important to understand and anticipate their origins and occurrence, particularly for the world's main agricultural commodities. In this study, we assess the impacts of yearly variations in regional maize productions and yields on global maize prices using several statistical and machine-learning (ML) methods. Our results show that, of all regions considered, Northern America is by far the most influential. More specifically, our models reveal that a yearly yield gain of +8% in Northern America negatively impacts the global maize price by about –7%, while a decrease of –0.1% is expected to increase global maize price by more than +7%. Our classification models show that a small decrease in the maize yield in Northern America can inflate the probability of maize price increase on the global scale. The maize productions in the other regions have a much lower influence on the global price. Among the tested methods, random forest and gradient boosting perform better than linear models. Our results highlight the interest of ML in analyzing global prices of major commodities and reveal the strong sensitivity of maize prices to small variations of maize production in Northern America.

This study analyses the quality of six regression algorithms in forecasting the monthly price of maize in its primary international trading market, using publicly available data of agricultural production at a regional scale. The forecasting process is done between one and twelve months ahead, using six different forecasting techniques. Three (CART, RF, and GBM) are tree-based machine learning techniques that capture the relative influence of maize-producing regions on global maize price variations. Additionally, we consider two types of linear models—standard multiple linear regression and vector autoregressive (VAR) model. Finally, TBATS serves as an advanced time-series model that holds the advantages of several commonly used time-series algorithms. The predictive capabilities of these six methods are compared by cross-validation. We find RF and GBM have superior forecasting abilities relative to the linear models. At the same time, TBATS is more accurate for short time forecasts when the time horizon is shorter than three months. On top of that, all models are trained to assess the marginal contribution of each producing region to the most extreme price shocks that occurred through the past 60 years of data in both positive and negative directions, using Shapley decompositions. Our results reveal a strong influence of North-American yield variation on the global price, except for the last months preceding the new-crop season.