Forecasting Food Prices with Transformer Models Using the Informer Approach on WFP Data

For food security, market stability, and policy planning, accurate food price forecasting is essential, especially in the face of global shocks and diverse market behavior. The World Food Programme Global Food Prices dataset, which covers 1990–2025 and comprises 529 country–commodity series across nine categories, is used in this study to assess sophisticated transformer-based forecasting models. Monthly aggregation, full reindexing, imputation of missing values, train-only winsorization, and standardized scaling are all part of the stringent preprocessing pipeline. Three methods are benchmarked: an Informer transformer with a long-sequence encoder–decoder design, a LightGBM model with lag and rolling features, and a naïve last observation baseline. Informer consistently achieves the lowest forecasting errors, as shown by rolling-origin backtesting on the 2021–2025 test set. Informer achieves RMSE 1.48 USD, SMAPE 4.49 percent, and MAPE 6.49 percent at a three-month horizon, while recording RMSE 1.04 USD, SMAPE 2.07 percent, and MAPE 3.37 percent at a one-month horizon. The findings indicate that attention-based architectures offer the most dependable and scalable accuracy for multi-country and multi-commodity food price forecasting, even though simple benchmarks are still competitive.