Forecasting Wholesale Prices of Canchan potatoes using artificial intelligence models with data from the Ministry of Agrarian Development and Irrigation of Peru (2020–2025)

Evangelista Gamarra M.; Chancan Labajos J.A.; Figueroa Rosas P.E.; Alvarez Manrique R.E. · Frontiers in Sustainable Food Systems (2026) · DOI: 10.3389/fsufs.2026.1883363

Resumen

Accurately forecasting agricultural prices is a critical factor for stabilizing markets and optimizing decision-making across supply chains. This paper proposes a supervised machine learning application focused on forecasting the monthly prices of the Canchan potato in Peru, using the open data portal of the Ministerio de Desarrollo Agrario y Riego [Ministry of Agrarian Development and Irrigation] (MIDAGRI) as a baseline. Under this approach, six methodological strategies were implemented and comparatively evaluated: SARIMAX (operated via walk-forward validation), RNNs, LSTM architectures, and their respective hybrid architectures using XGBoost for residual learning. The annual performance of each algorithm was evaluated using the RMSE, MAE, and MSE error metrics. To select the final configuration, the analysis was not limited to finding the lowest error in the test split; it also audited the mathematical stability between the training and testing metrics to ensure the system genuinely possessed generalization capabilities. While the pure RNN and LSTM models achieved marginally lower test-set errors on individual metrics, the hybrid LSTM-XGBoost model exhibited the smallest train-test performance gap among all architectures (ΔRMSE ≈ 0.008), indicating superior generalization stability. Based on this criterion, LSTM-XGBoost was selected as the most robust configuration (RMSE = 0.1086, MAE = 0.0874, MSE = 0.0118). This research contributes a reproducible methodological framework that can be replicated as a baseline to design forecasting tools for other agricultural crops of high economic impact in the region. Copyright