Work Performed and Main Achievements
The central scientific achievement of this project is the development of AGRICAF (Agricultural Commodity Forecasting Framework), a pipeline for the analysis and medium-term forecasting of global agricultural commodity prices. AGRICAF produces monthly price forecasts up to twelve months ahead using only publicly available data, and accompanies every forecast with explanations of the factors driving each prediction.
The framework operates in four stages. In the first stage, time series of potential explanatory variables are collected from publicly available sources, covering production figures, trade statistics, stock reports, and historical price data. In the second stage, a statistical screening process identifies which variables genuinely improve forecast accuracy for each commodity and time horizon, discarding those that add noise. In the third stage, a battery of ten statistical and machine learning models — including gradient boosting methods, random forests, linear and additive models, and time series decomposition techniques — is evaluated through cross-validation, with the best-performing model selected automatically for each combination of commodity, calendar month, and forecast horizon. In the fourth stage, every forecast is accompanied by explanations using model-agnostic interpretability methods, specifically Shapley values and LIME, organised across four levels of detail.
AGRICAF was developed and validated for five agricultural commodities: wheat, maize, soybean, arabica coffee, and sugar.
In parallel, the project produced a peer-reviewed scientific publication examining food prices under conflict conditions.
