A Machine-Learning Assessment of Climate Impacts on Global Agricultural Productivity

Climate change affects agricultural productivity through nonlinear and seasonally structured responses to temperature and precipitation. Global agricultural total factor productivity (TFP) attribution is commonly represented with a quadratic response to growing-season mean temperature and precipitation. We compare that benchmark with an expanded quadratic model and a Random Forest model using annual TFP growth for 168 countries, agricultural-area-weighted CRU TS observations, and eleven paired CMIP6 simulations with and without anthropogenic forcing. Under five-fold country-grouped cross-validation, the Random Forest shows higher out-of-sample predictive performance than either quadratic specification. Placebo reshuffling produces null R² distributions far below the observed value, while a year-by-region block bootstrap gives an average warming sensitivity near -0.9 percentage points of annual TFP growth per 1 °C. Held-out SHAP shows that precipitation, maximum temperature, and three-month climate windows contribute strongly to prediction, while accumulated local effects reveal nonlinear and direction-sensitive responses to year-to-year climate changes. Historical global attribution remains similar under the two response forms: the baseline quadratic gives -22.1% and the Random Forest -20.9% over 1962-2020. This agreement strengthens the finding that anthropogenic climate change has substantially reduced global agricultural productivity, while showing that a more flexible machine-learning response can reveal richer seasonal and nonlinear climate-TFP relationships. At the country level, attributed impacts are more dynamic and vary across rainfall-based moisture quintiles, with the Random Forest tending to attribute smaller losses than the quadratic in drier countries and larger losses in wetter countries.