Climate-Resilience and Cereal-Yield Joint Forecasting for Low- and Middle-Income Countries with Public Data and Grounded LLM Policy Explanation

Abstract

This paper reports a country-year forecasting experiment that links climate-resilience indicators, observed climate exposure, and cereal yields for low- and middle-income countries. The empirical panel combines the December 16, 2024 World Development Indicators archive for cereal yield, cereal production, cereal land and GDP, an ND-GAIN country-year file for vulnerability, readiness, ND-GAIN index and HDI, and QoG Environmental Indicators variables for annual temperature and rainfall. Country-specific temperature and rainfall anomalies are calculated against each country's 1981-2000 climate baseline. After complete-case filtering, the panel contains 39 countries, 702 complete rows, feature years 2002-2019, and target years 2003-2020. All features at year t are used to predict national cereal yield at year t+1. The evaluation uses a temporal split with feature years 2002-2016 for training and 2017-2019 for held-out testing, plus rolling-origin validation. Nine forecasting specifications are compared: persistence, country mean, panel fixed-effects ordinary least squares, ridge, Huber-ridge, regression tree, random forest, gradient boosting, and residual gradient boosting. The ridge panel model is the best held-out specification with RMSE 318.23 kg/ha, MAE 192.93 kg/ha, MAPE 9.08%, and R2 0.945. Recent cereal-yield history remains the strongest accuracy driver, while the climate and ND-GAIN variables add auditable exposure and adaptation context. A constrained policy-explanation layer then converts measured vulnerability, readiness, HDI, climate anomalies and yield volatility into country-specific adaptation explanations.