Reproducibility package: How much does machine learning add to municipal crop yield prediction? Evidence from Chiapas, Mexico, under spatially blocked validation

Reproducibility package for the manuscript "How much does machine learning add to municipal crop yield prediction? Evidence from Chiapas, Mexico, under spatially blocked validation" (Y. Reyes Suárez, Universidad Autónoma de Chiapas). The study quantifies how much of the performance attributed to machine learning in municipal crop yield prediction depends on the evaluation protocol, and under which conditions models add predictive value over simple decision rules. A municipal panel of maize grain, beans and coffee cherry was assembled for Chiapas, Mexico (2003-2023) from SIAP agricultural closures, with NASA POWER climate series bias-corrected against 152-171 National Meteorological Service stations, and geographic attributes from INEGI. CONTENTS 1. FASE_2_PUNTOS_6_10_POST_TEST_V4.zip — frozen evaluation package. Model, hyperparameters and seeds were fixed and predictions generated before any access to test labels. Includes evaluation code, metrics, stored predictions, interpretability outputs and an internal manifest of 52 files. SHA-256: 6bce313ed5690d8c299715194edb4e9045822bbe8ee491540c1496afb6242bb2 2. EVIDENCIA_REPLICABILIDAD_LOCAL_POST_TEST_V4.zip — independent numerical replication of all published metrics, comparisons and confidence intervals (tolerance 1e-10). 3. analisis_complementarios/ — analyses supporting the central results:- Evaluation-protocol comparison: random versus municipality-blocked cross-validation on identical data, predictors and hyperparameters. Quantifies the overestimation of R2 (0.140 for beans, 0.203 for maize grain, 0.354 for coffee cherry).- Block permutation of predictor groups (geography, climate, management), joint geography-climate permutation to quantify block overlap, and stability of the importance ranking across spatial folds (Kendall's tau).- Mean-based baselines: global training mean and municipal historical mean, in both evaluation scenarios.- Paired municipality-clustered bootstrap of the RMSE difference between models and baselines.- Five-fold spatial cross-validation with and without SoilGrids 2.0 soil covariates (negative result: no improvement).- Full predictor listing per crop and variant (409 records), municipalities of each partition also present in training, and elevation mismatch between municipal seats and reanalysis cells.- Digital elevation model and geometries used in the cartographic figures.- Verification script that recomputes every published metric from the stored predictions and cross-checks it against the manuscript. SOURCE DATA (public): SIAP municipal agricultural closures, NASA POWER, Mexican National Meteorological Service, INEGI, and SoilGrids 2.0 (ISRIC). REQUIREMENTS: Python >= 3.10 with pandas, numpy, scikit-learn, xgboost, shap, scipy and matplotlib. CHANGES FROM VERSION 2.0.0: adds the evaluation-protocol comparison, joint block permutation, training-overlap table, digital elevation model and new cartographic figures; description updated to the reframed manuscript.