Dataset for: Beyond Accuracy — Diagnosing Systematic Biases in Classical ML Weather Models Using SHAP Explainability

This dataset contains raw and processed meteorological data for Dhaka, Bangladesh, used in the study "Beyond Accuracy: Diagnosing Systematic Biases in Classical ML Weather Models Using SHAP Explainability." Source: NASA POWER (Prediction Of Worldwide Energy Resources) Location: Dhaka, Bangladesh (23.8103°N, 90.4125°E) Parameters: PRECTOTCORR (corrected precipitation), T2M (temperature at 2 meters) Temporal coverage: 1981–2025 Temporal resolution: Daily Preprocessing: Data cleaning involved resolving missing values and index normalization, followed by feature engineering to enforce a strict input schema. The entire pipeline is verified by an automated pytest suite to ensure reproducible data integrity. Usage: The dataset was used to train and evaluate a comprehensive suite of six machine learning models: Linear Regression (as a baseline), Decision Tree, Random Forest, XGBoost, LightGBM, and CatBoost. These models were developed to predict two key meteorological parameters: PRECTOTCORR (corrected precipitation) and T2M (temperature at 2 meters). To ensure training stability and optimal performance, input features were processed using specialized scalers for each target variable. Finally, SHAP explainability was applied across these diverse architectures to rigorously diagnose and compare geographic and seasonal biases in model predictions.