Processed REMMAQ data, code, and reproducibility artifacts for multi-horizon PM2.5 forecasting in Quito

This repository contains the processed data, Python code, reproducibility notebooks, model outputs, figures, and supporting artifacts associated with the study “Explainable multi-horizon fine particulate matter forecasting with spatial transfer and local calibration in Quito, Ecuador.” The analysis uses hourly air-quality and meteorological observations from five stations of the Metropolitan Air Quality Monitoring Network of Quito (REMMAQ), Ecuador, covering the period 2019–2025. The harmonized dataset contains 306,840 station-hour records and supports PM2.5 forecasting at 1-, 6-, 12-, and 24-hour horizons. The repository includes the processed unified dataset, data dictionary, reproducible Google Colab/Python notebook, scripts for model training and evaluation, CatBoost, LightGBM, XGBoost and Ridge comparisons, SHAP explainability analysis, split-conformal prediction intervals, leave-one-station-out spatial validation, feature-group ablation, local few-shot calibration, elevated-concentration diagnostics, block-bootstrap uncertainty estimates, and publication-quality figures. The chronological experimental design uses 2019–2022 for training, 2023 for model selection and calibration, and 2024–2025 as an independent test period. The processed dataset is distributed with a SHA-256 integrity check to ensure exact reproducibility. Raw observations originate from the official REMMAQ platform of the Secretaría de Ambiente del Distrito Metropolitano de Quito. This repository provides the processed and analysis-ready materials used to reproduce the results reported in the associated manuscript.