Accurate short-term particulate-matter forecasting is needed for proactive exposure management, but reported performance often relies on random data splits or station-specific models that do not reveal temporal leakage, spatial distribution shift, or forecast uncertainty. This study developed an explainable and uncertainty-aware framework for hourly fine particulate matter (PM2.5) forecasting across five stations of the Metropolitan Air Quality Monitoring Network of Quito, Ecuador. The harmonized 2019–2025 panel contained 306,840 station-hour records. Models were trained on 2019–2022, selected using 2023 only, and evaluated on an untouched 2024–2025 test set for 1-, 6-, 12-, and 24-h horizons. CatBoost, selected before test evaluation, obtained RMSE values of 6.48, 7.78, 8.06, and 8.04 µg m−3, respectively, and significantly outperformed persistence, seasonal persistence, and climatology. For observations above 25 µg m−3, RMSE increased to 11.64–16.18 µg m−3 and 88.2–98.3% of cases were underpredicted, revealing a systematic upper-tail limitation. Split-conformal intervals calibrated on 2023 achieved 89.11–90.43% coverage at the 90% nominal level and 95.15–95.51% at the 95% level. Under strict leave-one-station-out transfer, CatBoost reduced RMSE relative to persistence by 15.48%, 31.36%, 34.36%, and 19.13%, but failed to outperform persistence for Tumbaco at 1 h. Conformal coverage at unseen stations fell to 87.12% at the 90% level for 1-h forecasts, diagnosing spatial shift. SHAP and ablation analyses showed that recent PM2.5 history was the dominant information source, with meteorology providing its largest incremental benefit at 1 h. A smoothed hourly-bias calibration using 90 local days reduced aggregate zero-shot RMSE by 7.65%, 6.32%, 5.05%, and 2.94%; in Tumbaco, reductions reached 29.99%, 20.35%, 16.75%, and 12.31%. The framework offers a reproducible basis for transferable forecasting and local calibration, but the pronounced underestimation of high concentrations precludes interpreting it as a stand-alone warning system without additional extreme-event safeguards.