Madrid NO₂ Forecasting

Most forecasting portfolios overstate results by leaking future information into features or evaluating on random splits. I wanted the opposite: honest, reproducible time-series evaluation on official Madrid air-quality data, where every claim maps to an executable check.

An end-to-end pipeline from typed Python ingestion of Madrid Open Data (1.6M+ hourly observations, 2018 to 2025) through dbt-modeled PostgreSQL marts, into leakage-safe feature engineering, rolling-origin backtesting with a temporal embargo, and direct XGBoost models per horizon with conformal P10 to P90 prediction intervals.

• Typed Python ingestion normalizes official Madrid Open Data and Open-Meteo archives into partitioned Parquet and a PostgreSQL raw schema, with UTC normalization and explicit DST handling.
• dbt owns the analytical grain and data quality across staging, intermediate, and mart layers. Python owns features, backtesting, training, and prediction.
• Features are restricted to prediction time (lags and shifted rolling statistics only), enforced by pytest tests that fail if future data enters the features.
• Rolling-origin backtests compare direct per-horizon XGBoost models against seasonal-naive baselines, with versioned benchmark manifests protecting published metrics.

• 50% MAE improvement over the best naive baseline at 1 hour ahead, with honest, modest ~12% gains reported at 24 and 72 hours.
• A clean clone reproduces the demo path with one command, and CI runs the real PostgreSQL and dbt integration on every push.
• Forecasts ship with empirical conformal prediction intervals, reported by their worst fold rather than their average.

Add forecast-based weather features from archived weather forecasts (which would have been known at prediction time) and adaptive conformal calibration to fix undercoverage in seasonal transition folds.

• Python
• PostgreSQL
• dbt
• XGBoost
• pytest
• uv
• Docker
• GitHub Actions