- Tagline: Time series forecasting for environmental health
- Category: Featured Project
Context
Urban pollution forecasting is noisy and non-stationary but critical for early public health warning.
Role
Solo data scientist handling collection, feature engineering, modeling, and evaluation.
Work completed
- Built 2+ years hourly pollution + weather dataset pipeline
- Engineered lag/rolling/seasonality features and differencing for stationarity
- Combined LSTM + ARIMA for one-year horizon forecasts
Results
- RMSE around 15.9 micrograms per cubic meter
- Correlation around 0.88
- About 15% better than single-model baselines