ny-noise
Where New York is loud, and next week by ZIP · Data and Forecasting
Two models kept apart on purpose. A spatial model explains where noise complaints concentrate on a ~250 m hex grid from 30 measured features of the surroundings; on held-out areas it explains 78% of deviance, against 66% from housing units alone. A forecast gives next week's complaints for 177 ZIP codes with 80% ranges that covered 79.8% of the holdout; it beats repeating last week and does not clearly beat exponential smoothing, and the site says so. Wake the Block is a separate decibel model over every NYC building footprint, with residents scaled to the 2020 Census, that counts who a single night-time noise would wake.
View source on GitHubKey features
- Spatial cross-validation holds out whole ~36 km² areas, so no block is scored by a model that saw its neighbours
- The forecast is raced against naive baselines in the open, including the quarters it loses
- Wake the Block counts sleepers from footprints, roof heights and homes, with a range from measured window losses
Impact
Shows where complaints concentrate and what goes with them, and how far each model can be trusted.
Tech stack
Python, pandas, scikit-learn, React, TypeScript, deck.gl, GitHub Actions