NYC Taxi Demand Forecast

Hourly yellow-taxi pickups across New York City, forecast one week ahead per borough and compared across three models with rolling-origin backtests. Data: NYC TLC trip records + Open-Meteo historical weather, loaded into MySQL by this repo's pipeline. Source: wirkix/nyc-taxi-demand-forecast.

Run 4e2b3b9fabe5411ab211d6f02adee90c · trained through 2026-07-31 23:00 · 168h horizon · 4 backtest folds · generated 2026-09-23

Demand history

Daily pickups by borough, with the seasonal shape the models have to learn.

Does rain move demand?

Average pickups per hour of day, split by whether it was raining in that hour (Central Park, > 0.5 mm). This is the weather signal the gradient-boosting and Prophet models get as features.

Backtests

Each fold trains on everything before a cutoff and forecasts the following week; folds are the most recent non-overlapping weeks. WAPE = total absolute error ÷ total actual trips (lower is better). A model only earns its complexity if it beats the seasonal-naive baseline.

model	Gradient boosting	Prophet	Seasonal naive (last week)
series_id
Bronx	50.7%	63.8%	56.3%
Brooklyn	28.7%	46.9%	33.7%
Manhattan	9.2%	19.1%	15.2%
Queens	13.8%	16.7%	19.5%
total	8.7%	17.9%	14.5%

Next-week forecast

TLC publishes trip data about two months behind, so the "future" week here is the one right after the latest published month. It gets scored against actuals once the next month is released.