How long does a NYC taxi ride take?
Yellow-cab trips from the first half of 2016 (Kaggle's NYC Taxi Trip Duration data), cleaned, explored and modelled. Every number here is read from the pipeline's own output.
Trips per day
Cleaned trips by pickup day.
How long trips take
Share of cleaned trips by duration.
Traffic by hour and weekday
Duration distribution
Trips in two-minute bins; the last bar holds everything over an hour.
Where the trips are
Real streets underneath, trips on top. The dots are random trips coloured by how long they took; Density and Speed count every cleaned trip in 75 m cells.
Predict a trip
The trained LightGBM model runs in your browser: no server, the same features and trees as the Python pipeline (tests check they agree).
Model
Models on the same June test set
Test RMSLE; every model sees the same 25 features.
Other ways to score it
Error by trip length
Mean absolute error in minutes, June test trips.
Error by pickup hour
Mean absolute error in minutes, June test trips.
What the model relies on
Data quality
Rows removed, by rule
A row can break more than one rule.
Passengers per trip
Cleaned trips.
Why these rules
- Meter errors, not trips. Some raw trips last days (one runs for over 22 days). Trips over 3 hours or under 1 minute are dropped from training and the clean test.
- Impossible geometry. Pickups or drop-offs outside New York City, trips that never move, and straight-line speeds over 100 km/h are dropped.
- Scored both ways. The Model page also scores every June trip, dirty rows included, so the cleaning cannot flatter the result.