Estimate a trip
How long will it take?
The regression from the 2021 notebook, running in your browser: 579 features, the original coefficients from fold 1 of its 10-fold cross-validation. Across the folds it explained 36.7% of the variance with a typical error of 9.2 minutes, so each prediction now comes with a split-conformal prediction interval that held its stated coverage on 7.5 million held-out 2019 trips.
2021 model predicts
43.0min
Midtown Center → JFK Airport, Wednesday at 5 pm
90% prediction interval: 32.0 to 68.6 min
Split-conformal, from 687,938 held-out 2019 trips picked up in Manhattan with a similar prediction (bin 10 of 10). On 688,079 other Manhattan trips in the same bin, 90.0% fell inside (95% CI 89.8% to 90.2%, resampling whole days). Solid line: the prediction. Dashed blue line: the observed median at this hour. How the intervals are built
- 2026 refit, same features
- 42.7 min
- Observed median, 5 pm
- …
- Observed median, any hour
- …
- Straight-line distance
- 12.6 mi
Why that number
A linear model is a sum. The 2021 model starts from an intercept of 22.35 minutes and adds one coefficient per active feature. Bars show what each part of your trip adds or removes.
- Weather, events, collisions+1.62
- Weekday0.00
- Hour+0.98
- Rate code+9.29
- Passengers0.00
- Pickup zone0.00
- Vendor0.00
- Drop-off zone+8.78
- Store-and-forward flag0.00
- Intercept + all terms22.35 with 4 non-zero terms43.02
Red adds minutes, blue removes them. Zones, hours and weekdays are relative to the categories the penalty set to zero. Coefficients: coursework/10-folds-linear-regression.csv, fold 1.