ml-pipeline-nyc-taxi-trip-duration

Model card: NYC taxi trip duration

Generated automatically by python -m src.pipeline on 2026-10-17.

Model details

• Task: predict how long a New York City taxi trip will take (in seconds), at the moment it starts.
• Model: LightGBM gradient-boosted trees, trained on log(1 + trip_duration); predictions are converted back to seconds. Learning rate 0.05, 64 leaves, 1566 boosting rounds (chosen by early stopping on a validation set).
• Inputs (16 features, all known when the trip starts): distance_km, manhattan_km, bearing, pickup_latitude, pickup_longitude, dropoff_latitude, dropoff_longitude, pickup_hour, pickup_dayofweek, pickup_month, is_weekend, is_rush_hour, airport_trip, passenger_count, vendor_id, store_and_fwd_flag.
• Not used (data leakage): dropoff_datetime and any speed computed from the trip duration — both reveal the answer. A data contract and a unit test enforce this.

Intended use

• Estimating typical trip times for planning, dispatch analysis and dashboards.
• A portfolio demonstration of an end-to-end, tested ML pipeline.

Out of scope: pricing or other decisions about individual customers or drivers, real-time routing, and any period or city other than New York City, January–June 2016.

Training data

• Kaggle NYC Taxi Trip Duration competition data: 1,458,644 trips, January–June 2016.
• 15,704 invalid trips removed (1.08%): durations under 1 minute or over 6 hours, zero distance, coordinates outside NYC, zero passengers.
• Models are trained on a 25% random sample of the 1,442,940 clean trips.

Evaluation

Evaluation	MAE (s)	RMSE (s)	R²	Meaning
Random split (all months)	171.0	296.9	0.799	Same days appear in training and test, as in the course version
Random split, June trips only	181.6	311.7	0.809	Same model, scored only on June: shows how hard June is

The time-based split is the honest estimate: in real use the model always predicts trips from days it has never seen. Its error is higher for two reasons: June trips are harder (trips got longer every month of the period), and the model cannot learn June-specific conditions from June itself.

Comparison of all models (test set of the random split unless noted):

model	split	mae_s	rmse_s	rmsle	r2
Improved LightGBM (tuned)	random	171.0	296.9	0.3103	0.799

Known weaknesses

Long trips are underestimated. Error by actual trip length (random-split test set):

bucket	trips	mean_actual_min	mean_pred_min	bias_min	mae_min	pct_of_trips
1-5 min	10503	3.5	4.6	1.0	1.2	14.6
5-10 min	21717	7.5	8.0	0.5	1.6	30.1
10-20 min	25269	14.2	13.9	-0.3	2.6	35.0
20-30 min	9134	24.1	22.0	-2.1	4.4	12.7
30-60 min	5009	39.1	33.9	-5.2	7.9	6.9
1-6 h	515	74.2	53.6	-20.6	21.6	0.7

• The 80% of trips under 20 minutes are predicted within 1.2–2.6 minutes on average.
• Trips over 1 hour are predicted about 20.6 minutes too short on average. Such trips are rare (under 1% of the data), and their causes (accidents, detours, waiting) are not in the data.
• Distances are straight-line (haversine) distances, not road distances.

Do not use in unusual conditions

The model has no information about weather, events, road closures or live traffic. It should not be relied on during conditions like the January 2016 blizzard. Days with snow on the ground had about 16% fewer trips than dry days, so the training data contains very few snowy trips.

The slowest days in the data:

overall_rank	date	median_speed_kmh	snow_depth_cm
1	2016-01-26	10.55	43.18
2	2016-01-25	10.82	48.26
3	2016-06-08	10.91	0.0

Reproducing

python scripts/download_data.py
python -m src.pipeline
mlflow ui --backend-store-uri sqlite:///mlflow.db     # browse every logged run

This site is open source. Improve this page.