ayanjawed's picture
Model card: future dates, ± range, Open-Meteo licence
0a79302 verified
|
Raw History Blame Contribute Delete
4.27 kB
metadata
library_name: sklearn
tags:
  - tabular-regression
  - time-series
  - transportation
license: other

TransitPulse planning model

Predicts hourly bus ridership for one route (riders per route-hour) from information known in advance: route, date, hour, schedule and weather. Built for transport planners as a CT-378 Machine Learning course project.

Intended use

  • What-if planning for the 24 MTA bus routes it was trained on. It was trained on January to September 2025; the apps also accept later dates (up to 2027-06-30) with Open-Meteo forecast weather, but those are rough guesses, especially for March and October to December, which the model never saw.
  • Not for other routes or other cities (see Limitations).
  • The apps show each prediction with a ± 56 riders range (the validation MAE).

Model

sklearn.pipeline.Pipeline: one-hot encoding of bus_route and borough, then HistGradientBoostingRegressor (learning_rate 0.05, max_iter 500, max_leaf_nodes 63). Selected by validation MAE among models whose saved file is under 50 MB; a Random Forest was about as accurate but about 110 MB. Input columns: bus_route, borough, gtfs_n_stops, gtfs_max_stops_per_trip, sched_trips_hr, sched_headway_min, sched_trips_day, temperature_2m, precipitation, snowfall, wind_speed_10m, hour, day_of_week, is_weekend, hour_sin, hour_cos, is_am_peak, is_pm_peak, is_holiday, is_school_day.

Data (three sources plus calendar)

  1. MTA Bus Hourly Ridership: Beginning 2025, NY Open Data (dataset id gxb3-akrn), summed to one row per route per hour. Licence: the dataset page gives no licence information; it is covered by NY Open Data's terms of use.
  2. MTA GTFS static bus schedules (mta.info/developers). A 2026 snapshot, not the 2025 service that produced the ridership.
  3. Open-Meteo historical weather (New York City coordinates, one value per hour for the whole city). Licence: CC BY 4.0. Weather data by Open-Meteo.com.
  4. US federal holidays (Python holidays package) and a hand-typed, approximate NYC school calendar.

Eight two-week windows between January and September 2025, 24 routes, 64,512 route-hour rows. Missing hours are filled with 0 riders. Data are from New York because Karachi has no public passenger-count data.

Evaluation

Time-ordered split by whole days: train 2025-01-13 to 07-14, validation 07-15 to 09-04, test 09-05 to 09-28 (17 days, scored once). The final model is refit on train plus validation before testing.

MAE (riders) RMSE R2
Planning model (this model) 65.4 113.3 0.92
Baseline: training average for the route, hour and weekday/weekend 78.0 129.4 0.89

The average route-hour in the test period has 450 riders, so the error is about 14.5%. A separate history model (uses recent ridership) reaches MAE 39.0 but needs recent data, so it is not deployed. Other checks: leave-one-window-out MAE 53.0 vs baseline 68.9; unseen routes MAE 111 vs baseline 132.

Limitations

  • Only 24 routes and eight 2-week windows in one year; the test set is only 17 days, so small differences are not meaningful.
  • Some routes changed level during the year (Q58, Q10, Q65 about 35% lower; Q25 about 32% higher from summer). The model does not know about service changes; its error on these routes is about three times larger.
  • Weather helps a little, but partly because it hints at the season. It is not evidence that weather causes ridership changes.
  • Weather was one city-wide value, and the test used recorded (not forecast) weather, so real use will have higher error.
  • Schedule data is a 2026 snapshot; the school calendar is approximate; licence wording was not independently checked.
  • New York data only; no claim is made about Karachi.

How to load

import joblib
model = joblib.load("planning_model.joblib")  # needs scikit-learn 1.9.1

The input row must contain the columns listed above (see app/features.py in the project repository).