Uber TLC

Overview

Uber pickups in New York City per taxi zone, January to June 2015, built from the raw trip records that FiveThirtyEight obtained from the NYC Taxi & Limousine Commission. Every pickup (14.3 million) is counted in its taxi zone (locationID) and hour or day, with zero for periods without pickups. There are 262 zones, so 262 series. This is the Uber TLC dataset of the Chronos paper (Ansari et al., 2024); the counts match its published copy exactly.

Two configs:

Config	Length	Frequency	Prediction length	MASE seasonality
`hourly` (default)	4,344	`h`	24	24
`daily`	181	`D`	7	1

UberTLC.SETTINGS[config] gives (frequency, prediction length, seasonality). The splits follow the standard forecasting protocol used by Chronos and GluonTS: test holds each full series and the forecast target is its last prediction-length points; train holds the same series without them.

Data Structure

Key	Type	Description
`series`	list of float32	Pickup counts per hour or day
`item_id`	str	Taxi zone, `location_001` … `location_265`
`start`	str	ISO timestamp of the first point (`2015-01-01T00:00:00`)

Usage Example

import numpy as np
from stable_datasets.timeseries import UberTLC

daily = UberTLC("daily", split="test")
_, H, season = UberTLC.SETTINGS["daily"]
y = np.asarray(daily[0]["series"])
context, target = y[:-H], y[-H:]

References

Homepage: https://github.com/fivethirtyeight/uber-tlc-foil-response
Chronos: https://arxiv.org/abs/2403.07815

Citation

@misc{fivethirtyeight2015uber,
  title={Uber TLC FOIL Response},
  author={FiveThirtyEight},
  howpublished={\url{https://github.com/fivethirtyeight/uber-tlc-foil-response}},
  year={2015}
}