Traffic Map Movie Forecasting 2019

  • Study traffic movies of entire cities
  • Predict future traffic flows by high-resolution forecasts
  • Discover rules and patterns, new insights and understanding

The Challenge

Solving sustainable mobility

The dense and growing populations of cities face a mobility and sustainability challenge. Superior predictive models could offer a solution, by empowering the planning of Smart Cities and establishing better road networks for more efficient and sustainable mobility. Improved traffic predictions are of great social and environmental value, while also advancing our general ability to capture the simple implicit rules underlying a complex system and model its future states. 

The TRAFFIC4CAST competition reaches right to the heart of that issue. It challenges competitors to understand complex traffic systems, to be able to make predictions about future states of such systems

Traffic4cast Challenge 2019

Prediction of Short-term Traffic Flows
Kicking off a series of annual competitions, this year’s data is based on 100 billion probe points covering 3 entire cities. High-resolution 5-minute samples throughout a year show trends within and across weekdays, as well as seasonal effects.
Predict traffic flow volume, heading, and speed on whole city maps looking 15 minutes into the future!

The Data

Unique high-resolution traffic map movies

The traffic map movies were derived from positions reported by a large fleet of probe vehicles. Each movie frame summarizes GPS trajectories mapped to spatio-temporal cells. Traffic forecasting can then be seen as a ‘scene completion’ task. This representation is deliberately ignorant of the underlying road network. Predictions of future traffic thus do not depend on high-quality up-to-date road maps, which often are not available – especially in the fast-evolving cities of emerging economies.

Unprecedented scope and detail

Working with HERE Technologies we can provide industrial-scale real-world data for 3 full cities throughout a year. Samples every 5 minutes capture mornings, evenings, and rush-hour. In addition, we can learn weekday and seasonal effects. Overall, the data that we will share with the scientific community is based on the unprecedented number of over 100 billion (1011) probe points. The derived movies have multiple colour channels characterizing traffic volume, speed, and direction.

Here technologies

a special partnership with Here technologies makes this competition possible

Since 1985, HERE Technologies has been dedicated to advancing the science of location intelligence. After collecting over 100 billion probe-points of real-time and historic traffic data, HERE has partnered with IARAI to bring these data to the scientific community. Through efforts like these and their record as a leader in location services, HERE reaffirms their commitment to an autonomous world for everyone.


Benchmark and Robustness

Algorithms must make predictions for all three cities, where city-dependent parameters can be applied. The independent held-out testing set thus covers a range of different socio-economic areas, types of traffic, and response to weather patterns to identify approaches that will robustly generalize to new cities in different regions of the world.

Community engagement

Traffic4cast is a novel challenge that brings together several complementary fields from machine-learning and classical traffic research. The different perspectives in this interdisciplinary community will make for lively debates and lead to new insights.

Join us to explore unanswered questions, like what metrics might best identify award the correct prediction of relevant events, such as traffic jams! We will also jointly set optional additional goals, like journey time prediction, or the identification of simple rules underlying traffic patterns.


Our core metric will be the mean squared error of all predicted pixel values, the widely accepted ‘gold standard’ in video prediction tasks. In addition, we will publish filters applied to distinguish realistic traffic predictions from crude fakes.

We will incentivize and through lively debate in the community jointly develop complementary meaningful metrics and filters. Join our discussion on ideas like using the Frechet Inception Distance from classical image classification networks or a learned loss function from a GAN discriminator.


Core Competition

The three top-ranked teams in the core competition leaderboard are honoured at NeurIPS and receive $10,000, $5,000 and $2,000 value prizes, respectively.

Bonus Tasks

Traffic4cast is a novel challenge that brings together several complementary fields from machine-learning and classical traffic research. As an interdisciplinary community, we will jointly identify alternative metrics and complementary optional tasks of value that merit extra recognition, such as journey time prediction, or the identification of simple rules underlying traffic patterns.

Winners of these complementary competitions will be honoured at NeurIPS.

  • First and second bonus prizes: a 12 month Research Fellowship at IARAI, covering stipend and expenses
  • Third to sixth prize: Complimentary registration for NeurIPS 2019 / 2020.

Key Dates

- Early May -
Beta phase goes live
- 1 June -
Main competition goes live
- 15 July -
Leaderboard goes live
- 15 October -
Submissions deadline
- 30 October -

Award announcements, invitation to Traffic4cast Symposion

- 7 December -

Traffic4cast Symposion

8/10 - 12/14 December


Imprint | Terms and conditions

Log in with your credentials


Forgot your details?

Create Account