Skip to main content
Data / ML

Engineering Extreme Event Forecasting at Uber with Recurrent Neural Networks

9 June 2017 / Global
Featured image for Engineering Extreme Event Forecasting at Uber with Recurrent Neural Networks
Figure 1: The scaled number of trips taken over time in a city is part of the historical data backlog used to train our model. Notice a dip during NYE and then a sharp spike, indicating people taking rides home with Uber on New Year’s Day.
Figure 2: Our model was trained using a combination of exogenous variables, including weather (e.g., precipitation, wind speed, and temperature forecasts) and city-level information (e.g., trips in progress at any given time within a specific geographic area, registered Uber users, and local holidays or events).
Figure 3: The X and Y sliding windows consist of batch, time, features (for X) and forecasted features (for Y).
Figure 4: Our model is composed of manually derived time series features (left) and our proposed LSTM architecture with an automatic feature extraction model (right).5
Figure 5: The described model is trained offline and then exported to the target language for native execution.
Figure 6: Our new forecasting model dramatically outperformed our previous one.
Figure 7: A mock-up of the number of completed trips in one city over 200 days and our forecasts for that same data highlight our new model’s accuracy.
Slawek Smyl

Slawek Smyl

Slawek Smyl is a forecasting expert working at Uber. Slawek has ranked highly in international forecasting competitions. For example, he won the M4 Forecasting competition (2018) and the Computational Intelligence in Forecasting International Time Series Competition 2016 using recurrent neural networks. Slawek also built a number of statistical time series algorithms that surpass all published results on M3 time series competition data set using Markov Chain Monte Carlo (R, Stan).

Santhosh Shanmugam

Santhosh Shanmugam

Santhosh Shanmugam works as a senior data scientist on Uber's Marketplace team.

Posted by Nikolay Laptev, Slawek Smyl, Santhosh Shanmugam

Category: