Engineering Extreme Event Forecasting at Uber with Recurrent Neural Networks

Nikolay Laptev, Slawek Smyl, Santhosh Shanmugam
8 min readintermediate
--
View Original

Overview

The article discusses Uber's approach to extreme event forecasting using Recurrent Neural Networks (RNNs), specifically Long Short Term Memory (LSTM) architecture. It highlights the challenges of data sparsity during peak events and presents a new model that combines historical data with external factors to improve prediction accuracy.

What You'll Learn

1

How to implement an LSTM architecture for time series forecasting

2

Why data sparsity impacts forecasting accuracy during extreme events

3

How to incorporate external factors into forecasting models

Prerequisites & Requirements

  • Understanding of time series analysis and machine learning concepts
  • Familiarity with TensorFlow and Keras for model implementation(optional)

Key Questions Answered

What is the importance of event forecasting for Uber's operations?
Event forecasting is crucial for Uber as it helps anticipate user demand, allowing for better resource allocation and operational planning. By accurately predicting ride requests during peak times, Uber can optimize its services and improve user experience.
How does Uber's new forecasting model improve upon previous methods?
Uber's new forecasting model utilizes LSTM architecture, which allows for end-to-end modeling and better incorporation of external variables. This results in improved accuracy, achieving a 14.09% symmetric mean absolute percentage error improvement over the baseline LSTM architecture.
What challenges does Uber face in extreme event forecasting?
Uber faces challenges such as data sparsity, especially for infrequent events like New Year’s Eve, where historical data is limited. Additionally, external factors like weather and marketing changes complicate accurate predictions.

Key Statistics & Figures

SMAPE improvement
14.09%
Achieved over the baseline LSTM architecture
Percentage improvement over classical time series model
over 25%
This improvement was noted in comparison to Uber’s previous forecasting methods.

Technologies & Tools

Some links below are affiliate links. We may earn a commission if you make a purchase.

Key Actionable Insights

1
Implementing an LSTM architecture can significantly enhance forecasting accuracy for time series data.
This is particularly relevant for businesses that experience sporadic demand spikes, as it allows for better anticipation of user needs and resource allocation.
2
Incorporating external variables into forecasting models can lead to more accurate predictions.
Understanding the impact of factors like weather and local events can help businesses prepare for fluctuations in demand, improving service delivery.

Common Pitfalls

1
Relying solely on a vanilla LSTM implementation can lead to subpar performance.
This occurs because a vanilla model may not adapt well to various time series domains, resulting in poor accuracy. It's crucial to customize the model for specific use cases.

Related Concepts

Time Series Analysis
Machine Learning Techniques For Forecasting
Neural Network Architectures