Machine learning for time series: What works and what doesn't

Description

Time series data has many applications in industry, from analyzing server metrics to monitoring IoT signals and outlier detection. One of the most common tasks is to predict the future based on historical data.

Time series analysis has a strong mathematical background in the theory of stationary processes, but recently, machine learning methods, and in particular deep learning, have become quite popular. The flexibility and expressive power of these methods has allowed data scientists to tackle complex real-world problems.

Mikio Braun offers an overview of time series analysis with a focus on modern machine learning approaches and practical considerations, including recommendations for what works and what doesn’t. Mikio covers the mathematical underpinnings and explains how to reframe time series problems as general supervised learning problems and how to use nonlinear methods, including deep learning.

Using practical applications from the industry, Mikio shares a number of scenarios of increasing complexity and provides practical advice to get started, for example: using stacked recurrent neural networks autoencoders for extreme event forecasting at Uber, using deep learning for autoregressive demand forecasting at Amazon, and using recurrent neural networks for customer analysis at Zalando.