Practical Time Series Analysis

Chapter 10. Deep Learning for Time Series

Deep learning for time series is a relatively new endeavor, but it’s a promising one. Because deep learning is a highly flexible technique, it can be advantageous for time series analysis. Most promisingly, it offers the possibility of modeling highly complex and nonlinear temporal behavior without having to guess at functional forms—which could potentially be a game changer for nonstatistical forecasting techniques.

If you aren’t familiar with deep learning, here’s a one-paragraph summary (we’ll go into more details later). Deep learning describes a branch of machine learning in which a “graph” is built that connects input nodes to a complicated structure of nodes and edges. In passing from one node to another via an edge, a value is multiplied by that edge’s weight and then, usually, passed through some kind of nonlinear activation function. It is this nonlinear activation function that makes deep learning so interesting: it enables us to fit highly complex, nonlinear data, something that had not been very successfully done previously.

Deep learning has come into its own primarily within the past 10 years, as improvements in commercially available hardware have been coupled with massive amounts of data to enable this kind of heavy-duty model fitting. Deep learning models can have millions of parameters, so one way of understanding them is to dream up just about any graph you can think of, with all sorts of matrix multiplications and ...