Chapter 9. Machine Learning for Time Series

In this chapter, we will look at a few examples of applying machine learning methods to time series analysis. This is a relatively young area of time series analysis but one that has shown promise. The machine learning methods we will study were not originally developed for time series–specific data—unlike the statistical models we studied in the past two chapters—but they have proven useful for it.

This turn to machine learning is a shift from our previous work in forecasting in earlier chapters of this book. Up to this point we have focused on statistical models for time series forecasts. In developing such models, we formulated an underlying theory about the dynamics of a time series and the statistics describing the noise and uncertainty in its behavior. We then used the hypothesized dynamics of the process to make predictions and also to estimate our degree of uncertainty about the predictions. With such methods, both model identification and parameter estimation required that we think carefully about the best way to describe the dynamics of our data.

We now turn to methodologies in which we do not posit an underlying process or any rules about that underlying process. We instead focus on identifying patterns that describe the process’s behavior in ways relevant to predicting the outcome of interest, such as the appropriate classification label for a time series. We will also consider unsupervised learning for time series, in the ...

Get Practical Time Series Analysis now with O’Reilly online learning.

O’Reilly members experience live online training, plus books, videos, and digital content from 200+ publishers.