Machine Learning for Time-Series with Python

Forecast, predict, and detect anomalies with state-of-the-art machine learning methods

Ben Auffarth

• Publisher: Packt Publishing
• Year: 2021
• Print ISBN: 9781801819626
• eBook ISBN: 9781801816106
• Edition: 1
• Topic: Computer Science
• Subtopic: Artificial Intelligence (AI) & Semantics
• Index: Computer Science

Get better insights from time-series data and become proficient in model performance analysis

Key Features

• Explore popular and modern machine learning methods including the latest online and deep learning algorithms
• Learn to increase the accuracy of your predictions by matching the right model with the right problem
• Master time series via real-world case studies on operations management, digital marketing, finance, and healthcare

Book Description

The Python time-series ecosystem is huge and often quite hard to get a good grasp on, especially for time-series since there are so many new libraries and new models. This book aims to deepen your understanding of time series by providing a comprehensive overview of popular Python time-series packages and help you build better predictive systems.

Machine Learning for Time-Series with Python starts by re-introducing the basics of time series and then builds your understanding of traditional autoregressive models as well as modern non-parametric models. By observing practical examples and the theory behind them, you will become confident with loading time-series datasets from any source, deep learning models like recurrent neural networks and causal convolutional network models, and gradient boosting with feature engineering.

This book will also guide you in matching the right model to the right problem by explaining the theory behind several useful models. You'll also have a look at real-world case studies covering weather, traffic, biking, and stock market data.

By the end of this book, you should feel at home with effectively analyzing and applying machine learning methods to time-series.

What you will learn

• Understand the main classes of time series and learn how to detect outliers and patterns
• Choose the right method to solve time-series problems
• Characterize seasonal and correlation patterns through autocorrelation and statistical techniques
• Get to grips with time-series data visualization
• Understand classical time-series models like ARMA and ARIMA
• Implement deep learning models, like Gaussian processes, transformers, and state-of-the-art machine learning models
• Become familiar with many libraries like Prophet, XGboost, and TensorFlow

Who this book is for

This book is ideal for data analysts, data scientists, and Python developers who want instantly useful and practical recipes to implement today, and a comprehensive reference book for tomorrow. Basic knowledge of the Python Programming language is a must, while familiarity with statistics will help you get the most out of this book.

Preface

Time-series are ubiquitous in industry and in research. Examples of time-series can be found in healthcare, energy, finance, user behavior, and website metrics to name just a few. Due to their prevalence, time-series modeling and forecasting is crucial and it's of great economic importance to be able to model them accurately.

While traditional and well-established approaches have been dominating econometrics research and – until recently – industry, machine learning for time-series is a relatively new research field that's only recently come out of its infancy.

In the last few years, a lot of progress has been made in machine learning on time-series; however, little of this has been made available in book form for a technical audience. Many books focus on traditional techniques, but hardly deal with recent machine learning techniques. This book aims to fill this gap and covers a lot of the latest progress, as evident in results from competition such as M4, or the current state-of-the-art in time-series classification.

If you read this book, you'll learn about established as well as cutting edge techniques and tools in Python for machine learning with time-series. Each chapter covers a different topic, such as anomaly detection, probabilistic models, drift detection and adaptive online learning, deep learning models, and reinforcement learning. Each of these topics comes with a review of the latest research and an introduction to popular libraries with examples.

Who this book is for

If you want to build models that are reactive to the latest trends, seasonality, and business cycles, this is the book for you. This book is for data scientists, analysts, or programmers who want to learn more about time-series, and want to catch up on different techniques in machine learning.

What this book covers

Chapter 1, Introduction to Time-Series with Python, is a general introduction to the topic. You'll learn about time-series and why they are important, and many conventions, and you'll see an overview of applications and techniques that will be explained in more detail in dedicated chapters.

Chapter 2, Time-Series Analysis with Python, breaks down the steps for analyzing time-series. It explains statistical tests and visualizations relevant for making sense of and drawing insights from time-series.

Chapter 3, Preprocessing Time-Series, is about data treatment for time-series for traditional techniques and for machine learning. Methods such as naïve and Loess STL decomposition for seasonal and trend effects are covered, along with normalizations for values, as well as specific feature extraction techniques such as catch22 and ROCKET.

Chapter 4, Introduction to Machine Learning for Time-Series, deals with an overview of the state of the art for univariate and multivariate time-series forecasts and predictions.

Chapter 5, Forecasting with Moving Averages and Autoregressive Models, focuses on forecasting, mostly on univariate time-series (see Chapter 12, Multivariate Forecasting for multivariate time-series). Well-established traditional methods used in econometrics are introduced, explained, and applied on data sets.

Chapter 6, Unsupervised Methods for Time-Series, introduces anomaly detection, change detection, and clustering. The chapter reviews industry practices at major technology companies such as Facebook, Amazon, Google, and others, and gives practical examples for both anomaly detection and change detection.

Chapter 7, Machine Learning Models for Time-Series, reviews recent research on machine learning for time-series at institutes such as at the University of East Anglia and Monash University. Many techniques are summarized and compared throughout the chapter, and there's a practical section with many examples.

Chapter 8, Online Learning for Time-Series, introduces online learning, a topic often neglected. Online models continuously update their parameters based on latest samples, and some of them have mechanisms to deal with different kinds of drift – a common problem with time-series.

Chapter 9, Probabilistic Models for Time-Series, covers probabilistic models for time-series. This includes models with confidence intervals such as Facebook's Prophet, Markov Models, Fuzzy Models, and counter-factual causal models such as Bayesian Structural Time-Series Models as proposed by Google.

Chapter 10, Deep Learning for Time-Series, reviews recent literature and benchmarks for different tasks. The chapter explains techniques such as autoencoders, InceptionTime, DeepAR, N-BEATS, Recurrent Neural Networks, ConvNets, and Informer. Deep learning still hasn't completely caught up with more traditional or other machine learning techniques; however, the progress has been promising, and for certain applications such as multivariate predictions, deep learning techniques are emerging as the state of the art, as can be seen in competitions such as M4.

Chapter 11, Reinforcement Learning for Time-Series, gives an overview of basic concepts in reinforcement learning. It introduces techniques relevant for time-series such as bandit algorithms and Deep Q-Learning, and they are applied for a recommender system and for a trading algorithm.

Chapter 12, Multivariate Forecasting, gives practical examples for multivariate multistep forecasts of energy demand with deep learning models.

To get the most out of this book

• You should have a basic knowledge of Python to get started.
• All notebooks used in this book come with links to Google Colab, where you should be able to execute them.

Download the example code files

The code bundle for the book is also hosted on GitHub at https://github.com/PacktPublishing/Machine-Learning-for-Time-Series-with-Python. We also have other code bundles from our rich catalog of books and videos available at https://github.com/PacktPublishing/. Check them out!

Download the color images

We also provide a PDF file that has color images of the screenshots/diagrams used in this book. You can download it here: https://static.packt-cdn.com/downloads/9781801819626_ColorImages.pdf.

Conventions used

There are a number of text conventions used throughout this book.

CodeIn...

Table of contents

• Preface