实用时间序列分析(影印版)

Preface

1．TimeSeries：AnOverviewand aQuickHistory

The History of Time Series in Diverse Applications

Medicine as a Time Series Problem

Forecasting Weather

Forecasting Economic Growth

Astronomy

Time Series Analysis Takes Off

The Origins of Statistical Time Series Analysis

The Origins of Machine Learning Time Series Analysis

More Resources

2．FindingandWranglingTimeSeriesData

where to Find Time Series Data

Prepared Data Sets

Found Time Series

Retrofitting a Time Series Data Collection from a Collection of Tables

A Worked Example：Assembling a Time Series Data Collection

Constructing a Found Time Series

Timestamping Troubles

Whose Timestamp

Guesstimating Timestamps to Make Sense of Data

What’s a Meaningful Time Scale

Cleaning Your Data

Handling Missing Data

Upsampling and Downsampling

Smoothing Data

Seasonal Data

Time Zones

Preventing Lookahead

More Resources

3．ExploratoryDataAnalysisforTimeSeries

Familiar Methods

Plotting

Histograms

Scatter Plots

Time Series-Specific Exploratory Methods

Understanding Stationarity

Applying Window Functions

Understanding and Identifying Self-Correlation

Spurious Correlations

Some Useful Visualizations

lD Visualizations

2D Visualizations

3D Visualizations

More Resources

4．SimulatingTimeSeriesData

What’S Special About Simulating Time Series

Simulation Versus Forecasting

Simulations in Code

Doing the Work Yourself

Building a Simulation Universe That Runs Itself

A Physics Simulation

Final Notes on Simulations

Statistical Simulations

Deep Learning Simulations

More Resources

5．StoringTemporalData

Defining Requirements

Live Data Versus Stored Data

Database Solutions

SQL Versus NoSQL

Popular Time Series Database and File Solutions

File Solutions

NumPv

Pandas

Standard R Equivalents

Xarray

More Resources

6．StatisticaIModelsforTimeSeries

Why Not Use a Linear Regression

Statistical Methods Developed for Time Series

Autoregressive Models

Moving Average Models

Autoregressive Integrated Moving Average Models

Vector Autoregression

Variations on Statistical Models

Advantages and Disadvantages of Statistical Methods for Time Series

More Resources

7．StateSpaceModels for TimeSeries

State Space Models：Pluses and Minuses

The Kalman Filter

Overview

CodefortheKalmanFilter、

Hidden Markov Modds

HOW the Model Works

HOWWeFittheModel

Fitting an HMM in Code

Bayesian Structural Time Series

Code forbsts

More Resources

8．Generating and Selecting FeaturesforaTimeSeries

Introductory Example

General Considerations When Computing Features

The Nature of the Time Series

Domain Knowledge

External Considerations

A Catalog of Places to Find Features for Inspiration

Open Source Time Series Feature Generation Libraries

Domain-Specific Feature Examples

How to Select Features 0nce You Have Generated Them

Concluding Thoughts

More Resources

9．Machine LearningforTime Series

Time Series C：lassification

Selecting and Generating Features

Decision Tree Methods

Clustering

Generating Features from the Data

TemporaUy Aware Distance Metrics

Clustering Code

More Resources

10．Deep LearningforTimeSeries

Deep Learning Concepts

Programming a Neural Network

Data，Symbols，Operations，Layers，and Graphs

Building a Training Pipeline

Inspecting Our Data Set

Steps of a Training Pipeline

Feed Forward Networks

A Simple Example

Using an Attention Mechanism to Make Feed Forward

Networks More Time—Aware

CNNS

A Simple Convolutional Model

Alternative Convolutional Models

RNNS

Continuing Our Electric Example

The Autoencoder Innovation

Combination Architectures

Summing Up

More Resources

11．Measuring Error

The Basics：HoW to Test Forecasts

Model-Specific Considerations for Backtesting

When Is Your Forecast Good Enough

Estimating Uncertainty in Your Model with a Simulation

Predicting Multiple Steps Ahead

Fit Directlv to the Horizon of Interest

Recursive Approach to Distant Temporal Horizons

Multitask Learning Applied to Time Series

Model Validation Gotchas

More Resources

1 2．Performance Considerations in Fitting and Serving Time Series Models

Working with Tools Built for More General Use Cases

Models Built for Cross．Sectional Data Don’t Share”Data Across Samples

Models That Don’t Precompute Create Unnecessary Lag Between

Measuring Data and Making a Forecast

Data Storage Formats：Pluses and Minuses

Store Your Data in a Binary Format

Preprocess Your Data in a Way That Allows Yon to“Slide”Over It

Modi研ng Your Analysis to Suit Performance Considerations

Using A11 Your Data Is Not Necessarily Better

Complicated Models Don’t Always Do Better Enough

A Brief Mention of Alternative High—Performance Tools

More Resources

13．HealthcareApplications

Predicting the Flu

A Case Study of Flu in 0ne Metropolitan Area

What Is State of the Art in Flu Forecasting

Predicting Blood Glucose Levels

Data Cleaning and Exploration

Generating Features

Fitting a Model

More Resources

14．FinanciaIApplications

Obtaining and Exploring Financial Data

Preprocessing Financial Data for Deep Learning

Adding Quantities of Interest to Our Raw Values

Scaling Quantities of Interest Without a Lookahead

Formatting 0ur Data for a Neural Network

Building and Training an RNN

More Resources

15．TimeSeriesforGovernment

Obtaining Governmental Data

Exploring Big Time Series Data

Upsample and Aggregate the Data as We Iterate Through It

Sort the Data

0nline Statistical Analysis of Time Series Data

Remaining Questions

Further Improvements

More Resources

16．TimeSeriesPackages

Forecasting at Scale

Google’S Industrial In．house Forecasting

Facebook’S Open Source Prophet Package

Anomaly Detection

Twitter’s Open Source AnomalyDetection Package

Other Time Series Packages

More Resources

17．ForecastsAbout Forecasting

Forecasting as a Service

Deep Learning Enhances Probabilistic Possibilities

Increasing Importance of Machine Learning Rather Than Statistics

Increasing Combination of Statistical and Machine Learning Methodologies

More Forecasts for Everyday Life

Index