book

Practical Time Series Analysis

by Aileen Nielsen

October 2019

Beginner to intermediate

497 pages

14h

English

O'Reilly Media, Inc.

Book available

Read now

Unlock full access

Who Should Read This BookExpected BackgroundWhy I Wrote This BookNavigating This BookOnline ResourcesConventions Used in This BookUsing Code ExamplesO’Reilly Online LearningHow to Contact UsAcknowledgments
The History of Time Series in Diverse ApplicationsMedicine as a Time Series ProblemForecasting WeatherForecasting Economic GrowthAstronomyTime Series Analysis Takes OffThe Origins of Statistical Time Series AnalysisThe Origins of Machine Learning Time Series AnalysisMore Resources
Where to Find Time Series DataPrepared Data SetsFound Time SeriesRetrofitting a Time Series Data Collection from a Collection of TablesA Worked Example: Assembling a Time Series Data CollectionConstructing a Found Time SeriesTimestamping TroublesWhose Timestamp?Guesstimating Timestamps to Make Sense of DataWhat’s a Meaningful Time Scale?Cleaning Your DataHandling Missing DataUpsampling and DownsamplingSmoothing DataSeasonal DataTime ZonesPreventing LookaheadMore Resources
Familiar MethodsPlottingHistogramsScatter PlotsTime Series–Specific Exploratory MethodsUnderstanding StationarityApplying Window FunctionsUnderstanding and Identifying Self-CorrelationSpurious CorrelationsSome Useful Visualizations1D Visualizations2D Visualizations3D VisualizationsMore Resources
What’s Special About Simulating Time Series?Simulation Versus ForecastingSimulations in CodeDoing the Work YourselfBuilding a Simulation Universe That Runs ItselfA Physics SimulationFinal Notes on SimulationsStatistical SimulationsDeep Learning SimulationsMore Resources
Defining RequirementsLive Data Versus Stored DataDatabase SolutionsSQL Versus NoSQLPopular Time Series Database and File SolutionsFile SolutionsNumPyPandasStandard R EquivalentsXarrayMore Resources
Why Not Use a Linear Regression?Statistical Methods Developed for Time SeriesAutoregressive ModelsMoving Average ModelsAutoregressive Integrated Moving Average ModelsVector AutoregressionVariations on Statistical ModelsAdvantages and Disadvantages of Statistical Methods for Time SeriesMore Resources
State Space Models: Pluses and MinusesThe Kalman FilterOverviewCode for the Kalman FilterHidden Markov ModelsHow the Model WorksHow We Fit the ModelFitting an HMM in CodeBayesian Structural Time SeriesCode for bstsMore Resources
Introductory ExampleGeneral Considerations When Computing FeaturesThe Nature of the Time SeriesDomain KnowledgeExternal ConsiderationsA Catalog of Places to Find Features for InspirationOpen Source Time Series Feature Generation LibrariesDomain-Specific Feature ExamplesHow to Select Features Once You Have Generated ThemConcluding ThoughtsMore Resources
Time Series ClassificationSelecting and Generating FeaturesDecision Tree MethodsClusteringGenerating Features from the DataTemporally Aware Distance MetricsClustering CodeMore Resources

Deep Learning ConceptsProgramming a Neural NetworkData, Symbols, Operations, Layers, and GraphsBuilding a Training PipelineInspecting Our Data SetSteps of a Training PipelineFeed Forward NetworksA Simple ExampleUsing an Attention Mechanism to Make Feed Forward Networks More Time-AwareCNNsA Simple Convolutional ModelAlternative Convolutional ModelsRNNsContinuing Our Electric ExampleThe Autoencoder InnovationCombination ArchitecturesSumming UpMore Resources
The Basics: How to Test ForecastsModel-Specific Considerations for BacktestingWhen Is Your Forecast Good Enough?Estimating Uncertainty in Your Model with a SimulationPredicting Multiple Steps AheadFit Directly to the Horizon of InterestRecursive Approach to Distant Temporal HorizonsMultitask Learning Applied to Time SeriesModel Validation GotchasMore Resources
Working with Tools Built for More General Use CasesModels Built for Cross-Sectional Data Don’t “Share” Data Across SamplesModels That Don’t Precompute Create Unnecessary Lag Between Measuring Data and Making a ForecastData Storage Formats: Pluses and MinusesStore Your Data in a Binary FormatPreprocess Your Data in a Way That Allows You to “Slide” Over ItModifying Your Analysis to Suit Performance ConsiderationsUsing All Your Data Is Not Necessarily BetterComplicated Models Don’t Always Do Better EnoughA Brief Mention of Alternative High-Performance ToolsMore Resources
Predicting the FluA Case Study of Flu in One Metropolitan AreaWhat Is State of the Art in Flu Forecasting?Predicting Blood Glucose LevelsData Cleaning and ExplorationGenerating FeaturesFitting a ModelMore Resources
Obtaining and Exploring Financial DataPreprocessing Financial Data for Deep LearningAdding Quantities of Interest to Our Raw ValuesScaling Quantities of Interest Without a LookaheadFormatting Our Data for a Neural NetworkBuilding and Training an RNNMore Resources
Obtaining Governmental DataExploring Big Time Series DataUpsample and Aggregate the Data as We Iterate Through ItSort the DataOnline Statistical Analysis of Time Series DataRemaining QuestionsFurther ImprovementsMore Resources
Forecasting at ScaleGoogle’s Industrial In-house ForecastingFacebook’s Open Source Prophet PackageAnomaly DetectionTwitter’s Open Source AnomalyDetection PackageOther Time Series PackagesMore Resources
Forecasting as a ServiceDeep Learning Enhances Probabilistic PossibilitiesIncreasing Importance of Machine Learning Rather Than StatisticsIncreasing Combination of Statistical and Machine Learning MethodologiesMore Forecasts for Everyday Life

Content preview from Practical Time Series Analysis

Chapter 17. Forecasts About Forecasting

There are many good quotes about the hopelessness of predicting the future, and yet I can’t help wanting to conclude this book with some thoughts about what’s coming.

Forecasting as a Service

Because time series forecasting has fewer expert practitioners than other areas of data science, there has been a drive to develop time series analysis and forecasting as a service that can be easily packaged and rolled out in an efficient way. For example, and as noted in Chapter 16, Amazon recently rolled out a time series prediction service, and it’s not the only company to do so. The company’s model seems deliberately general, and it frames forecasting as just one step in a data pipeline (see Figure 17-1).

These forecasting-as-a-service modeling endeavors aim for a good enough general model that can accommodate a variety of fields without making terribly inaccurate forecasts. Most of them describe their models as using a mix of deep learning and traditional statistical models. However, because the service is ultimately a black box, it will be difficult to understand what could make forecasts go wrong or even to retrospectively investigate how they can be improved. This means there is a reasonably high quality level for the forecasts but probably a performance ceiling as well.

This service can be valuable for companies that need many forecasts but do not have the personnel available to generate them individually. However, for companies that have ...