Start free trial
Trending New Genres Authors
EnglishEnglish
EspañolSpanish
简体中文Chinese
繁體中文Chinese (Traditional)
FrançaisFrench
DeutschGerman
日本語Japanese
PortuguêsPortuguese
ItalianoItalian
한국어Korean
РусскийRussian
NederlandsDutch
العربيةArabic
PolskiPolish
हिन्दीHindi
Tiếng ViệtVietnamese
SvenskaSwedish
ΕλληνικάGreek
TürkçeTurkish
ไทยThai
ČeštinaCzech
RomânăRomanian
MagyarHungarian
УкраїнськаUkrainian
IndonesiaIndonesian
DanskDanish
SuomiFinnish
БългарскиBulgarian
עבריתHebrew
NorskNorwegian
HrvatskiCroatian
CatalàCatalan
SlovenčinaSlovak
LietuviųLithuanian
SlovenščinaSlovenian
СрпскиSerbian
EestiEstonian
LatviešuLatvian
فارسیPersian
മലയാളംMalayalam
தமிழ்Tamil
اردوUrdu
Searching...
SoBrief
Practical Time Series Analysis

Practical Time Series Analysis

Prediction with Statistics and Machine Learning
Build forecasts you can trust by catching data leaks and testing models against the future.
by Aileen Nielsen 2019 497 pages
3.77
64 ratings
Listen
Immersive
V2.3
Amazon Kindle Audible
Try Full Access for 3 Days
Unlock listening & more!
Continue

Key Takeaways

Treat every ordered axis as a history, not just a clock

Order carries information. Time series analysis looks for patterns in observations arranged along an ordered axis, then uses them to understand behavior or predict what comes next. That axis may be calendar time, the sequence of a gesture, or even wavelength in a wine spectrum. The crucial feature is meaningful order, not a clock on the wall.

Data access shaped the field. In 1662, London haberdasher John Graunt used parish death records to build early life tables, estimating how survival varied with age. His work shows that useful forecasting depends not just on clever mathematics, but on reliable records collected over time. Weather prediction likewise advanced when telegraphs made observations from distant places available together.

Analysis

The broad definition makes time series methods useful across fields that might otherwise seem unrelated. Spectra, motion traces, and patient records can all contain structure in their ordering. The historical examples also make a crucial point about infrastructure: a sophisticated model cannot recover a history that was never recorded. One caution is that ordered data are not automatically a genuine temporal process. A wavelength axis may support curve comparison, but causal claims about events unfolding in time require more than sequence alone.

Inspect the data pipeline before trusting what a timestamp claims

A timestamp needs provenance. It can mark when an event occurred, when an app recorded it, or when a server received it. A mobile food diary might assign identical times to several meals because those entries were uploaded together, not because they happened simultaneously. Time-zone information, daylight saving changes, and reporting delays can further alter what a timestamp means.

Absence needs interpretation. A missing week of email opens could mean zero opens, a broken data feed, or an employee who had not yet joined the list. These cases should not be treated alike. In the nonprofit example, the analyst aligned weekly email counts with donations, filled missing weeks with zero only after checking the data, and excluded apparent weeks before each member became active.

Analysis

This is closely related to the distinction between event time and processing time in distributed computing: systems often record when information arrives separately from when the underlying event happened. That distinction can change conclusions about behavior, service delays, or treatment effects. Missingness can carry meaning too. In healthcare, for instance, a test may be skipped because a clinician considers a patient stable. Treating such a skipped test as a random gap can erase the decision process that produced the data.

Make every predictor available at the moment its forecast is made

Future information leaks backward. Lookahead is any information used in a model before it could realistically have been known. In the nonprofit example, a member's current status cannot safely explain donations from years earlier if that status was assigned later. Likewise, if email openings are summarized for a whole week, that summary cannot predict a donation made midweek without risking knowledge of events that occurred afterward.

Preprocessing can leak, too. Centering values using the mean of the complete data set, smoothing with observations on both sides of a point, and selecting model settings after inspecting later test periods can all give a model an unfair preview. The danger is especially high when data arrives late or models take time to train. Audit each shift, imputation, and transformation against the actual production timeline.

Analysis

The issue is not merely a programming bug. It is a mismatch between the information set used in evaluation and the information set available in practice. This is central to causal inference as well: conclusions depend on what was knowable before an outcome occurred. A useful operational habit is to record data availability times, not just event timestamps, and test the pipeline with small hand-checked examples. A surprisingly excellent feature deserves scrutiny before celebration, since leakage often looks like genuine predictive power.

Plot changes and seasonal cycles before trusting a correlation

Start with the shape. A time plot can reveal trends, recurring seasonal patterns, shifts in measurement, or missing stretches. The airline passenger series, for example, rises over time and shows larger seasonal swings as its level grows. A scatter plot alone may hide those dynamics. For European stock indexes, correlations between daily changes appeared strong when measured at the same time, but largely disappeared when one series was shifted to test whether it predicted the other.

Check stability and dependence. Stationarity means a series has stable statistical behavior over time, particularly in its mean and variance. Differencing can remove some trends, while transformations such as logarithms can reduce changing variance. Autocorrelation describes how observations relate across time lags; partial autocorrelation helps distinguish direct lag relationships from correlations passed along through intervening observations. Tests inform judgment, but neither plots nor tests settle every case.

Analysis

The warning about spurious correlation is especially valuable because trend makes unrelated series move together. This is a form of shared-time confounding: two curves may rise together because both respond to broad change, not because one drives the other. Differencing can expose whether a relationship survives after removing levels, but it can also discard meaningful long-run structure. The right diagnostic depends on the question. Prediction, causal explanation, and description are different goals, even when they begin with the same plot.

Choose the simplest model that captures the process you can defend

Statistical models make their assumptions visible. An autoregressive model predicts from a series' own past values; a moving-average model uses past shocks or errors. ARIMA combines these ideas with differencing to address nonstationarity. The Box-Jenkins approach iterates between inspecting the data, fitting a model, and checking whether its residual errors still contain temporal patterns. Parsimony matters: in an example with simulated data, an ARIMA(2,0,1) described the process well, while adding complexity brought little improvement.

Use state models when the mechanism matters. A Kalman filter estimates an unseen state from noisy measurements and known dynamics, much like estimating a spacecraft's position from imperfect sensors. Hidden Markov Models infer transitions among discrete hidden states, such as possible market regimes. These methods can incorporate useful knowledge, but extra flexibility brings more parameters, computational cost, and opportunities to overfit.

Analysis

Model choice is a form of scientific argument. ARIMA says that a small set of temporal dependencies is an adequate description; a state space model says the observed measurements arise from an underlying process that can be specified. Those assumptions can help interpretation, but they are not guaranteed truths. A useful model is one whose assumptions survive checks against the domain and whose residual patterns are acceptable. Mathematical elegance does not excuse a model that is poorly matched to the decision it is meant to support.

Compress time series into features only when the lost order is affordable

Features make familiar algorithms usable. A feature is a summary such as a trend, amplitude, variability, or number of peaks. These summaries can turn a long series into a manageable row of inputs for a decision tree or other machine learning method. In an EEG example, a few statistical features helped classify different recording types. But automated tools can create thousands of candidates, including irrelevant or unstable ones. The book notes that Catch22 reduced a feature set from more than 4,000 to 22, cutting computation by a factor of 1,000 with a 7% reduction in classification accuracy in the reported task.

Shape sometimes matters more than summary. Dynamic time warping compares curves while flexibly aligning their sequences, useful when similar patterns occur at different speeds. On handwritten-word profiles, clustering with this method scored 0.828 in homogeneity, compared with 0.508 for the selected features, though calculating pairwise distances was much more expensive.

Analysis

Feature engineering is a compression problem: it trades raw detail for a representation that a model can handle. The right amount of compression depends on what distinguishes the classes. If the timing of a peak matters, a histogram may destroy the signal; if the overall distribution matters more than peak location, the same histogram may help. This resembles representation learning in vision, where a model must preserve task-relevant invariants while ignoring nuisance variation. Feature importance should therefore be tested against the intended decision, not treated as an abstract ranking.

Use deep learning as a candidate, not as a synonym for better

Results depend on the data. Neural networks can represent complicated nonlinear patterns without requiring the analyst to specify every rule. Recurrent networks process sequences step by step, while convolutional networks can identify local patterns. Yet the electricity example produced a test correlation of about 0.66 for a feed-forward network, 0.56 for a simple convolutional model, and 0.36 for the tested recurrent model. A hybrid that combined a convolutional and recurrent network with an autoregressive component reached about 0.76.

The hybrid mattered. The comparison suggests that preserving a simple statistical signal alongside a flexible neural model can be effective. It does not establish that this architecture is best for other data. Deep learning still requires careful scaling, time-aware preparation, sufficient data, tuning, and comparison with simpler baselines. Its ability to fit complex patterns does not make it immune to assumptions or leakage.

Analysis

The uneven results are a useful corrective to architecture worship. Model families encode different biases: a feed-forward network can exploit cross-variable patterns, while recurrence imposes a particular way of carrying information through sequence. Neither bias is universally appropriate. The hybrid result also reflects a broader principle in statistical learning: combining a strong low-complexity component with a flexible residual learner can stabilize predictions. But one benchmark is not a universal ranking. Small changes in data, preprocessing, or evaluation can reverse the apparent winner.

Evaluate forecasts by moving through time, not by shuffling observations

Test the future as it would arrive. Random train-test splits can let later conditions influence model selection for earlier periods. Instead, backtesting repeatedly fits on earlier data and evaluates on the next period, then moves the cutoff forward. Training windows can expand when older history remains relevant, or move when behavior is changing. Models should also be timestamped so an evaluation never uses a model before it could have been trained.

Demand a useful baseline. Compare forecasts with a simple temporal rule, such as predicting that the next value equals the latest observed value. Inspect errors over time, test important outliers, and use a metric tied to the real cost of mistakes. A blood-glucose model that predicts ordinary values reasonably well may still fail its vital purpose if it misses dangerously low readings.

Analysis

Chronological evaluation is a practical response to distribution shift, the possibility that the data-generating process changes between training and deployment. It is not enough to preserve order if the test periods omit the conditions that matter most. A public-health system may care disproportionately about epidemic peaks, while a glucose-warning tool must prioritize dangerous lows. Evaluation metrics are therefore policy choices as well as statistical summaries. The clearest score is not necessarily the most relevant one.

Design storage and computation around the decisions the data must support

Data has a lifecycle. High-volume event streams may be valuable as short-lived raw records but more useful long term as aggregates. A web service, for example, may need minute-level page-load data during an incident, then only hourly summaries for ongoing monitoring. Repeated state values may not need repeated storage, and stale data may be safely removed if its future use is understood. Databases, flat files, and specialized time-series systems each trade flexibility, speed, and ease of maintenance.

Choose by workload, not fashion. Time-series systems often prioritize rapid writes, chronological access, and bulk deletion. Binary files can improve storage and input speed for repeated model training, while SQL remains useful when series must join with related business tables. For huge government files, reading only the needed columns can be enough: the book processed selected fields from a 20-million-row NYC 311 dataset on a modest laptop.

Analysis

Computational efficiency is part of analytical validity. If a pipeline cannot be rerun, audited, or updated within the decision window, its theoretical accuracy has limited practical value. Data retention also has ethical and legal dimensions, particularly for health or location records: keeping everything is not a neutral choice. The 311 example shows that dataset size is not the same as task size. A focused question, a narrow column selection, and an appropriate summary can often outperform a costly effort to ingest the entire universe.

Analysis

The book's strongest claim is practical rather than algorithmic: time series analysis is a discipline of respecting information order. That principle governs timestamp interpretation, imputation, feature construction, train-test design, and deployment. The sophisticated model is downstream of these choices. A model can be mathematically correct and still answer a fictional question if its inputs contain future knowledge or if its timestamps describe upload time rather than event time.

The book also resists a common false choice between classical statistics and machine learning. ARIMA offers transparent short-memory structure; state space methods represent latent dynamics and noisy measurement; tree models and neural networks can capture nonlinear patterns; ensembles can combine imperfect forecasts. The electricity case makes the point concretely: a hybrid outperformed the tested standalone networks, but the result is a lesson in empirical comparison, not proof that hybrids always win.

A useful extension is to treat a forecasting pipeline as a measurement system with an explicit information contract. For each prediction time, specify which events have occurred, which records have arrived, which transformations have completed, and which model version is available. This resembles causal inference's discipline of defining what precedes an outcome, and distributed computing's distinction between event time and processing time. Making that contract explicit makes leakage easier to detect and model performance easier to reproduce.

There is also a productive tension between automation and expertise. Automated packages make forecasts accessible and consistent at scale, but they cannot automatically know whether an extreme glucose value is a device error, whether an unexpected flu peak is plausible, or whether a missed observation reflects clinical judgment. More data and computation do not eliminate the need for domain reasoning. The durable lesson is to automate repeatable mechanics while keeping consequential assumptions visible, testable, and matched to the cost of being wrong.

Summary in 30 Seconds
Time series depend on meaningful order, so verify what timestamps and missing values represent before modeling. Prevent leakage by using only information available at prediction time, including during preprocessing. Plot trends and seasonal patterns, then compare simple statistical models with machine learning and deep learning. A reported feature set cut from over 4,000 variables to 22 reduced computation 1,000-fold with a 7% accuracy loss. Backtest chronologically and compare with a simple baseline.
Contains spoilers

Last updated:

Report Issue

Review Summary

3.77 out of 5
Average of 64 ratings from Goodreads and Amazon.

Practical Time Series Analysis receives mixed reviews. Some readers find it a good introductory book covering basics and best practices, while others criticize it for lacking practical examples and containing errors. Positive aspects include its overview of techniques, data sources, and references. Critics note poor modeling results and undigested material. Some readers suggest it's better as a reference companion to hands-on practice. The book's mathematical explanations are considered weak, with readers recommending online alternatives for deeper understanding. Overall, it's seen as a broad overview rather than a practical guide.

Your rating:
4.31
200 ratings
Want to read the full book?

Glossary

found time series

series assembled from existing data

A found time series is a time series assembled from data that was collected for another purpose, rather than designed specifically for temporal analysis. For example, transaction records with timestamps can be aggregated into weekly customer activity. Constructing one requires checking what timestamps mean, aligning different frequencies, and deciding how to represent periods with no recorded events.

lookahead

using future information too early

Lookahead occurs when a model, preprocessing step, or evaluation uses information that would not have been available at the time of a real prediction. It can enter through future observations used for smoothing or imputation, later data used to select model settings, or timestamps that conceal reporting delays. It commonly makes test results appear better than production performance.

stationarity

stable statistical behavior over time

A stationary series has statistical properties that do not systematically change with time. In the weak form emphasized in the book, its mean and variance remain stable, and its dependence structure is time-invariant. Many classical forecasting models assume stationarity. Trends, changing variance, and fixed seasonal patterns can violate that assumption, sometimes requiring transformations or differencing.

autocorrelation function

correlation across time lags

The autocorrelation function, or ACF, measures the correlation between observations in a series separated by a specified lag. Its pattern can reveal persistence, seasonality, or dependence across time. ACF plots are used in exploratory analysis and model diagnosis, but trends and nonstationarity can make their apparent correlations difficult to interpret.

partial autocorrelation function

direct correlation at each lag

The partial autocorrelation function, or PACF, estimates the relationship between observations at a given lag after accounting for the intervening lags. It can help distinguish a direct dependence at a particular lag from correlation that is transmitted through shorter lags. Analysts often inspect it when selecting the order of an autoregressive model.

ARIMA

autoregression with differencing and errors

An autoregressive integrated moving average model combines dependence on earlier values, dependence on earlier forecast errors, and differencing to address nonstationarity. Its order is written as (p, d, q), where the terms specify autoregressive order, differencing count, and moving-average order. ARIMA is a widely used statistical forecasting method, particularly for relatively small datasets.

state space model

hidden dynamics behind observations

A state space model separates an underlying system state from the measurements that reveal it imperfectly. It specifies how the state changes over time and how observations depend on that state. This structure supports filtering, forecasting, and smoothing, and can represent measurement error, external forces, and time-varying behavior.

dynamic time warping

shape comparison with flexible alignment

Dynamic time warping compares two ordered series by finding an alignment that minimizes differences between their values while allowing some points to match multiple points. It can compare curves of unequal length or patterns that unfold at different speeds. The method is useful for shape-based classification and clustering, though pairwise computation can be expensive.

backtesting

testing models on rolling past periods

Backtesting evaluates a forecasting process by repeatedly fitting a model on data available up to a historical cutoff and testing it on a later period. The cutoff then moves forward and the process is repeated. This preserves the order in which information would arrive and provides a more realistic estimate of deployment performance than random splitting.

FRESH

feature selection by hypothesis tests

FRESH stands for feature extraction based on scalable hypothesis tests. It evaluates generated features against a target using statistical tests, then applies a multiple-testing procedure to control false discoveries when deciding which features to retain. The approach is designed to make large-scale time-series feature selection more manageable, but its output still depends on the relevance and quality of the candidate features.

P-square algorithm

streaming quantile estimation

The P-square algorithm estimates selected quantiles from a stream of observations without storing the complete dataset. It maintains a small set of markers representing quantile positions and values, updating them as new observations arrive. The book modifies the method by discounting older observations so estimates can adapt when the distribution changes over time.

FAQ

1. What is Practical Time Series Analysis: Prediction with Statistics and Machine Learning by Aileen Nielsen about?

  • Comprehensive time series guide: The book provides a thorough overview of time series analysis, covering the entire workflow from data collection and preprocessing to modeling and forecasting.
  • Bridges traditional and modern methods: It uniquely combines classical statistical models (like ARIMA) with modern machine learning and deep learning techniques, offering practical advice and code examples.
  • Real-world applications: Case studies span domains such as healthcare, finance, and government, demonstrating the versatility of time series methods.
  • Focus on practical implementation: The author emphasizes hands-on coding, performance considerations, and common pitfalls, making it a valuable resource for practitioners.

2. Why should I read Practical Time Series Analysis by Aileen Nielsen?

  • Fills a knowledge gap: Time series analysis is often underrepresented in data science resources; this book provides a centralized, accessible guide.
  • Balanced theory and practice: It offers both mathematical foundations and practical coding examples, suitable for readers who want to understand both the "why" and "how."
  • Addresses real-world challenges: The book discusses issues like data leakage, lookahead bias, and causality, helping readers avoid common mistakes.
  • Covers modern techniques: It introduces deep learning models and advanced machine learning methods, which are often missing from traditional time series texts.

3. What are the key takeaways from Practical Time Series Analysis by Aileen Nielsen?

  • End-to-end workflow: The book covers the full pipeline, from data handling and preprocessing to modeling, evaluation, and deployment.
  • Model diversity: Readers learn about statistical models, machine learning, and deep learning approaches, understanding their strengths and limitations.
  • Feature engineering importance: The book highlights the critical role of domain-informed feature generation and selection in time series modeling.
  • Practical pitfalls and solutions: It provides actionable advice on avoiding lookahead bias, handling missing data, and optimizing performance for real-world applications.

4. What background knowledge is recommended for reading Practical Time Series Analysis by Aileen Nielsen?

  • Coding familiarity: Some experience with R and Python, especially libraries like NumPy, Pandas, scikit-learn, and data.table, is helpful.
  • Statistics and ML basics: Understanding of variance, correlation, probability distributions, clustering, decision trees, and neural networks is recommended.
  • Supplementary resources: The book offers brief overviews and links to free online tutorials for readers needing to strengthen foundational knowledge.
  • No advanced prerequisites: While some background is useful, the book is designed to be accessible to data scientists and supervisors alike.

5. How does Practical Time Series Analysis by Aileen Nielsen define and explain time series data?

  • Chronological data focus: Time series analysis is about extracting meaningful information from data points arranged in chronological order.
  • Broad interpretation of time: The book notes that time series methods apply to any ordered data, not just temporal, such as frequency-ordered spectrographs.
  • Causality and trends: Time series uniquely address questions of causality, trends, and future event likelihood, which cross-sectional data cannot.
  • Importance in forecasting: The book emphasizes the role of time series in diagnosing past behavior and predicting future outcomes.

6. What are the main statistical models for time series covered in Practical Time Series Analysis by Aileen Nielsen?

  • Autoregressive (AR) models: These predict future values based on past observations, with detailed explanations of AR(1) and AR(p) processes.
  • Moving Average (MA) models: MA models use past error terms to model current values, with guidance on parameter selection and limitations.
  • ARIMA and extensions: The book explains ARIMA models, which combine AR and MA components with differencing to handle nonstationarity, and discusses automated fitting methods.
  • State space models: It covers Kalman filters, Hidden Markov Models (HMMs), and Bayesian Structural Time Series (BSTS) for modeling hidden states and system dynamics.

7. How does Practical Time Series Analysis by Aileen Nielsen explain stationarity and its importance?

  • Stationarity defined: A stationary time series has stable statistical properties (mean, variance) over time; nonstationary series exhibit trends or changing variance.
  • Testing for stationarity: The book introduces the Augmented Dickey–Fuller (ADF) test and explains mathematical conditions for stationarity in AR models.
  • Modeling implications: Many time series models require stationarity; the book discusses transformations like differencing and log transforms to achieve it.
  • Weak vs. strong stationarity: It clarifies the difference, with weak stationarity requiring constant mean and variance, and strong stationarity demanding the entire distribution remains unchanged.

8. What are the key machine learning and deep learning methods for time series in Practical Time Series Analysis by Aileen Nielsen?

  • Tree-based models: Random forests and gradient boosted trees (e.g., XGBoost) are used for classification and forecasting, with practical examples.
  • Clustering and distance metrics: The book covers clustering with feature-based and raw time series distances, introducing dynamic time warping (DTW) and alternatives to Euclidean distance.
  • Deep learning architectures: It discusses feedforward networks, convolutional neural networks (CNNs), recurrent neural networks (RNNs), and attention mechanisms, explaining their suitability for temporal data.
  • Hybrid and advanced models: The book explores hybrid models like LSTNet, combining convolutional, recurrent, and autoregressive components for complex patterns.

9. How does Practical Time Series Analysis by Aileen Nielsen approach feature generation and selection for time series?

  • Domain knowledge emphasis: The book stresses the importance of features informed by domain expertise, such as time-of-day or seasonal effects.
  • Automated feature extraction: Tools like tsfresh, Cesium, and tsfeatures are reviewed for generating thousands of statistical and periodicity features.
  • Time-aware features: Lagged values, rolling statistics, and categorical time indicators are recommended to capture temporal dynamics.
  • Feature selection techniques: Methods like recursive feature elimination and FRESH are discussed to identify the most predictive features and avoid overfitting.

10. What are the main challenges and solutions for handling missing data, lookahead bias, and model validation in Practical Time Series Analysis by Aileen Nielsen?

  • Missing data strategies: The book covers imputation methods like forward fill, moving averages, and interpolation, with cautions about their pitfalls.
  • Lookahead bias warning: It emphasizes avoiding future information leakage into training or evaluation, which can invalidate models.
  • Time-aware validation: Rolling window backtesting is recommended to mimic real forecasting scenarios, rather than random splits that violate temporal order.
  • Model-specific considerations: The book notes that even models trained chronologically can suffer from data leakage if not carefully managed.

11. How does Practical Time Series Analysis by Aileen Nielsen address performance optimization and practical implementation?

  • Data overlap inefficiency: The book discusses how overlapping time windows can inflate dataset size and suggests non-overlapping windows or generator-based feeding.
  • Precomputation for speed: Precomputing parts of models, such as unrolling RNNs for known past data, can reduce prediction latency in real-time applications.
  • Efficient data storage: Using binary formats (e.g., pickling in Python) improves I/O speed and reduces storage size compared to CSVs.
  • Production considerations: The book prepares readers for real-world challenges, including performance bottlenecks and deployment issues.

12. What tools, packages, and resources does Practical Time Series Analysis by Aileen Nielsen recommend for time series forecasting and anomaly detection?

  • Automated forecasting frameworks: Google’s in-house system and Facebook’s Prophet package are highlighted for handling seasonality, holidays, and level changes.
  • Anomaly detection tools: Twitter’s AnomalyDetection package is recommended for identifying outliers in seasonal time series.
  • Feature extraction and modeling: The book points to R and Python packages like tsfeatures, tsfresh, and bsts for feature extraction and state space modeling.
  • Encouragement to explore: Readers are encouraged to leverage the extensive R and Python ecosystems for their time series projects.

About the Author

Aileen Nielsen is the author of Practical Time Series Analysis. As an expert in time series analysis, Nielsen aims to provide readers with a comprehensive overview of modern techniques used in data science and related fields. Her approach focuses on presenting big-picture concepts and addressing common problems in time series analysis. Nielsen's writing style is geared towards giving readers a general understanding of the subject matter, though some reviewers note that the book lacks in-depth practical examples. Despite this, Nielsen's work is recognized for its coverage of important topics in time series analysis and its potential value as a reference for practitioners in the field.

Download PDF

To save this Practical Time Series Analysis summary for later, download the free PDF. You can print it out, or read offline at your convenience.
Download PDF
File size: 0.39 MB     Pages: 21

Download EPUB

To read this Practical Time Series Analysis summary on your e-reader device or app, download the free EPUB. The .epub digital book format is ideal for reading ebooks on phones, tablets, and e-readers.
Download EPUB
File size: 1.34 MB     Pages: 28
Want to read the full book?
Follow
Listen
Home
Swipe
Library
Get App
Try Full Access for 3 Days
Listen, bookmark, and more
Compare Features Free Pro
📖 Read Summaries
Read unlimited summaries. Free users get 3 per month
🎧 Listen to Summaries
Listen to unlimited summaries in 40 languages
—
❤️ Unlimited Bookmarks
Free users are limited to 4
—
📜 Unlimited History
Free users are limited to 4
—
📥 Unlimited Downloads
Free users are limited to 1
—
Risk-Free Timeline
Today: Get Instant Access
Listen to full summaries of 26,000+ books. That's 12,000+ hours of audio!
Day 2: Trial Reminder
We'll send you a notification that your trial is ending soon.
Day 3: Your subscription begins
You'll be charged on Nov 24,
cancel anytime before.
Consume 2.8× More Books
2.8× more books Listening Reading
Our users love us
1M+ readers
Trustpilot Rating
TrustPilot
4.6 Excellent
This site is a total game-changer. I've been flying through book summaries like never before. Highly, highly recommend.
— Dave G
Worth my money and time, and really well made. I've never seen this quality of summaries on other websites. Very helpful!
— Em
Highly recommended!! Fantastic service. Perfect for those that want a little more than a teaser but not all the intricate details of a full audio book.
— Greg M
Save 62%
Yearly
$119.88 $44.99/year/yr
$3.75/mo
Monthly
$9.99/mo
Start a 3-Day Free Trial
3 days free, then $44.99/year. Cancel anytime.
Unlock a world of fiction & nonfiction books
26,000+ books for the price of 2 books
Read any book in 10 minutes
Discover new books like Tinder
Request any book if it's not summarized
Read more books than anyone you know
#1 app for book lovers
Lifelike & immersive summaries
30-day money-back guarantee
Download summaries in EPUBs or PDFs
Cancel anytime in a few clicks

Why one summary?

Scanner
Find a barcode to scan

⭐
⭐
⭐
We have a special gift for you
Open
38% off
DISCOUNT FOR YOU
⭐
⭐
⭐
$79.99
$49.99/year
Only $4.16 per month
Continue
2 taps to start, super easy to cancel
Settings
General
Widget
Loading...