Time Series Analysis: Forecasting and Control (Forecasting & control) [Hardcover]

In the early 1970s I was working on practical forecasting methods to apply to the U.S. Army supply depot workloads. Exponential smoothing was the commonly used "automatic" technique (once smoothing constants have been determined) that had great advantages over the informal methods used by the Army. Then someone told me that Box-Jenkins techniques were more general and powerful. I got a copy of the first edition published in 1970 and found that I could read and understand it even though I had little statistical training. I had a bachelors degree in mathematics. I got to appreciate the book even more when I took a short course from George Box, George Tiao and David Pack based on the book. I began to grasp some of the key ideas of stationary and nonstationary time series and learned about model selection, diagnostic checking and estimation. This started my interest in becoming a statistician and gave me the practical side of time series analysis first. I later specialized in it and got a Ph.D. in statistics.
Gwilym Jenkins died many years prior to this edition and Box's colleague Greogory Reinsel took on the task of helping to revise and update it.

It retains its original flavor. It is an applied book with many practical and illustrative examples. It concentrates on the three stages of time series analysis: modeling building, selection, estimation and diagnostic checking and how to iterate the process toward a good solution. The ARIMA time series models are what are considered. The theory of stationary and nonstationary time series is introduced to motivate interpretation of autocorrelation and partial autocorrelation in the model identification phase. Operator notation is introduced and used throughout the book to simplify equations. For me it helped simplify things and illuminate some concepts. But many readers found it difficult and confusing. the book is very systematic and practical. Many of the examples are real examples from Box's work in the chemical industry and his consulting during his career at the University of Wisconsin and also the consulting experience of Gwilym Jenkins in England.

The publishers and some amazon reviewers say that this edition is a major revision. The second edition published in 1976 was criticized for being essentially a reprint of the first. Although there is a new chapter 12 on intervention analysis and outlier detection it mainly is an expansion of ideas already discussed in the first edition. Theoretical results are kept aside in appendices as in previous editions.

This is not an up-to-date text on the theory of time series. It deals strictly with the time domain approach and does not include recent advances including nonlinear and bilinear models, models with non-Gaussian innovations and bootstrap or other resampling methods.

To get a balanced approach that includes the theory for frequency and time domain approaches the book by Shumway, the latest edition of the Brockwell and Davis text and the latest edition of Fuller's text are appropriate. For a graduate course I taught at UC Santa Barbara in 1981 I used the first edition of Fuller's book. Anderson provides a thorough account of the time domain theory. Excellent texts that specialize in the frequency domain approach are Bloomfield's second edition and the two volume book by Priestley. Brillinger's text is also worthwhile for those interested in spectral theory (frequency domain statistics).

Although there are many things that is text does not cover, it remains the classical text on a rich class of time domain methods that are still very practical. This is a text I bought for reference even though I still have the first edition.

Box-Jenkins is THE definitive, foundational text in time series analysis. Mastery of this volume requires extensive graduate level understanding of mathematical statistics. While difficult even for intermediate statistical practitioners, this text is necessary for any professional who examines time series data and well worth the considerable effort to acquire mastery.

I read the first edition of this book, it is very clear and it is plenty of insights. My only concern is that the space given to non-linear time series in this version (3ed) might no be sufficient; any way, for me, it is a six star book.