Changelog
Changes made since the last print edition (2018)
- Typos fixed, and some wording improved for clarification or accuracy.
- Packages updated to latest CRAN versions.
- Code updated to ensure it works with latest CRAN packages.
Chapter 1
- Corrected description of Babylonian sheep liver forecasting (was “distribution of maggots in a rotten sheep’s liver”; now “appearance of a sheep’s liver”). Thanks to Srikanth Reddy for pointing out the error.
- Corrected statement about Emperor Constantius II, and provided footnote to source.
- Fixed date of the Vagrancy Act and provided a quote and footnote to source.
Chapter 3
- Added equation for the Bickel & Doksum (1981) transformation.
- Discussion of portmanteau tests no longer uses degrees of freedom based on model parameters, except for ARIMA models.
- Added clarification that \(m\) denotes the seasonal period in the seasonal naïve prediction interval formula.
Chapter 5
- Added description of confounder variables.
- Replaced section on confounded predictors with correlated predictors.
- Boston marathon example now restricted to data from 1924 onwards, when the course was lengthened from 24.5 to 26.2 miles. Knot locations updated accordingly.
Chapter 6
- Corrected error in description of remainders in classical decomposition: the 2009 values for the electrical equipment index were described as “large negative” but should have been described as “below 1” (since this is a multiplicative decomposition).
- Updated explanation of
t.windowands.windowarguments in STL decomposition.
Chapter 7
- Fixed error in AICc formula: the numerator of the correction term now correctly includes a factor of 2, i.e., \(\text{AIC}_c = \text{AIC} + \frac{2k(k+1)}{T-k-1}\).
- Extended the SES oil production example to show forecasts five years ahead (previously three).
Chapter 8
- Added subsection on portmanteau tests of residuals for ARIMA models.
- Added the missing third condition for AR(2) stationarity: \(\phi_2 - \phi_1 < 1\).
- Fixed error in description of AR(1) behaviour: when \(\phi_1 < 0\), the process oscillates around the mean (not “between positive and negative values”).
- Added paragraph explaining how
auto.arima()selects the orders \(p\), \(q\), \(P\) and \(Q\) for seasonal ARIMA models.
Chapter 9
- Added footnote noting that the term “dynamic harmonic regression” is also used in the literature for harmonic regression with time-varying parameters.
- Clarified how non-stationary
xregvariables are handled inauto.arima(). - Clarified statements about the role of Fourier terms in dynamic regression models.
Chapter 10
- Corrected the dimension of the unit vector \(\bm{1}\) in the structural scaling specification from \(n\) to \(m\) (the number of bottom-level series).
Chapter 12
- Added note that forecasts from Croston’s method are biased, citing Syntetos & Boylan (2001).