Forecasting and Decomposition

Rob J Hyndman
Forecasting using
4. White noise and time series decomposition

OTexts.com/fpp/2/6
OTexts.com/fpp/6/
Forecasting using R 1
Student award
To students who make the most contribution to
the class, as voted by their peers.
Contributions can be on Piazza or during the
webex online sessions.
Contribute questions, answers, comments,
code suggestions, etc.
Looking for engaged learners, not experts.
At end of course, all students to vote for best
contributor.
$100 to the top voted student and $50 to
the second student (plus t-shirts).
Time zones
All sessions are at UTC 22:00

for the entire course.
Be aware when your time zone changes.
Most of Europe changed last Sunday. So classes
are now one hour earlier for most European
residents.
Most of USA and Canada change next Sunday.
So classes will be one hour earlier for most
North American residents from next week.
Outline
1 White noise
2 Time series decomposition
3 Seasonal adjustment
4 Forecasting and decomposition
Forecasting using R White noise 4

Example: White noise
White noise
2
1
0
x
−1
−2
−3
0 10 20 30 40 50
Time
White noise
2
White noise data is uncorrelated across

1
time with zero mean and constant variance.

(Technically, we require independence as
0
well.)
x
−1
−2
−3
0 10 20 30 40 50
Time
White noise
2
White noise data is uncorrelated across

1
time with zero mean and constant variance.

(Technically, we require independence as
0
well.)
x
−1
Think of white noise as completely

−2
uninteresting with no predictable patterns.

−3
0 10 20 30 40 50
Time
r1 = 0.013
0.4
r2 = −0.163
r3 = 0.163
0.2
r4 = −0.259
ACF
0.0
r5 = −0.198
r6 = 0.064
r7 = −0.139 −0.2
r8 = −0.032
−0.4
r9 = 0.199 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
r10 = −0.240 Lag
Sample autocorrelations for white noise series.

For uncorrelated data, we would expect each
autocorrelation to be close to zero.

Sampling distribution of autocorrelations
Sampling distribution of rk for white noise data is

asymptotically N(0,1/T).
95% of √
all rk for white noise must lie within
±1.96/ T.
If this is not the case, the series is probably not
WN.
√
Common to plot lines at ±1.96/ T when
plotting ACF. These are the critical values.


95% of √
±1.96/ T.
WN.
√


95% of √
±1.96/ T.
WN.
√


95% of √
±1.96/ T.
WN.
√

Autocorrelation
Example:
0.4
T = 50 and so
0.2
critical √
values at
±1.96/ 50 =
ACF
0.0
±0.28.
−0.2
All autocorrelation
coefficients lie within −0.4
these limits, 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Lag
confirming that the

data are white noise.
(More precisely, the data cannot be
distinguished from white noise.)
Example: Pigs slaughtered
Number of pigs slaughtered in Victoria
110
100
thousands
90
80
1990 1991 1992 1993 1994 1995

0.2
ACF
0.0
−0.2
0 10 20 30 40
Lag

Monthly total number of pigs slaughtered in the

state of Victoria, Australia, from January 1990
through August 1995. (Source: Australian Bureau of
Statistics.)
Difficult to detect pattern in time plot.
ACF shows some significant autocorrelation at
lags 1, 2, and 3.
r12 relatively large although not significant.
This may indicate some slight seasonality.
These show the series is not a white noise series.


Statistics.)
lags 1, 2, and 3.


Statistics.)
lags 1, 2, and 3.


Statistics.)
lags 1, 2, and 3.


Statistics.)
lags 1, 2, and 3.


Statistics.)
lags 1, 2, and 3.

ACF of residuals
We assume that the residuals are white noise
(uncorrelated, mean zero, constant variance). If
they aren’t, then there is information left in the
residuals that should be used in computing
forecasts.
So a standard residual diagnostic is to check
the ACF of the residuals of a forecasting
method.
We expect these to look like white noise.
Dow-Jones naive forecasts revisited
ŷt|t−1 = yt−1
et = yt − yt−1
ACF of residuals
forecasts.
method.
ŷt|t−1 = yt−1
et = yt − yt−1
ACF of residuals
forecasts.
method.
ŷt|t−1 = yt−1
et = yt − yt−1
ACF of residuals
forecasts.
method.
ŷt|t−1 = yt−1
et = yt − yt−1
ACF of residuals
forecasts.
method.
ŷt|t−1 = yt−1
et = yt − yt−1
Forecasting Dow-Jones index
50
Change in Dow−Jones index
0
−50
−100
0 50 100 150 200 250 300
Day

Forecasting Dow-Jones index
0.15
0.10
0.05
ACF
−0.05
−0.15
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Lag

Example: Dow-Jones residuals
0.15
0.10
0.05
ACF
−0.05
−0.15
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Lag
These look like white noise.

But the ACF is a multiple testing problem.

Example: Dow-Jones residuals
0.15
0.10
0.05
ACF
−0.05
−0.15
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Lag
These look like white noise.

But the ACF is a multiple testing problem.

Portmanteau tests
Consider a whole set of rk values, and develop a
test to see whether the set is significantly different
from a zero set.

Portmanteau tests
from a zero set.
Box-Pierce test
h
X
Q=T rk2
k =1
where h is max lag being considered and T is
number of observations.
My preferences: h = 10 for non-seasonal data, h = 2m
for seasonal data.
If each rk close to zero, Q will be small.
If some rk values large (positive or negative), Q will be
large.
Portmanteau tests
from a zero set.
Box-Pierce test
h
X
Q=T rk2
k =1
for seasonal data.
large.
Portmanteau tests
from a zero set.
Box-Pierce test
h
X
Q=T rk2
k =1
for seasonal data.
large.
Portmanteau tests
from a zero set.
Box-Pierce test
h
X
Q=T rk2
k =1
for seasonal data.
large.
Portmanteau tests
from a zero set.
Ljung-Box test
h
X
∗
Q = T (T + 2) (T − k)−1 rk2
k =1
for seasonal data.
Better performance, especially in small samples.

Portmanteau tests
If data are WN, Q∗ has χ2 distribution with
(h − K ) degrees of freedom where K = no.
parameters in model.
When applied to raw data, set K = 0.
For the Dow-Jones example,
res <- residuals(naive(dj))
# lag=h and fitdf=K

> Box.test(res, lag=10, fitdf=0)
Box-Pierce test
X-squared = 14.0451, df = 10, p-value = 0.1709
> Box.test(res, lag=10, fitdf=0, type="Lj")
Box-Ljung test
Portmanteau tests
# lag=h and fitdf=K

Box-Pierce test
Box-Ljung test
Portmanteau tests
# lag=h and fitdf=K

Box-Pierce test
Box-Ljung test
Exercise
1 Calculate the residuals from a seasonal
naive forecast applied to the quarterly
Australian beer production data from
1992.
2 Test if the residuals are white noise.

Exercise
1 Calculate the residuals from a seasonal
naive forecast applied to the quarterly
Australian beer production data from
1992.
2 Test if the residuals are white noise.
beer <- window(ausbeer,start=1992)
fc <- snaive(beer)
res <- residuals(fc)
Acf(res)
Box.test(res, lag=8, fitdf=0, type="Lj")

Outline
1 White noise
Forecasting using R Time series decomposition 20

Time series decomposition
Yt = f (St , Tt , Et )
where Yt = data at period t
St = seasonal component at period t
Tt = trend-cycle component at period t
Et = remainder (or irregular or error)
component at period t
Additive decomposition: Yt = St + Tt + Et .
Multiplicative decomposition: Yt = St × Tt × Et .



Additive model appropriate if magnitude of

seasonal fluctuations does not vary with level.
If seasonal are proportional to level of series,
then multiplicative model appropriate.
Multiplicative decomposition more prevalent
with economic series
Logs turn multiplicative relationship into an
additive relationship:
Yt = St × Tt × Et ⇒ log Yt = log St + log Tt + log Et .







History of time series decomposition
Classical method originated in 1920s.

In R: decompose().
Census II method introduced in 1957.
Basis for modern X-12-ARIMA method.
STL method introduced in 1983. It only
allows additive decomposition.
In R: stl().
TRAMO/SEATS introduced in 1990s.


In R: decompose().
In R: stl().


In R: decompose().
In R: stl().


In R: decompose().
In R: stl().


In R: decompose().
In R: stl().


In R: decompose().
In R: stl().

Classical decomposition
90
70
data
50
30
10
5
seasonal
0
−5
−10
60
trend
50
40
30
10
remainder
0
−10
1975 1980 1985 1990 1995
time

STL decomposition
90
70
data
50
30
10
5
seasonal
0
−10 −5
60
trend
50
40
5 10
remainder
−5 0
−15
1975 1980 1985 1990 1995
time

Examples: US house sales
Sales of new one−family houses, USA
90
80
70
Total sales
60
50
40
30
1975 1980 1985 1990 1995

Sales of new one−family houses, USA
90
80
70
Total sales
60
50
40
30
1975 1980 1985 1990 1995

90
70
Data
50
30
1975 1980 1985 1990 1995
Time
5
Seasonal
0
−10
1975 1980 1985 1990 1995

90
70
Data
50
30
1975 1980 1985 1990 1995

15
5
Error
−5
−15
1975 1980 1985 1990 1995
Time

90
70
Data
50
30
1975 1980 1985 1990 1995

15
5
Error
−5
−15
1975 1980 1985 1990 1995
Time

90
70
data
50
30
10
5
seasonal
0
−10 −5
60
trend
50
40
5 10
remainder
−5 0
−15
1975 1980 1985 1990 1995
time

Time series decomposition in R
plot(decompose(hsales))
plot(stl(hsales,s.window="periodic"))
plot(stl(hsales,s.window=15))

Euro electrical equipment
Electrical equipment manufacturing (Euro area)
130
120
110
New orders index
100
90
80
70
60
2000 2005 2010

Electrical equipment manufacturing (Euro area)
130
120
110
New orders index
100
90
80
70
60
2000 2005 2010

120
100
data
80
60
10
5
seasonal
0
−10
−20
110
100
trend
90
80
6
4
remainder
2
0
−4
−8
2000 2005 2010
time
120
fit <- stl(elecequip, s.window=5)
100
data
plot(fit)
80
60
10
5
seasonal
0
−10
−20
110
100
trend
90
80
6
4
remainder
2
0
−4
−8
2000 2005 2010
time
15
10
5
Seasonal
0
−5
−10
−15
−20
J F M A M J J A S O N D
Seasonal sub-series plot of the seasonal component
15
monthplot(fit$time.series[,"seasonal"],
10
ylab="Seasonal")
5
Seasonal
0
−5
−10
−15
−20
J F M A M J J A S O N D
Seasonal sub-series plot of the seasonal component
120
100
data
80
60
10
5
seasonal
0
−5
−15
110
100
trend
90
80
10
5
remainder
0
−5
2000 2005 2010
time
120
fit <- stl(elecequip, t.window=15,
100
data
s.window="periodic", robust=TRUE)
80
plot(fit)
60
10
5
seasonal
0
−5
−15
110
100
trend
90
80
10
5
remainder
0
−5
2000 2005 2010
time
120
100
data
80
plot(fit)
60
10
5
seasonal
0
t.window
−5
controls
−15
wiggliness of
110
trend
100
trend
component.
90
80
10
s.window
5
remainder
controls
0
variation in
−5
seasonal
2000 2005 2010
component. time
120
100
data
80
plot(fit)
60
10
5
seasonal
0
t.window
−5
controls
−15
wiggliness of
110
trend
100
trend
component.
90
80
10
s.window
5
remainder
controls
0
variation in
−5
seasonal
2000 2005 2010
component. time
Outline
1 White noise
Forecasting using R Seasonal adjustment 41

Seasonal adjustment
Useful by-product of decomposition: an easy
way to calculate seasonally adjusted data.
Additive decomposition: seasonally adjusted
data given by
Yt − St = Tt + Et
Multiplicative decomposition: seasonally

adjusted data given by
Yt /St = Tt × Et

Seasonal adjustment
data given by
Yt − St = Tt + Et

Yt /St = Tt × Et

Seasonal adjustment
data given by
Yt − St = Tt + Et

Yt /St = Tt × Et

Electrical equipment manufacturing
130
120
110
New orders index
100
90
80
70
60
2000 2005 2010

Seasonally adjusted series
Seasonal adjustment in R
seasadj(obj)
where obj is the output from stl() or
decompose().
Example
plot(hsales,col="gray")
fit <- stl(hsales,s.window=15)
hsales.sa <- seasadj(fit)
lines(hsales.sa, col="red")

Seasonal adjustment in R
seasadj(obj)
where obj is the output from stl() or
decompose().
Example
plot(hsales,col="gray")
fit <- stl(hsales,s.window=15)
hsales.sa <- seasadj(fit)
lines(hsales.sa, col="red")

Outline
1 White noise
Forecasting using R Forecasting and decomposition 45

Forecasting and decomposition
Forecast seasonal component by repeating the
last year
Forecast seasonally adjusted data using
non-seasonal time series method. E.g.,
Holt’s method — next topic
Random walk with drift model
Combine forecasts of seasonal component with

forecasts of seasonally adjusted data to get
forecasts of original data.
Sometimes a decomposition is useful just for
understanding the data before building a
separate forecasting model.
last year

last year

last year

last year

last year

Seas adj elec equipment
Electrical equipment manufacturing
130
120
110
New orders index
100
90
80
70
60
2000 2005 2010

Naive forecasts of seasonally adjusted data
120
110
100
New orders index
90
80
70
2000 2005 2010

Forecasts from STL + Random walk
120
100
New orders index
80
60
40
2000 2005 2010

How to do this in R

eeadj <- seasadj(fit)

plot(naive(eeadj), xlab="New orders index")
fcast <- forecast(fit, method="naive")

plot(fcast, ylab="New orders index")

Decomposition and prediction intervals
It is common to take the prediction

intervals from the seasonally adjusted
forecasts and modify them with the
seasonal component.
This ignores the uncertainty in the
seasonal component estimate.
It also ignores the uncertainty in the
future seasonal pattern.


seasonal component.


seasonal component.

Forecasting and Decomposition

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Forecasting and Decomposition

Uploaded by

Copyright:

Available Formats

Rob J Hyndman

4. White noise and time series decomposition

All sessions are at UTC 22:00

2 Time series decomposition

4 Forecasting and decomposition

Forecasting using R White noise 4

White noise data is uncorrelated across

time with zero mean and constant variance.

White noise data is uncorrelated across

time with zero mean and constant variance.

Think of white noise as completely

uninteresting with no predictable patterns.

r10 = −0.240 Lag

Sample autocorrelations for white noise series.

Forecasting using R White noise 6

Sampling distribution of rk for white noise data is

Forecasting using R White noise 7

Sampling distribution of rk for white noise data is

Forecasting using R White noise 7

Sampling distribution of rk for white noise data is

Forecasting using R White noise 7

Sampling distribution of rk for white noise data is

Forecasting using R White noise 7

confirming that the

1990 1991 1992 1993 1994 1995

Forecasting using R White noise 9

Forecasting using R White noise 10

Monthly total number of pigs slaughtered in the

Forecasting using R White noise 11

Monthly total number of pigs slaughtered in the

Forecasting using R White noise 11

Monthly total number of pigs slaughtered in the

Forecasting using R White noise 11

Monthly total number of pigs slaughtered in the

Forecasting using R White noise 11

Monthly total number of pigs slaughtered in the

Forecasting using R White noise 11

Monthly total number of pigs slaughtered in the

Forecasting using R White noise 11

0 50 100 150 200 250 300

Forecasting using R White noise 13

Forecasting using R White noise 14

These look like white noise.

Forecasting using R White noise 15

These look like white noise.

Forecasting using R White noise 15

Forecasting using R White noise 16

Forecasting using R White noise 16

# lag=h and fitdf=K

# lag=h and fitdf=K

# lag=h and fitdf=K

Forecasting using R White noise 18

Forecasting using R White noise 19

2 Time series decomposition

4 Forecasting and decomposition

Forecasting using R Time series decomposition 20

Forecasting using R Time series decomposition 21

Forecasting using R Time series decomposition 21

Forecasting using R Time series decomposition 21

Additive model appropriate if magnitude of

Yt = St × Tt × Et ⇒ log Yt = log St + log Tt + log Et .

Forecasting using R Time series decomposition 22

Additive model appropriate if magnitude of