Comment
forthe
moving
average
smoothing
filter onedoes
not have
enough
values ofthe
original
times series inthe
windowcentred
at time t if t < gti or t > n-g
e.g. . if 9=2 then the full filtered
output
cannot becomputed
at t- 1,2 , n - I , h s! K : '! ! ! . -i -. . . - . . . E o • o d . I 2 3 . -. n -I n
The text uses x, for times t - I and ka for times t s n . So
initially
the value x, dominates
the
output
and
atthe
end thevalue
Indominates
the
output
.Method
Exponential
Smoothing
#
'
Hee we set a
parameter
2 Eco, D . Set Tn , = K, . Then fo- t > I
int
= 2 It t C l - 2)int
... Note that
by
recursion ,Mnt
= 2kt t Cl -2) Tht
. , = a Ketel - 4) (2kt -i t Cl -2)Tht
-s)
= 2kt t dCi -4) Kt -i t C l -2) 'Tht
-z :=L
It tall -21kt
-it 2 ( i -25kt
-z t ... t 2 Cl -a) t-Sca
t Cl -2)
" K , This iscomputed
fo- t--2 , . . , n .If one had observations
going infinitely
back into thepast
then onecould continue
the
above recursion forever and obtainTnt
= 2kt tall -2) It -, t 2 ( I - 4) ' Kt-z t . . - -=£72
Cl -2) " It. -j = aj It -j ,where
aj =L Ci -a) "which
is acausal
linear filterwith
coefficients { 9J}
.
The
exponential
smoother
can bethought
of as anapproximation
to
this
causal filter.As 2 increases
the
output Tht
becomes lesssmooth
(
when
2=1hit
= It for all t ). The choice of 2 is
subjective
( as isthe
choiceof
q for
the
moving
averagesmoothing
filter) , butthe
same commentapplies
forthe
choice
of
2 :the
output
should lookreasonably
smooth
while
following
the
general
trend as seenby
eyewhen
plotting
Another
qualitative
guideline
one can use to choose asmoothing
parameter
( g or d)
is to , in addition to the previous comment,
look at the
residuals
oneobtains
by
subtracting
the estimatedtrend
fromthe
times series- if one can choose the
parameter
so that thesample
ACF ofthe
residuals indicates
i i d noisethen
this isjustification
that
this is a
good
choice
ofthe
parameter
.-more
generally
, if theresiduals
lookstationary
(
e.g. , the
sample
Acf of
the
residuals looks like the true Acf of some model for astationary
time seriesthat
we know of)
then
this
may alsobe
Treudf-liminatcoisbydiffe-en.ci#
Differencing
and
Backshift
Operators
-The
differencing
operator
is denotedby
Fp ( " nabla
")
and
itoperates
as Tf Xt = Xe - Xt . , l first differences)
. Theback
shift
operator
is denotedby
Band
itoperates
asB Xt =
Xt
-I . Powers
of
these
operators
arecompositions
of
the
operator
the number of times as inthe
power
Ci.e.,
the
operator
isapplied
K times ,where
k isthe
power)
.e.ge . D -Xt =
17117
Xt)
=D ( Xt -Xt- i)
= Xe -Xt- * -(
Xt- i -Xt - z)
= Xt - 2kt -i t Xt-zB'
Xt = B ( B Xt)
= B Xt -I = Xt -z we take Bo = I ,where
I Xt = Xt .The difference
operator
can beexpressed
as= I
-B since C l -
B)
Xt = Xt - Xt - 1
Expressions
involving
the
backshift
operator
B can bemanipulated
the
same aspolynomials
in powersof
B .e.ge
.TTXE
= C l -B)
' Xt = ( I -2B t B')
Xt =Xt
- 2 Xt -itXt
-zFor trend
elimination
, one canshow
that
if onehas
atime series
with
a trend and astationary
part
,
say
Xt
=mt t
Yt
,where
Mt isa Kth
degree
polynomial
and
{
4+3
isweakly
stationary
then
17
"me isconstant
andTHE
isweakly
stationary
.Example
If me = Co t C it is linear ,then
-Dmt = Co t Cit - (co t Ci (t-
D
)
= cwe can state
the
general
caseand
prove itin
the
following
theorem
If Mt = Co t att . . . t Ck tk is a 4thdegree
polynomial
,then
17"
Mt = K! CK .
Prout
.Proof
by
induction . Wejust
showed
that
the
ka caseis true lice, Eco
tGt
)
= Ci)
. Assume
the
theorem
holdsfor
polynomials
oforders
1 , 2, -. , K -I . Then T" ( co t Gt t . -tenth
)
= T "-' (cot at t . - t Ck-it " -' tcut
k)
=(
Ck -D ! Cx -i tti
-'cat"
)
by
the
inductionhypothesis
.--
wait
"III.
III.
a-us
=p " -'(
Cath
-CkEE
(ki
)
c-pit
"- i)
(by
Binomial
theorem
)
=D" -'(
Kc" t" ' - C C-dit
" -i)
= k Ck Ck- D! = Ck k! , since for j e k- I ,17k
- ' qti
= Tl " -'DJ
ajtj
=tyre
- i -Jajj
!
= 0 Thus,
differencing
K times willeliminate
any
polynomial
trend ofdegree
K or less .A
couple
ofobservations
ondifferencing
:①
Ok is a causal linearfiltering
operation
.Indeed
, 17"
Xt
= C l - B)
" Xt =(4)
HI"-it
-B)
i]
Xt =EC
C-Di
Bi
Xt
i --o =ii.
(7)
I-Di
Xt - i ,which
is a causal linear filter with filtercoefficients
a. =L , ai -(
ki
)
L-1)
i , i ←I , . .. , K , and aj = O fo-j
# {oil, .... , K}
. .⑦
Differencing
isnot
agood
way
to do trendestimation
.For
example
, if Xt = co t Cit tYt
. Thent
Xt = C , t Ye-Ye, and the linear trend is
eliminated
.However
,
the
trend estimate would be
hit
= Xt - T Xt = Xt - ( Xt - Xt - i)
=Xt
-i .Thus
,the
trend estimate is
just
theoriginal
time seriesshifted
back
intime
by
one unit ,which
does follow thegeneral
linear trendbut is not a
good
estimate of a line .So
,
differencing
is useful forremoving
lowfrequency
components
of the time series (such
as trends)
, but does so in acrude way
,
and
removes more thanjust
the trend . If oneonly
cares that what is left over looksstationary
,and
not inestimating
the
trend itself ,then
differencing
isfine
and isa
useful
building
block for morecomplicated
models
. It is a
simple operation
and
easily
reversed forprediction
purposes : one
does
prediction
onthe
differenced
data
using
aprediction
based
on a model for
stationary
time series,
and
then
reversethe
differencing
toget
the
prediction
inthe
original
time scaleof
the
data
.Reversing
the
differencing
!If