time series analysis - lecture 2 a general forecasting principle set up probability models for which...
Post on 19-Dec-2015
213 views
TRANSCRIPT
Time series analysis - lecture 2
A general forecasting principle
Set up probability models for which we can derive analytical expressions for
and estimate the unknown parameters in that expression
Use estimated autocorrelations to select a suitable model class
,...),,|(ˆ321 ttttt YYYYEY
Time series analysis - lecture 2
The first order autoregressive model: AR(1)
...,2,1,),(
1
)()(
1)(
11 where,
2
1
kYYCorr
VarYVar
cYE
YcY
ktkt
tt
t
ttt
The autocorrelation tails off exponentially
Time series analysis - lecture 2
Datasets simulated by using a random number
generator and first order autoregressive models
-3
-2
-1
0
1
2
3
1 13 25 37 49 61 73 85 97 109
AR(1)
-8-6-4-20246
1 13 25 37 49 61 73 85 97 109
AR(1)
= 0.2 = 0.9
Time series analysis - lecture 2
The first order autoregressive model
- prediction
...,2,1,ˆ
11 where,1
kYY
YY
tk
kt
ttt
Time series analysis - lecture 2
Weekly SEK/EUR exchange rate Jan 2004 - Oct 2007
8.9
9
9.1
9.2
9.3
9.4
9.5
9.6
9.7
02/01/2004 01/01/2005 01/01/2006 01/01/2007 01/01/2008
SE
K/E
UR
exc
han
ge
rate
2018161412108642
1.00.80.60.40.20.0
-0.2-0.4-0.6-0.8-1.0
Lag
Auto
corr
ela
tion
Time series analysis - lecture 2
Daily SEK/EUR exchange rate Jan 2004 - Oct 2007
757065605550454035302520151051
1.00.80.60.40.20.0
-0.2-0.4-0.6-0.8-1.0
LagAuto
corr
ela
tion
8.8
8.99
9.19.2
9.39.4
9.59.6
9.7
02/01/2004 01/01/2005 01/01/2006 01/01/2007 01/01/2008
SE
K/E
UR
ex
ch
an
ge
ra
te
Time series analysis - lecture 2
No. registered cars
0
50000
100000
150000
200000
250000
300000
350000
400000
1970 1980 1990 2000 2010
No
. reg
iste
red
car
s
10987654321
1.00.80.60.40.20.0
-0.2-0.4-0.6-0.8-1.0
Lag
Auto
corr
ela
tion
Time series analysis - lecture 2
Weekly SEK/EUR exchange rate Jan 2004 - Oct 2007
AR(1) model
200180160140120100806040201
9.6
9.5
9.4
9.3
9.2
9.1
9.0
8.9
Time
SEK
/EU
R
Time Series Plot for SEK/ EUR(with forecasts and their 95% confidence limits)
Final Estimates of Parameters
Type Coef SE Coef T P
AR 1 0.9537 0.0218 43.84 0.000
Constant 0.426548 0.002833 150.56 0.000
Mean 9.21262 0.06119
Time series analysis - lecture 2
Daily SEK/EUR exchange rate Jan 2004 - Oct 2007
AR(1) model
Final Estimates of Parameters
Type Coef SE Coef T P
AR 1 0.9835 0.0061 161.78 0.000
Constant 0.151965 0.000800 189.86 0.000
Mean 9.21316 0.04853
9008007006005004003002001001
9.7
9.6
9.5
9.4
9.3
9.2
9.1
9.0
8.9
Time
SEK
/EU
RO
WARNING *
Back forecasts
not dying out
rapidly
Time series analysis - lecture 2
The general auto-regressive model: AR(p)
{Yt} is said to form an AR(p) sequence if
where the error terms t are independent
and N(0;)
,...11 tptptt YYY
Time series analysis - lecture 2
The backshift operator B
The backshift operator B is defined by the relation
It follows that
and we can write the AR(p) model:
1 tt YBY
)...1(... 111 ttp
pptptt YBBYYY
...,2,1, kYYB kttk
Time series analysis - lecture 2
The characteristic equation of an
AR(p) sequence
The equation
is called the characteristic equation of an AR(p) sequence
If the roots of this equation are all outside the unit circle, then the AR(p) sequence is stationary
It returns repeatedly to zero, and the autocorrelations tail off with increasing time lags
0...1 1 pp xx
Time series analysis - lecture 2
Stationarity of an
AR(p) sequence
p=1Characteristic equation:
The root is outside the unit circle if
p=2Characteristic equation:
The roots of this equation are
all outside the unit circle if:
01 1 x
01 221 xx
11 1
11
1
1
2
12
21
Time series analysis - lecture 2
The moving average model: MA(q)
{Yt} is said to form an MA(q) sequence if
where the error terms t are independent
and N(0;)
,...11 qtqtttY
Time series analysis - lecture 2
The general auto-regressive-moving-average
model ARMA(p,q)
{Yt} is said to form an ARMA(p,q) sequence if
where the error terms t are independent
and N(0;)
,...... 1111 qtqttptptt YYY
tq
qtp
p BBYBB )...1()...1( 11
Time series analysis - lecture 2
Typical auto-correlation functions of
ARMA(p,q) sequences
AR(p): Auto-correlations tail off gradually with increasing time-lags
MA(q): Auto-correlations are zero for time lags greater than q
ARMA(p,q): Auto-correlations tail off gradually with time-lags greater than q
Time series analysis - lecture 2
Partial auto-correlation
The partial auto-correlation between Yt and Yt+m represents the remaining correlation after Yt+1, …, Yt+m-1 have been fixed
t t+m
Time series analysis - lecture 2
Typical partial auto-correlation functions of
ARMA(p,q) sequences
AR(p): Partial auto-correlations are zero for time lags greater than p
MA(q): Partial auto-correlations tail off gradually with increasing time-lags
ARMA(p,q): Partial auto-correlations tail off gradually with time-lags greater than p
Time series analysis - lecture 2
Consumer price index and its first order differences
0
50
100
150
200
250
300
1980 1985 1990 1995 2000 2005 2010Co
nsu
mer
pri
ce in
dex
(19
80=
100) Consumer_price_index
-4
-2
0
2
4
6
8
1980 1985 1990 1995 2000 2005 2010F
irst
ord
er d
iffe
ren
ces
of
con
sum
er p
rice
ind
ex (
1980
=10
0)
Diff_Consumer_price
Time series analysis - lecture 2
Consumer price index - first order differences
10987654321
1.00.80.60.40.20.0
-0.2-0.4-0.6-0.8-1.0
LagAuto
corr
ela
tion
-4
-2
0
2
4
6
8
1980 1985 1990 1995 2000 2005 2010
Fir
st o
rder
dif
fere
nce
s o
f co
nsu
mer
pri
ce in
dex
(19
80=
100)
Diff_Consumer_price
Time series analysis - lecture 2
Consumer price index – predictions using an AR(1)
model fitted to first order differences of the original data
300250200150100501
300
250
200
150
100
Time
Consu
mer_
pri
ce_in
dex
Final Estimates of Parameters
Type Coef SE Coef T P
AR 1 0.1701 0.0561 3.03 0.003
Constant 0.49788 0.06307 7.89 0.000
Differencing: 1 regular difference
Number of observations: Original series 312, after differencing 311
Residuals: SS = 382.264 MS = 1.237
Time series analysis - lecture 2
The first order integrated autoregressive model: ARI(1)
11 where,1
1
ttt
ttt
wcw
YYw