Tehnic de ajustare care se poate aplica unei serii de timp cu o pronunat component aleatoare.
Este utilizat n principal la realizarea de predicii, atunci cnd alte metode (ARIMA, Trend
determinist etc.) nu dau rezultate.
Context
Fie seria de timp A1 , A2 , ,An..
Modelele de predicie au drept scop estimarea urmtoarelor componente:
1. Media
2. Trendul
3. Componenta sezonier
4. Componenta ciclic
Procesul de predicie
1. Se estimeaz parametrii modelului folosind date istorice.
2. Se testeaz modelul prin back-testing.
3. Se utilizeaz modelul pentru realizarea de predicii n viitor.
1 n
At Ft
n t 1
1 n
( At Ft ) 2
n t 1
1 n
( At Ft ) 2
n t 1
(A
Ft )
t 1
Ft 1
At At 1 At k 1
k
Exponential Smoothing
Ft 1 Ft ( At Ft ) At (1 ) Ft
-
0<<1 - constanta de netezire; valori marisenzitivitate mai mare, valori micistabilitate mai
mare
Yt 1 At (1 ) At 1 (1 ) 2 At 2 (1 ) 3 At 3
-
Exemplu
Month
Apr-07
May-07
Jun-07
Jul-07
Aug-07
Sep-07
Oct-07
Nov-07
Dec-07
Actual
115
111
120
99
132
120
141
116
141
SMA (6
months)
104.8
108.0
111.7
111.5
109.5
114.8
116.2
120.5
121.3
ES( = 0.2)
104.8
106.8
107.7
110.1
107.9
112.7
114.2
119.5
118.8
160
140
120
100
Actual
80
SMA (6 months)
60
ES( = 0.2)
40
20
0
1
MAD
MSE
SFE
TS
12.30
210.52
76.67
6.23
Exponential Smoothing
13.53
250.43
92.37
6.83
Modele nestaionare
Ft n E t nTt
3
unde
E t At (1 ) ( E t 1 Tt 1 )
Tt ( E t E t 1 ) (1 )Tt 1
-
Ft n E t nTt S t n p
Unde:
E t ( At S t p ) (1 ) ( E t 1 Tt 1 )
Tt ( E t E t 1 ) (1 )Tt 1
S t ( At E t ) (1 ) S t p
-parametrul (0 < < 1) este folosit pentru estimarea componentei sezoniere.
Holt-Winters Multiplicative Method
Ft n ( E t nTt ) St n p
Et
At
(1 ) ( Et 1 Tt 1 )
St p
Tt ( E t E t 1 ) (1 )Tt 1
St
At
(1 ) S t p
Et
Filtrul Hoddrick-Prescott
-
metod utilizat pentru estimarea trendului n serii care prezint component ciclic
yt = t + ct
T 1
t min{ ( y t t ) 2 [( t 1 t ) ( t t 1 )]2 }
1
PIB
40,000
36,000
32,000
28,000
24,000
20,000
16,000
00
01
02
03
04
05
06
07
08
09
10
11
12
13
14
Sezonalitate:
- Stochastic SARIMA
- Determinist medii sezoniere + trend sau sinusoide
Dac f(.) este o serie determinist, atunci f(.) este periodic cu perioada s dac
f t f t k s , k 0,1,2, .
Dac Yt este un process stochastic, acesta este sezonier (periodic) de perioad s dac Yt i Yt+ks au
aceeai distribuie.
Sezonalitate + trend: Yt t
Dit at .
i 1
Yn h n h i Di ,n h a n h .
i 1
Predicia: Yn h n h
D
i
i ,n h
i 1
Exemplu
Numrul de pasageri transportai pe aeroporturile din Romnia
PASAGERI
3,500,000
3,000,000
2,500,000
2,000,000
1,500,000
1,000,000
500,000
1
1
4q
5q
0
0
20
20
1
6q
0
20
q
07
0
2
1
1
8q
9q
0
0
20
20
1
1
1q
0q
1
1
20
20
q
12
0
2
3q
01
1
4q
1
20
LNPAS
15.2
14.8
14.4
14.0
13.6
13.2
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
Coefficient
Std. Error
t-Statistic
Prob.
C
@TREND
13.74419
0.030092
0.066799
0.002739
205.7531
10.98836
0.0000
0.0000
R-squared
Adjusted R-squared
S.E. of regression
Sum squared resid
Log likelihood
F-statistic
Prob(F-statistic)
0.746513
0.740331
0.222848
2.036113
4.564032
120.7441
0.000000
14.37612
0.437319
-0.119257
-0.037341
-0.089049
1.092921
Corelograma reziduurilor
i 1
Coefficient
Std. Error
t-Statistic
Prob.
@TREND
D1
D2
D3
D4
48471.73
517092.2
963172.0
1345242.
712056.6
2725.701
86194.22
87943.46
89741.39
90442.35
17.78322
5.999151
10.95217
14.99021
7.873044
0.0000
0.0000
0.0000
0.0000
0.0000
R-squared
Adjusted R-squared
S.E. of regression
Sum squared resid
Log likelihood
Durbin-Watson stat
0.917148
0.908427
221437.0
1.86E+12
-587.5961
0.758846
1906305.
731756.8
27.56261
27.76740
27.63813
Coefficient
Std. Error
t-Statistic
Prob.
@TREND
D1
D2
D3
D4
0.029384
13.55297
13.80789
13.97890
13.69018
0.001964
0.062092
0.063352
0.064647
0.065152
14.96505
218.2721
217.9545
216.2330
210.1258
0.0000
0.0000
0.0000
0.0000
0.0000
R-squared
Adjusted R-squared
S.E. of regression
Sum squared resid
Log likelihood
Durbin-Watson stat
0.879620
0.866949
0.159517
0.966941
20.57423
0.118787
10
14.37612
0.437319
-0.724383
-0.519592
-0.648862
15.5
15.0
14.5
.3
14.0
.2
13.5
.1
13.0
.0
-.1
-.2
-.3
04
05
06
07
08
Residual
09
10
Actual
11
Fitted
Corelograma reziduurilor
11
12
13
14
Predicia
6,000,000
5,000,000
4,000,000
3,000,000
2,000,000
1,000,000
0
04
05
06
07
08
09
10
L
PASAGERI
11
12
13
14
U
PASAGERIF
7,000,000
Forecast: PASAGERIF
Actual: PASAGERI
Forecast sample: 2004Q1 2014Q4
Included observations: 43
Root Mean Squared Error 305878.4
Mean Absolute Error
254579.7
Mean Abs. Percent Error
13.74182
Theil Inequality Coefficient 0.073921
Bias Proportion
0.006671
Variance Proportion
0.104974
Covariance Proportion
0.888355
6,000,000
5,000,000
4,000,000
3,000,000
2,000,000
1,000,000
0
04
05
06
07
08
09
PASAGERIF
10
11
2 S.E.
12
12
13
14
B 1 B
P B s 1 1 B s 2 B 2 s P B sP
Q
2 B 2 s Q B sQ .
Exemplu: SARIMA(0,0,1)12=SMA(1)12
Yt 0 at at 12 .
13
E(Yt) = 0.
Var Yt 1 2 a2
, k 12
ACF : k 1 2
.
0, altfel
Exemplu: SARIMA(1,0,0)12
1 B Y
12
0 at - model AR sezonal.
Yt Yt 12 0 at .
Condiia de staionaritate:||<1.
E Yt
Var Yt
ACF : 12 k k , k 0,1, 2, .
0
.
1
a2
.
1 2
SARIMA(p, d, q)(P,D,Q)s : p B P B s 1 B 1 B s Yt 0 q B Q B s at .
d
Condiie: rdcinile polinoamelor (B); (Bs); (B) i (Bs) snt n afara cercului unitate!
Fie Wt = (1 B)(1 B12)Yt, unde = (1 B) este diferena standard, iar 12= (1 B12)
este diferena sezonier.
Wt 1 B 1 B12 at
Wt at at 1 at 12 at 13 .
Wt ~ I 0
14
1 2 1 2 a2 , k 0
2
2
1 a , k 1
k 1 2 a2 , k 12 .
2
a , k 11,13
0, o.w.
1 2 , k 1
, k 12
.
k 1 2
, k 11,13
2
2
1 1
0, o.w.
SARIMA(1, 1, 1)(1,1,1)12.
15
PASAGERI
1,200,000
1,000,000
800,000
600,000
400,000
200,000
5
00
1
2
1
6m
0
0
7
00
1
m
2
8
00
1
m
1
9m
0
20
0
01
1
m
m
11
0
2
1
2
2
01
1
m
2
3
01
1
2
4
01
.4
.2
.0
-.2
-.4
-.6
2005
2006
2007
2008
2009
2010
16
2011
2012
2013
2014
e. Corelograma seriei Xt
17
Coefficient
Std. Error
t-Statistic
Prob.
C
MA(1)
SMA(12)
-0.002467
-0.741865
-0.932179
0.000847
0.067615
0.030537
-2.913685
-10.97193
-30.52573
0.0044
0.0000
0.0000
R-squared
Adjusted R-squared
S.E. of regression
Sum squared resid
Log likelihood
F-statistic
Prob(F-statistic)
Inverted MA Roots
0.643076
0.636008
0.095244
0.916222
98.48852
90.98681
0.000000
.99
.50-.86i
-.50+.86i
-.99
.86-.50i
.50+.86i
-.50-.86i
.86+.50i
.00+.99i
-.86+.50i
18
-0.000836
0.157868
-1.836318
-1.760037
-1.805414
1.702707
.74
-.00-.99i
-.86-.50i
Dependent Variable: X
Method: Least Squares
Date: 03/09/15 Time: 21:50
Sample (adjusted): 2007M03 2014M09
Included observations: 91 after adjustments
Convergence achieved after 31 iterations
MA Backcast: 2006M02 2007M02
Variable
Coefficient
Std. Error
t-Statistic
Prob.
C
AR(1)
SAR(12)
MA(1)
SMA(12)
-0.003168
0.378653
-0.330655
-0.850327
0.974889
0.002761
0.135663
0.056928
0.079750
0.015391
-1.147497
2.791137
-5.808277
-10.66237
63.34215
0.2544
0.0065
0.0000
0.0000
0.0000
R-squared
Adjusted R-squared
S.E. of regression
Sum squared resid
Log likelihood
F-statistic
Prob(F-statistic)
Inverted AR Roots
Inverted MA Roots
0.742181
0.730190
0.077057
0.510648
106.7001
61.89199
0.000000
.88+.24i
.38
-.24-.88i
-.88-.24i
.96+.26i
.71+.71i
-.26-.96i
-.96+.26i
-0.005014
0.148348
-2.235168
-2.097208
-2.179510
1.975411
.88-.24i
.24-.88i
-.64+.64i
.64-.64i
.24+.88i
-.64-.64i
.64-.64i
-.24+.88i
-.88+.24i
.96-.26i
.26-.96i
-.71-.71i
.85
.26+.96i
-.71-.71i
.71-.71i
-.26+.96i
-.96-.26i
Model final:
12
19
Predicia
.6
.4
.2
.0
-.2
-.4
.2
-.6
.1
.0
-.1
-.2
-.3
2007
2008
2009
2010
Residual
2011
Actual
2012
2013
2014
Fitted
1,400,000
Forecast: PASAGERIF
Actual: PASAGERI
Forecast sample: 2005M01 2014M09
Adjusted sample: 2007M03 2014M09
Included observations: 91
Root Mean Squared Error 50852.49
Mean Absolute Error
37763.02
Mean Abs. Percent Error
5.429306
Theil Inequality Coefficient 0.032662
Bias Proportion
0.000374
Variance Proportion
0.003338
Covariance Proportion
0.996288
1,200,000
1,000,000
800,000
600,000
400,000
200,000
2007
2008
2009
2010
PASAGERIF
2011
2012
2 S.E.
20
2013
2014
1,400,000
1,200,000
1,000,000
800,000
600,000
400,000
I
II
III
IV
II
2012
III
IV
2013
UPPER
PASAGERI
21
II
2014
PASAGERIF
LOW ER
III