TAHAP PEMBENTUKAN MODEL ARIMA
KONSEP DASAR
ARIMA
STASIONER
Kondisi stasioner terdiri atas dua hal, yaitu:
1. Stasioner dalam rata-rata
Jika kondisi stasioner dalam rata-rata tidak terpenuhi
diperlukan proses pembedaan (differencing)
𝑧𝑡 = 1 − 𝐵 𝑑 𝑍𝑡
𝑑 : orde differencing
𝐵 : Backshift operator (operator mundur) yang didefinisikan dengan 𝐵 𝑑 𝑍𝑡 = 𝑍𝑡−𝑑
Proses differencing pada orde pertama merupakan
selisih antara data ke-t dengan data ke t-1, yaitu:
∆𝑍𝑡 = 𝑍𝑡 − 𝑍𝑡−1
Sedangkan bentuk differencing orde kedua adalah:
∆2 𝑍𝑡 = ∆𝑍𝑡 − ∆𝑍𝑡−1 = 𝑍𝑡 − 𝑍𝑡−1 − 𝑍𝑡−1 − 𝑍𝑡−2
= 𝑍𝑡 − 2𝑍𝑡−1 + 𝑍𝑡−2
STASIONER
2. Stasioner dalam varians
Jika kondisi stasioner dalam varians tidak terpenuhi,
Box & Cox memperkenalkan transformasi pangkat
(power transformation):
𝜆
𝜆 𝑍𝑡 −1
𝑍𝑡 =
𝜆
Dengan melihat λ=0 yang berkorespondensi dengan
transformasi logaritma, dinotasikan sbb:
𝜆
𝜆 𝑍𝑡 −1
lim 𝑇 𝑍𝑡 = lim 𝑍𝑡 = lim = ln 𝑍𝑡
𝜆→0 𝜆→0 𝜆→0 𝜆
STASIONER
2. Stasioner dalam varians
Nilai λ Transformasi
1
-1,0
𝑍𝑡
1
-0,5
𝑍𝑡
0,0 ln 𝑍𝑡
0,5 𝑍𝑡
1,0 𝑍𝑡
STASIONER
2. Stasioner dalam varians
Ketentuan untuk menstabilkan variansi:
a) Transformasi hanya boleh dilakukan untuk deret data
yang positif
b) Transformasi dilakukan sebelum melakukan differencing
dan pemodelan derete waktu
c) nilai λ dipilih berdasarkan SSE dari deret hasil
transformasi. Nilai SSE terkecil memberikan hasil
variansi paling konstan
𝑛
2
𝑆𝑆𝐸 𝜆 = 𝑍𝑡 𝜆 − 𝜇
𝑡=1
d) Transformasi tidak hanya menstabilkan varians, tetapi
juga dapat menormalkan distribusi
STASIONER
Data stasioner terbagi menjadi dua, yaitu:
1. Stasioner kuat (strickly stasioner)
Data dikatakan stasioner kuat jika distribusi
gabungan dari Zt1,Zt2,…,Ztn sama dengan
distribusi gabungan dari Zt1+t,Zt2+t,…,Ztn+k
untuk setiap t1,t2,…,tn dan k.
2. Stasioner lemah (weakly stasioner)
Data dikatakan stasioner lemah jika rata-
rata hitung data konstan, dan
autokovariansnya merupakan fungsi dari lag
STATIONARY AND NONSTATIONARY
TIME SERIES
Stationer
Nonstationer
THE FIRST DIFFERENCES: ZT = Y2T – Y2T-1
Nonstationer
Differences
Stationer
SAMPLE AUTOCORRELATION FUNCTION
(ACF)
Untuk data series Z1, Z2, …, Zn :
ACF UNTUK DATA STASIONER
dies down
1 (exponential)
1 cuts off
0 0
8 Lag k 8 Lag k
no
oscillation
-1 -1
dies down dies down
1 (exponential) 1 (sinusoidal)
0 0
8 Lag k 8 Lag k
-1 oscillation -1
DYING DOWN FAIRLY QUICKLY VERSUS EXTREMELY
SLOWLY
1 Dying down fairly quickly stationary time
series (usually)
0 Lag k
8
-1
Dying down extremely slowly nonstationary time
1 series (usually)
0 Lag k
8
-1
SAMPLE PARTIAL AUTOCORRELATION FUNCTION
(PACF)
Untuk data series Z1, Z2, …, Zn : Corr(Zt,Zt-k|Zt-1,…,Zt-k+1)
CALCULATION OF PACF AT LAG 1, 2 AND 3
The sample partial autocorelations at lag 1, 2 and 3
are:
MINITAB OUTPUT OF STATIONARY TIME SERIES
ACF PACF
Dying down fairly quickly Cuts off after lag 2
MINITAB OUTPUT OF NONSTATIONARY TIME SERIES
ACF PACF
Dying down extremely slowly Cuts off after lag 2
EXPLANATION OF ACF … [MINITAB OUTPUT]
+ +
t/2 . se(rk) t/2 . se(rk)
GENERAL THEORETICAL ACF AND PACF OF ARIMA
MODELS
Model ACF PACF
MA(q): moving average of order q Cuts off Dies down
after lag q
AR(p): autoregressive of order p Dies down Cuts off
after lag p
ARMA(p,q): mixed autoregressive- Dies down Dies down
moving average of order (p,q)
AR(p) or MA(q) Cuts off Cuts off
after lag q after lag p
No order AR or MA No spike No spike
(White Noise or Random process)
THEORETICALLY OF ACF AND PACF OF FIRST-
ORDER MOVING AVERAGE MODEL OR MA(1)
The model
Zt = + at – 1 at-1
Invertibility condition : –1 < 1 < 1
Theoretically of ACF Theoretically of PACF
THEORETICALLY OF ACF AND PACF OF FIRST-ORDER
MOVING AVERAGE MODEL OR MA(1) … [GRAPHICS
ILLUSTRATION]
ACF PACF
ACF PACF
SIMULATION EXAMPLE OF ACF AND PACF OF FIRST-
ORDER MOVING AVERAGE MODEL OR MA(1) … [GRAPHICS
ILLUSTRATION]
THEORETICALLY OF ACF AND PACF OF SECOND-
ORDER MOVING AVERAGE MODEL OR MA(2)
The model
Zt = + at – 1 at-1 – 2 at-2
Invertibility condition : 1 + 2 < 1 ; 2 1 < 1 ; |2| < 1
Theoretically of ACF Theoretically of PACF
Dies Down (according to a
mixture of damped exponentials
and/or damped sine waves)
THEORETICALLY OF ACF AND PACF OF SECOND-ORDER
MOVING AVERAGE MODEL OR MA(2) … [GRAPHICS ILLUSTRATION]
… (1)
ACF PACF
ACF PACF
THEORETICALLY OF ACF AND PACF OF SECOND-ORDER
MOVING AVERAGE MODEL OR MA(2) … [GRAPHICS ILLUSTRATION]
… (2)
ACF PACF
ACF PACF
SIMULATION EXAMPLE OF ACF AND PACF OF SECOND-ORDER
MOVING AVERAGE MODEL OR MA(2) … [GRAPHICS
ILLUSTRATION]
EXAMPLE: IDENTIFICATION STEP [STATIONARY, ACF AND
PACF]
ACF Stationer time series PACF
Dies down Cuts off after lag 2
[sinusoidal]
PENDUGAAN
PARAMETER
1. Metode Momen
2. Metode Least Square (Conditional
Least Square)
3. Metode Maximum Likelihood
1. METODE MOMEN
Taksiran parameter berdasarkan
hubungan pada:
𝑛
𝑡=1 𝑍𝑡
𝜇=𝑍=
𝑛
dan
𝜌𝑘 = 𝑟𝑘
1. METODE MOMEN
Persamaan Yule Walker
k 1, 2,...., p; 0 1, k k
1 1 2 1 ... p p 1
2 1 1 2 ... p p 2
. . . . . . . . . .
p 1 p 1 2 p 2 ... p
Karena nilai ρ1,ρ2,…,ρp tidak diketahui, koefisien
tersebut diganti dengan penaksirnya, yaitu
r1,r2,…,rp
1. METODE MOMEN
A. Parameter AR
Persamaan Yule Walker dapat dipecah untuk
𝜙1 , 𝜙2 , … , 𝜙𝑝 untuk memperoleh taksiran awal model
AR. Melalui prosedur tersebut, diperoleh nilai taksiran
parameter model AR(1)
𝜙1 = 𝑟1
Untuk model AR(2), persamaan Yule Walker akan
menghasilkan:
𝜌1 = 𝜙1 + 𝜙2 𝜌1
𝜌2 = 𝜙1 𝜌1 + 𝜙2
Sehingga diperoleh taksiran parameter model AR(2)
𝑟1 1 − 𝑟2
𝜙1 =
1 − 𝑟12
𝑟2 − 𝑟12
𝜙2 =
1 − 𝑟12
1. METODE MOMEN
B. Parameter MA
Untuk model MA(1) persamaan Yule Walker akan
menghasilkan:
−𝜃1
𝜌1 =
1 + 𝜃12
Sehingga diperoleh taksiran parameter model MA(1):
−1 ± 1 − 4𝑟12
𝜃1 =
2𝑟1
Untuk model MA(2) persamaan Yule Walker akan
menghasilkan:
−𝜃1 − 𝜃1 𝜃2
𝜌1 =
1 + 𝜃12 + 𝜃22
−𝜃2
𝜌2 =
1 + 𝜃12 + 𝜃22
2. METODE LEAST SQUARE
Model AR(1):
Zt Zt 1 at
Bentuk persamaan Least Square:
S , a = Zt Zt 1
T T 2
2
t
t 2 t 2
Diperoleh taksiran parameter:
Zt Z Zt 1 Z
T
ˆ t 2
Zt 1 Z
T 2
t 2
2. METODE LEAST SQUARE
Model MA(1):
Zt at at 1 at Zt at 1
a0 0
a1 z1
a2 z2 a1
a3 z3 a2
. . . . .
aT zT aT 1
Tentukan nilai yang me min imumkan persamaan
T
S* at2
t 1
3. METODE MAXIMUM LIKELIHOOD
Model AR(1)
Z t 1 Z t 1 a t
Z t 1 Z t 1 a t
Z t 1 1 1 Z t 1 a t
Z t 0 1 Z t 1 a t
Dimana: 0 1 1
a t ~ IIND 0, a2
Fungsi kepadatan peluang untuk data pertama:
Z1 0 / 1 1 2
f Z1 : 0 , 1 , a
1
2
Exp
2 a2 / 1 1
2 a / 1 1
2 2
3. METODE MAXIMUM LIKELIHOOD
Model AR(1)
Fungsi kepadatan peluang untuk data kedua:
Z1 0 1 Z1 2
1
f Z 2 : 0 , 1 , a
2
Exp
2 2
a 2 a2
Fungsi likelihood untuk model AR(1) dapat diperoleh
dengan mengalikan seluruh fungsi kepadatan peluang
sbb:
LZ1 , Z 2 ,..., Z T : 0 , 1 , a2 f Z1 : 0 , 1 , a2 .f Z 2 : 0 , 1 , a2 ....f Z T : 0 , 1 , a2
LZ1 , Z 2 ,..., Z T : 0 , 1 , f Z1 : 0 , 1 , . f Z t : 0 , 1 , a2
T
2 2
a a
t 2
Misalkan :
3. METODE MAXIMUM LIKELIHOOD
Model AR(1)
Misalkan:
2
1 Z / 1
f Z1 : 0 , 1 , a2 Exp 1 2 0 1 A dan
2 a2 / 1 1
2 a / 1
2
1
~
1
2
Z t 0 Z
1 t 1 2
Exp t 2 B maka
2 2 2 2
a a
L Z1 , Z 2 ,.., Z T : 0 , 1 , a2 A.B
3. METODE MAXIMUM LIKELIHOOD
Model AR(1)
Persamaan tsb dapat dilakukan dengan
mentransformasikan ke dalam bentuk persamaan
logaritma menjadi:
ln f ln f Z : , ,
T
L 0 , 1 , a2 Z1 : 0 , 1 , a2 t 0 1
2
a
t 2
Z1 0 / 1 1
2
T 1 T 1
1
2
1
2
2
ln 2 ln a / 1 1
2
2 a2 / 1 12
2
ln 2
2
ln a2
1 T
2 t
Z 0 Z
1 t 1 2
2 a t 2
T 1 ln 2 T 1 ln 2
T
ln f Z t : 0 , 1 , 2 1
t 0 1 t 1
Z Z 2
2 a2
a a
2 2 t 2
PEMERIKSAAN
DIAGNOSTIK
1. Uji kesignifikanan parameter
2. Uji kesesuaian model
2. UJI KESIGNIFIKANAN PARAMETER
t-values and prob-values for testing parameter model
ARIMA
Parameters
ARIMA
model estimates
2. UJI KESESUAIAN MODEL
a) Uji sisa white noise
ACF of
residual
b) Uji kenormalan sisaan
PERAMALAN
THE GENERAL [NONSEASONAL] ARIMA(P,D,Q) MODELS
The Model is
where
is an appropriate pre-
differencing transformation
Do not need
pre-differencing
Yt = Original transformation
Data
FORECASTING OF ARIMA(P,D,Q) MODEL
Forecasting of AR(1) model
or
Forecasting of MA(1) model
EXAMPLE: ARIMA(1,0,1) MODEL
Yt = Original
Data
The Model is
Zt = Yt
where
and
Therefore,
EXAMPLE: ARIMA(1,0,1) MODEL … [OTHER CALCULATION]
Yt = Original
The Model is Data
Zt = Yt
p=1 d=0 q=1
Therefore,
EXAMPLE: ARIMA(1,1,1) MODEL … [NONSTATIONARY MODEL]
Yt = Original
Data
The Model is
Zt = Yt – Yt-
1
where
and
Therefore,
Mean (Zt)
PEMILIHAN
MODEL TERBAIK
KRITERIA PEMILIHAN MODEL
Berdasarkan pada residual
AIC M n ln ̂ a2 2M
Akaike’s AIC
SBC( M ) n ln ˆ a2 M ln n
Schwart’z SBC
Criterion
1
1 , p 0
n
CAT p
Parzen’s CAT p
1 1 1 , p 1, 2,3,...
Criterion n j 1 ˆ j
2
ˆ 2p
CONTOH PENERAPAN
EXAMPLE 1: DAILY READINGS OF VISCOSITY OF CHEMICAL
PRODUCT XB-77-5 [BOWERMAN AND O’CONNELL, PG. 471]
EXAMPLE 1: IDENTIFICATION STEP [STATIONARY, ACF AND
PACF]
ACF PACF
Stationer time series
Dies down Cuts off after lag 2
[sinusoidal]
EXAMPLE 1: ESTIMATION AND DIAGNOSTIC
CHECK STEP
Estimation
and Testing
parameter
Diagnostic
Check (white
noise residual)
EXAMPLE 1: DIAGNOSTIC CHECK STEP … [NORMALITY TEST OF
RESIDUALS]
EXAMPLE 1: FORECASTING STEP [MINITAB
OUTPUT]
CALCULATION: FORECASTING (FITS AND FORECAST)
[CONTINUED]
EXAMPLE 2: WEEKLY SALES OF ULTRA SHINE TOOTHPASTE (IN
UNITS OF 1000 TUBES) [BOWERMAN AND O’CONNELL,
PG. 478]
t Yt t Yt t Yt
1. 235.000 31. 551.925 61. 846.962
2. 239.000 32. 557.929 62. 853.830
3. 244.090 33. 564.285 63. 860.840
4. 252.731 34. 572.164 64. 871.075
5. 264.377 35. 582.926 65. 877.792
… … … … … …
… … … … ... …
26. 517.237 56. 805.844 86. 996.291
27. 524.349 57. 815.122 87. 1003.100
28. 532.104 58. 822.905 88. 1010.320
29. 538.097 59. 930.663 89. 1018.420
30. 544.948 60. 839.600 90. 1029.480
EXAMPLE 2: IDENTIFICATION STEP [STATIONARITY AND
ACF]
Dying down
extremely slowly
ACF
Nonstationary
time series
EXAMPLE 2: IDENTIFICATION STEP … DIFFERENCE [WT = YT – YT-
1]
Stationary
Wt AR(1)
time series
or
Yt
ARI(1,1)
ACF PACF
Dies down Cuts off after lag 1
EXAMPLE 2: ESTIMATION AND DIAGNOSTIC
CHECK STEP
Yt = 3.0232 + 0.6591 Yt-1 + at
Estimation
and Testing
parameter
Diagnostic
Check (white
noise residual)
EXAMPLE 2: DIAGNOSTIC CHECK STEP … [NORMALITY TEST OF
RESIDUALS]
EXAMPLE 2: FORECASTING STEP [MINITAB
OUTPUT]
COMPARISON: ARIMA VERSUS TREND ANALYSIS
ARIMA(1,1,0)
MSE = 7.647
PLOT COMPARISON: ARIMA VERSUS TREND ANALYSIS
ARIMA(1,1,0)
MSE = 7.647
Trend Analysis
MSE = 598.212
Forecast
comparison
PLOT RESIDUAL COMPARISON: ARIMA VERSUS TREND
ANALYSIS
ARIMA(1,1,0)
MSE = 7.647
Trend Analysis
MSE = 598.212