BAYES ESTIMATION FOR ARMA MODEL FORECASTING UNDER NORMAL-GAMMAR PRIOR.

Bayes Estimation for ARMA Model Forecasting

under Normal-Gamma Prior
Zul Amryl'* and Adam Baharum2
1

2

Department of Mathematics, State University of Medan, lndonesia
School of Mathematical Sciences, Universiti Sains Malaysia, Pulau Pinang, Malaysia
*Corresponding
Author: zu l.a ryrrv@gmpil,cop

Presented

On The 3'd International Seminar on Operation Research.

To Celebrate The 50th Department of Mathematics University of
Sumatera Utara, Medan, lndonesia.

August 21-23,20L5.


Bayes Estimation for

ARMA Moder Forecasting under
Normal-Gamma Prior
Zul Amryl,- and Adam Baharum2

2schoo,",T:ffi:ll:1iT:*H',il1,::?:i,"Ji;Jil:,$H:fr1;i,Ti#x'ilMa'aysia
-Corresponding

Abstract

Author: zJl.am{r@gmail.com

This paper presents a Bayesian approach to find the Bayes estimator for the point forecast
and forecast variance

of

ARMA model under normal-gamma prior assumption with quadratic loss function in the form mathematical

expression. The
conditional posterior predictive density be obtained based the posterior under normal-gamma
prior and the conditional

predictive density' whereas the marginal conditional posterior predictive density
be obtain-ed based the conditional posterior

predictive density. Fufihermore, neither the point forecast nor the forecast variance
are both derived from the marginal
conditional posterior predictive density.

Keywords ARMA

1.

Model, Bayes Theorem, Normal-Gamma prior

Introduction

Bayes theorem calculates the posterior distribution as proportional to the product

of a prior distribution and the likelihood
function' The prior distribution is a probability model desiribing the knowledge about ihe parameters
before observing the
currently by the available data [1 ]. Main idea of Baye_sian forecaiting is the prrii.tiu"
distribution of the future given thJfast
data follows directly from the joint probabilistic model. The predictlve distiibution
is derived from the sampling predictive
density, weighted by the posterior dishibution
[4]. This pup"i ir done refer to Liu [15] discussed the Bayesian u"nutyiri, ro,
one-step ahead forecast in ARMA model and Liu
[16] also discussed the multiperiod ftrecast for ARX model employed the
Bayesian approach. others the paper related to this research are Amry and Baharum
[2],Fan& yao [7], Kleibergen & Hoek
[13]' and Uturbey [20] also discussed the Bayesian analysis for ARMA model. This'paper focuses to find the mathematical
expression of the Bayes estimator for the point forecast and forecast variance
of ARMI model under normal-gamma prior
assumption with quadratic loss function

2. Materials and Methods
The materials in this paper are some theories in mathematics and statistics such

as the ARMA model, Bayes theorem,
repeated integration, gamma distribution, and the univariate student's t-distribution.
The method is study of literature by
applying the Bayesian analysis under normal-gamma prior assumption.
2.1. ARMA model
The ARMA (1t, q)model [15] defined by:

y, =f.o y,-,
i=t

where {e'}
parameters.

is

sequence

of i i d

* j=tfe,r, , * r,


(2.1)

'

normal random variables with e,-N(0,rt1,

r>0

and unknown,

0,

and

Q

are

2.2.Bayes theorem

Suppose there are two discrete random variables
atd. )r, then the joint probability function can be written
p( x,y) = p( xl y) p,( y) and the marginal probability density function
of
[18] is:
p*(x)= z, p(x,y) 2,p(x|y) p,( y)
(2.2)
Bayes rule for the conditional p( y I x
is

X

:

)

X

p(


IfIis

ylx)= p(x,y) :
P*(r)

p,(y) _ p(xly) p,(y)
p,(x)
Z,p(xly)pr(y)

p(xly)

continuous, the Bayes theorem can be stated as:

,r

r'

(2.3)

r1=- PQ-2)-P'1! tj p(xly) p,r y )dy


(2.4)

by using the proportional notation (oc which can be expressed by:
)

p(ylx) 6 p(x)y)p,(y)

where p( y I x) is posterior distribution, p(x

I

I

(2.s)

is likelihood function and p,( y )is prior distribution.

2.3. Gamma distribution


A positive random quantity

/

is said to have a gamma distribution with paramet er n

>0

density function [17]:

e@=#ia*'
the notation isQ

- c(n,d)

the mean is

o(a)=]

and


d>0 if it has the probability

expf-4al

(2.6)

is var(Q)=ft

and the variance

2.4. Univariate student-t distribution
A random quantity, x, is said to have a student-t distribution on n degrees
of freedom with mode
r > 0if it has the probability density function [17]:

r( n*1),1

p1x


-

y) l)-1,*6- r
r

.,

r')'

rl;)@:

Thenotationisx-t,(p,r),tnemeanis

r(x)=p

-

p and scale parameter

,,,

|'

(2.7)

l

andthevariance

isvar{x)=J7-,ifn>2

3. Results
3.1. Likelihood Function for ARMA model
The k-step-ahead point forecast ofl ,* t , defined by

:

i(k)=E(Y,*ls]:)

(3.1)

where Sj =(y,, y,,..., y,*o_,)
Based the equation (2.

l)

be obtained residuals as:
e, = y,

-tt,y,-,

-f,o,r,_,

(3.2)

€,: e where r: min(0, p+l-q),
byBox&Jenkins[4],thelikelihoodfunctionfory:(Q1,Q2 ,000t,02,..,0)andr
based sj
By conditioning the first p observations and lerling eo:eo-t:...:

/ r s])x
-\

L\v,

I .1,-r ,f p
,
t,ll
r , p^rl-iLLlr,-Lo,t
-2t,,

one may approximate

t,

,n'k-t'

(3.3)

))j

The equation (3.3) can be expressed as:
/

\

r(p.rlsjJ
where B,

: (y,, !tt,

(n.k-ttp

(
cr , *fl-tllti
z

_1._o_,

I

..., lt+t_p, e6 er_t, ..., €*t_q)

I t=p+r

-

2y'

nrk-t

n+k-

|

ll

y @, a,_,)' ll
L /,8,_,+ t=p.r
t=p. r
J)

(3.4)

By

letting

Yp

!p+t

!n*-2

lp-t

Yp

!n+k,3

/ t

/z

/n+k-t-p

€o

€p*t

€n+k-2

p-t

€,

€n+k-j

U=


.'li';'i:l

W:U(.f,and V:Uxo

en+k-t-p

p*t-q € p*z-q
p _c_r


where 6,=y,-2d,y,,-ZF,aFj,t:p+l,p+2,...,n, 6,
a,, a,-,,..., a,-oo.

& d, aremaximumlikelihoodestimator of Q, & 0,,and

lutua.a iru'
A,

=y,

-Fr

B,_,

(3.5)

the likelihood function in equation (3.4) can be expressed as:

t\v.r

(n-k-tt-p

{

ll

|

lj

-f*r-r
rrpl-=llyl
2I,A;-zy'v+v'wvll

s))ne,

3.2. Posterior distribution under normal-gamma

(3.6)

prior

According to Broemeling and Shaarawy's suggestion [5], the normal-gamma prior of parameters Pand ris:

€(Y,r)=6,(Y )r).

n

(,(r)

r'f-'*o{-1lr'g*-v'ep-p'ev+pre1.t+2Bj}
'l 2'

,.tl

)

where f, - ,Ut,Gd-')

,

I , - c.tu (a, p) , Q is apositive definite matrix of the order (p + q),
By applying Bayes theorem to equation (3.6) and (3.7), the posterior
!o and r-, is:

of

n+k-t-p'2d+p_t

q.

p:w+e

and. K

=')t*,,ri

+ 1tr e1t +

2

p

andB are parameters.

1= r(v,r1s)), a(v,,)

n(Y,r-t1si

where

a

(

,t n
*pI-jlv,ev-v,
(v+ep)-(y+q1,fv+x)j

r,

(3.8)

.

3.3. Conditional posterior predictive density
Based

on

", = ,, -f,O,r,-,

-fr,",-,

with

e,

- u(o,r-,

)

be obtained f (et1s:,y,r-t )= (zo,

,)r *o{_LrG),}

Ifexpressed in y, is:

LI

f(y,lS),V,r-' 1=(2r, ' ''-t
)' uP\-jl,

P

-LQ,Y,-, -

Based the equation (3.9), be obtained the conditional predictive density

f

(

y.,o

I

s),y,

t -' ) =

(z

o,-' I i

*r{-

P,tr,,f' I

of y , *

(3.e)

p..

}lr. " - fit, r.. 0,, - f ,, n,-,)' }

p
o
l'l
*,'L*, f-f;Lt,.r _l,., t,-o-, _\t,,,*_,

)]

I_

* ': *r{-+l ,.-r -lf.r*.---, .ir,"..--,ll'l
"L
[

Bychanging f4,y,,o-,+f,o,r,*-, to:

\,=i

i-t

)))

(3.1 0)

!n+k-l

!

Qt!*t-t *02!n,*-z t .,.1-dp!,*-p

where

B*k-t=0*u,,

+ete,+k-t t0z€n+*-z +,..+eqe,+k_q

!n+k-2,..., !o+r-p,e,11-1,
-f

en+k-2,

6,.0 | s;, v,

-,
"

: (Ot Q, ,

0,

e

j e2

n+k-2

!n+*-p

0r)

€ n+k-t


n+k-z



n*-q

: V, B,*O-'

..., e,*o_o), theequation (3.10) can be written as:

*,i

*o

*

*p7-ilf,-r

)

,1

{-

!r[ r,. r -

r,

u,

* _,]']

- zv, a,-* t rn+t

* @, u*r-,)'lj

f ,r *v, Ry 2v7 t,,r_,
* ,iI *pl-ily',-o
,,-r lI

(3.1l)

where R = Bn+r_t@ BI*o_, and (v, n,*o_,)t =v, nv
Basedontheequation(3.8)and(3'll),beobtainedtheconditionalposteriorpredictivedensity

f o(t,*ls:,,Y,r-t

)

n o(v,

'"

ofy,t

lsj ) f (y,.ols),v)1 I

B,u-, y*o)- f)
,-o-,-o-r".0-, I lv'av -Yr 1lt
T'
*ofil(y +ep)r BI**., !,.* 1v * yl.r * rcll
+Q1t+

0.,

t"Lll

whereG:P+R

+

(3.12)

3.4. Marginal conditional posterior predictive density

The marginal conditional posterior predictive density

predictive density in equation (3.12):
_f

,(t,.rls,)
=

I[
_:-

=

-f ,f

of Y ,* r be obtained by integrating

the conditional posterior

t,,rls:,y,r1) dydr

ii''-*'-{-

;l:;,o.*r;'.' :;r':)l ;; i;-,:r' : -)}

n

*

o,

z--' ^'_l(r_o-,(1tt+g1t1+n,+k.trn+k)), c(v_c_,(1v+e1t)+8,*r,y,_r)\_f ,...,
* J71,--x-4:]:!_,
J'
o'
"-'l'i, )nu,* B,t*-t !n-t ), c-, (rt +gpt* B,-*-r
t,,* )*y,,,0 + K
)o'
(n+k-l- p+2d )+l
2

*l'.0-'-0.'"

*u -Q - Bl,o-,

G-t

B,*r-t

_K-{V+ep)t
+
(n+ k

- t - p+ 2")

U

f'

(t:,0-, c-, g, *

p1.,1)]

(3.1 3)

Gotv+gp1

-

B|*, B,;J

The marginal conditional posterior predictive density of y,*1
is a univariate student,s t_distribution on (n+k_l_p+2a)
degrees of freedom with mode
(t
a1-o-,
G-' Bn*o-,)" (u:.r-, c' 1v + g1t1) and scale parameter
tr= -

_
T=

(n + k

K-(V+eu)rG^(V+Ou)
v\
-"
- t- p+ za)Q -

ffi+;J'

where

co =G-t *Q

-

Bl.o-,G-'8,-r-,)-'

(c-'ao

'1

3.5. Point forecast and forecast variance
For quadratic loss function, the point forecast of Y
, *p is the posterior mean of the marginal conditional posterior predictive,
that is:

a(a., t s; )=(, -

Bf,*o_, G-,

B,*o_,)-' (r:.0_, c-' 1v + gp 1)

(3.14)

and the forecast variance

of

I,

* r is:

K-(V+

(n+k-1- p+2a)
r*(v,,.r 1s')=

1r

Golv +gpy

+k-l-p+2a

=(n + k

- 3-

p

+ za)-' (x - p,

+

gp 1, c o1v

+

gp 1)

Q

- ul,*,

G -,

B,*o_,),

(3. 1 5)

4. Conclusion and computational procedure
This paper analyzes how to find mathematical expression ofthe point forecast in equation (3.14)
and the forecast variance
in equation (3.15) based the marginal conditional posterior prediitive density in aquation
The conditional posterior
f:.r:1.
predictive density be obtained by multiplying the normal-gamma prior to the condiiional piedictive
density. By iniegrating
the conditional posterior predictive density to paramaters Y and t respectively, obtained
tire marginal .onditionut po-sterior
predictive density. Furthermore, the point forecast and the forecast uuiiun.. are derived
based the mean and the variance of
marginal conditional posterior predictive density that has the univariate student's t-distribution. procedure
to compute the
point forecast and the forecast variance are as follows:
Defines:

Yp

!p+t

!n+r-:

lpt

Yp

!n+k,3

!t

y.

,vh+x t_p

::::

U=

ao
6



o_,

iii.
iv.

XO:

€n+k-2

o

€ n+k-3

p*z-q

en+k-t-q

:::

e

ll.

p*l




l';:.:,1

l:l
lr..r

I

It, Q, a and

B

B,* -, =\y *o -,' !.nr-l,..., / n+t-t-p, € n+k,t,

€ n+*-2,

..., r,*-,u)

Compute:
v.

W:ULf

vi.

V:Ux6

P:IY+Q
^.f'
viii. K =
yl * p'ep + 2p

vii.

ix.

R= B,,r_,& BI*_,

x.

G:P+R

xi.

G-1

xii.

Go

xiii.
xiv.

= G-t + Q - Bl.o_,G-'a,,r_,)-'(c-'nc-')

E(y,.k s: )
)

:

Q

- a:.r-, G-' B,*0,,)-' (u:.,,, c-, 1v + gp 1)

va,(r.ul $) = (, + k - 3 - p + 2a)-'

(*

- f n + ep

f

G,(v + ep ))

Q

-

B!,,r_, G-,

B.*_,)-,

References

lll

Alba, E. And Mendoza, M. Bayesian Forecasting Methods for Short
Time Series. Article

l2l

Amry, Z' and Baharum, A. (2015)' Bayesian Multiperiod Forecasting
for Arma Model under Jeftey,s prior.
Mathematics and Statistics 3(3), 65-70.

13l

Bain, L'J.

and

Engelhardt, M'

(2006). Introduction

to Probability and Mathematical statistics, 2nd. Duxbury

Press, Belmont, California.

I4l

Bijak, J' (2010). Bayesian Forecasting and Issues
Southampton, I-19.

of

Uncertainty. Centrefor population change, (Jniversity

of

t5l Broemeling,L.and Shaarawy(1988).TimeSer_iesA_BayesianAnalysisinTimeDomain, inBayesian
analysis of Time series and Dynamic Moders, (J.c, spail, ed.; l-2l,MarcelDekker,
New york.

(2004). optimql Statistical decisions.

t6l

DeGroot, M.H'

17l

{an' c' a,nd Yao, s'(2008). Bayesian Approach for ARMA Process and Its Application. International Business
Reseqrch,

John Wiley

&

Son. Inc. New yersey

1(4),49-55.

t8l

Gelman, A.(2008), objections to Bayesian

statistics, Boyesian Anatysis 3(3), 445_ 450.

t9] Geweke, J and whiteman, C., (2004), Bayesian Forecasting,Deparfment

of Economics, University of Iowa.

[10] Hussein, H' M' A' (2008)' Bayesian Analysis_of the Autoregressive-Moving Average Model
with Exogenous Inputs
Using Gibb Sampling. Journar of Appried sciences Reseirch, a(r2),
tss-5-rsg2.

ll

ll

Ismail' M'A' and Amin, A.A. (2010). Gibbs Sampling for SARMA
Models. paper, Faculty of Commerce, statistics
Department, Menoufia (Jniversity, Egpt, l_15.

tl2l

Ismail, M'A' and Charif, H'A. (2003). Bayesian Inference for Thresold
Moving Average Models. METR7NInternqtional Journal ofStatistics, vol.
Ilg_132.

LXI,

tl3l

Kleibergen, F. and Hoek, H.

(1996). Bayesian Analysis of ARMA model using Non informative prior.
I_24

Paper, Econometric Institute, Erasmus University, Rotterdqm,

ll4]
II

Lee,

J'c', Lin, T.I. and Hsu, Y.L' (2004). Bayesian Analysis of Box-cox transformed
linear mixed models with

ARMA(p,q) dependence. Journal of Statiitical Planning and Inference,
t33(2005), qiS-.aSl.

5]

Liu, S. I. ( 1995). Comparison of Forecasts for ARMA Models Between
A Random coefficient Approach and
A Bayesian Approach. Commun. Statist. _ Theory Meth., 24(2),
3l 9_333

tl6l Liu,

S' I. (1995). Bayesian

Multiperiod Forecasts

2 I 1_224.

[17]

Pole, A', west, M' and Harrison , J.
Chapman and Hall, New york.

tl8l

Ramachandran, K'M' and Tsokos, c.P.
Academic press. San Diego, California.

[19]

Safadi, T' and Morettin, P.A.

w (2006).

ARX Models . Ann. Inst. statist. Mqth. vol. 47, no.2,

(lgg4). Apptied Bayesian

(2000)'

Identification

Bayesian

of

Forecasting

and

Time series Analysis.

(2009). Mathemqtical statistics with Applications, Elsevier
Analysis

sankhya: The Indian Journar ofstatistics, vorume

I20l

for

6i,

of

series B,

Threshold Autoregressive Moving Average Models.

pt. 3, 353-37r.

Methods Applied
^u**t9ty'
9''
Intetnational conference on Probqbilistic uethLi-'ippti"d to power to Srreamflow Data.
systems KTH ( l l-15

June 2006, Stockholm, Sweden), t_7

ARMA Model by Bayesian

(t)

o
FcsP

jj

\a
(g

7-E

ER
()rJ

9r,
cg

>8

({< ()

vc)

€>C/]

2!u
==

c.l

\z

^\JL.r

at

cd

TIr

&
C)

.l_)

(-

\/
a

C.)

^7

P
F

63

.Fl

H

tr

,k

C€

C)

c+

f-l

U
&

a1
ts

o

flfltl-.I
t)''v4

k
I

$-r

€tn
=F
;cg9
h€cl
qo".i

NE-^c..
-sxg:*
\'vrx5

lL-

s\

ru

e

'F!

\h
*)trtr

-'1

{t!20.8
Ql
q.s2o

-SLo
*t 'i
-!=:''e'==
-\

r-.,,

i6 E
{_)

6

i-

o
-l

F
bO

=,
rh

--=
-i-t €

SFTfi ?;e

FN
q
Itt

l\

C)

ca

a

rfrrt

E-

r

=
CA

@
s

t{
tv
-)
()

.E

E
e

H

C)

io

r,'l

;

5
E
o

@

r-l

:

r,,:;ffi

;E

F

i)
ll

E

ud

'F{

.\---'-.,

C)

tn
t-i
IJ

'..--c=
o

lt
\\

Lampiran 8
LEMBAR
HASIL PENITAIAN SEJAWAT SEBIDANG ATAU PEER REVIEW
KARYA ILMIAH : PROSIDING INTERNASIONAL

"Bayes Estimation for ARMA Model Forecasting Under Normal-Gamma

Judul Makalah

Prior"
Penulis Makalah

Zul Amry and, Adam Baharum

ldentitas Makalah

a. Judul Prosiding

lnternational Seminar on Operational Research
(lnteriOR 2015)

b. ISBN
c. Tahun Terbit

d. Penerbit
e. Jumlah halaman

f.
Kategori Publikasi Makalah

(beri /pada kategori yang tepat)
Hasil Penilaia n Peer Review

August 2015
USU Medan

WEB Laman

{
l-l

Prosiding Forum llmiah lnternasional
Prosiding Forum llmiah Nasional

:

Komponen
Yang Dinilai

Nilai Maksimal Prosiding
lnternasional
Nasaional

{

Nilai Akhir Yang

E

Diperoleh

0,f

d.

Kelengkapan unsur isi artikel (10%)

&f

b.

Ruang lingkup dan kedalaman pembahasan (30%)

2,7

2,f

c.

Kecukupan dan kemutahiran data/informasi dan
metodoloei (30%)
Kelengkapan unsur dan kualitas penerbit (30%)

7,V

2

2,7

2,f

d.

Totat =

f

(lOO%l

/,d

Nilai Pengusul

Medan,

September 2016

Reviewer 1,

\'\
ri
JJ

M.S
986011001
rsitas Sumatera Utara

Prof. Dr. Tulus, M.Si
NrP. 19620901 198803 1002

Unit Kerja : Guru Besar FMIPA USU

Lampiran 8
TEMBAR
HASIL PENILAIAN SEJAWAT SEBIDANG ATAU PEER REVIEW
KARYA ILMIAH : PROSIDING INTERNASIONAL

"Bayes Estimation for ARMA Model Forecasting Under Normal-Gamma

Judul Makalah

Prior"
Penulis Makalah

Zul Amry and, Adam Baharum

ldentitas Makalah

a. Judul Prosiding

lnternational Seminar on Operational Research
(lnteriOR 2015)

b. ISBN
c. Tahun Terbit

d. Penerbit
e. Jumlah halaman

f.
Kategori Publikasi Makalah
{beri /pada kategori yang tepat)
Hasil Penilaia n Peer Review

WEB Laman

{
|_l

Prosiding Forum llmiah lnternasional
Prosiding Forum llmiah Nasional

:

Komponen
Yang Dinilai
a.

Kelengkapan unsur isi artikel (10%)

b.

Ruang lingkup dan kedalaman pembahasan (30%)

c.

Kecukupan dan kemutahiran data/informasi dan

metodoloei(30%)
d.

Kelengkapan unsur dan kualitas penerbit (30%)

Total =

August 201.5
USU Medan

(LOO%I

Nilai Maksimal Prosiding
lnternasional
Nasaional
./

tr

Nilai Akhir Yang
Diperoleh

0,9

a(

L+

2,t

AT

248

v{

2,+

I

8

Nilai Pengusul

Medan,

September 2016
Reviewer 2,

Mengetahui:
,E€kanF,KlP

N Medan,

Dr. Firmansyah, M.Si

NtP, 19551-0251985031002
Unit Kerjai-Univ. Muslim Nusantara

NrP. 19671110 199303 1003

Unit Kerja: Univ. Muslim Nusantara

Lampiran

I

LEIVIBAR

HASIL PENILAIAN SEJAWAT SEBIDANG ATAU PTER REVIEW
KARYA ILMIAH : PROSIDING INTERNASIONAL

"Bayes Estirnation for ARMA Model Forecasting Under Normal-Gamma

Judul Makalah

Prior"
Penulis Makalah

Zul Amry and, Adam Baharum

ldentitas Makalah

a. Judul Prosiding

lnternational Seminar on Operational Research
(lnteriOR 2015i

b" tsBN
c. Tahun Terbit

d. Fenerbit
e. Jumlah halaman

f"
Kategori Publikasi Makalah

(beri "'pada kategori yang tepat)
Hasil Penilaian Peer Review

August 2015
USU Medan

WEB Laman

./
|_l

Prosiding Forum llmiah lnternasional
Prosiding Forum llmiah Nasional

:

Nilai Maks
Komponen
Yang Dinilai

;:lK"l""s@
b.
c.

d.

l"t"r*si"*a
./

Nasaional

n

Nilai Akhir Yang
Diperoleh

\f

otl

Ruang lingkup dan kedalaman pembahasan (30%)

g,T

L,f

Kecukupan dan kemutahiran data/informasi dan
metodoloei (30%)
Kelengkapan unsur dan kualitas penerbit (30%)

a,T

Ltl

L,T

L

Totat =

(LOO%I

L

Nilai Pengusul

Medan, September 2016
Reviewe 3

NtP. 19570804 198503 1002
Unit Kerja : Guru Besar FMIPA UNIMED