Graeme Walsh
May 14, 2013
Abstract
The following document is meant to give an brief indication of how to
go about producing point and interval forecasts from an ARIMA model.
The examples in the document are based on an AR(1) model. The same
method presented here can be used produce forecasts from other ARIMA
models. Note, however, that the equations will be different for each model
and that a little bit of algebra is required. This document is a far from
complete treatment of the subject of forecasting ARIMA models. I refer
the interested reader to Pankratz (1983)1 for more details.
Introduction
Suppose we have identified an AR(1) model to represent a time-series. The
model can be written as:
zt = C + 1 zt1 + at
(1)
Point Forecasts
The easiest way to generate point forecasts is to work with the model in difference form as shown in Equation 1.
We need to write an expression for our first forecast. We do this by altering
the subscripts in Equation 1 as follows:
zt+1 = C + 1 zt + at+1
(2)
1 Pankratz, A., 1983. Forecasting with univariate Box-Jenkins models: Concepts and cases.
New York: Wiley.
Now, our forecast of zt+1 , denoted zt (1), is the conditional mathematical expectation of zt+1 given information at time t, or, It . That is:
zt (1)
= E(zt+1 |It )
(3)
= C + 1 z t
(4)
Notice that at+1 is unknown at time t, hence we set its expected value equal to
zero.
A similar process can be used to find more point forecasts. Lets find the
next three point forecasts.
The first one is given by:
zt+2
C + 1 zt+1 + at+2
(5)
zt (2)
= C + 1 zt (1) + at+2
(6)
= E(zt+1 |It )
(7)
= C + 1 zt (1)
(8)
zt+3
= C + 1 zt+2 + at+3
(9)
= C + 1 zt (2) + at+3
(10)
zt (3)
= E(zt+2 |It )
(11)
= C + 1 zt (2)
(12)
= C + 1 zt+3 + at+4
(13)
= C + 1 zt (3) + at+4
(14)
= E(zt+3 |It )
(15)
= C + 1 zt (3)
(16)
Example
Using the data, we can now create four point forecasts. They are found as
follows.
From Equation 4:
z61 = z60 (1)
= C + 1 z60
(17)
1.92091 + (0.690)(6.2)
(18)
6.198910
(19)
From Equation 8:
z62 = z60 (2)
= C + 1 z61
(20)
1.92091 + (0.690)(6.198910)
(21)
6.1982
(22)
= C + 1 z62
(23)
1.92091 + (0.690)(6.1982)
(24)
6.1977
(25)
= C + 1 z63
(26)
1.92091 + (0.690)(6.1977)
(27)
6.1973
(28)
(29)
Example
Using our data, this means that:
0
(30)
1 = 0.690
21
31
(31)
2
= (0.690) = 0.4761
(32)
= (0.690)3 = 0.3285
(33)
(34)
where a is the standard deviation of the shocks. In practice, the RMSE can
replace a .
Example
Assume that the RMSE for our estimated AR(1) is 3.37903. The estimated standard deviation of the forecast errors for the first four forecasts can be calculated
as follows.
The first one is:
[et (1)]
=
a (1)1/2
(35)
(36)
3.37903
[et (2)]
(37)
2 1/2
3.37903(1 + (0.690) )
(38)
4.1053
(39)
[et (3)]
a (1 + 12 + 22 )1/2
(40)
2
2 1/2
(41)
4.4093
(42)
[et (4)]
=
a (1 + 12 + 22 + 32 ))1/2
(43)
(44)
4.5469
(45)
(46)
(47)
Example
Using our data, the 80% confidence interval forecasts are as follows:
z60 (1)
6.198910
1.28
[e60 (1)]
(48)
1.28(3.37903)
(49)
z60 (2)
1.28
[e60 (2)]
(50)
6.198910
1.28(4.1053)
(51)
z60 (3)
1.28
[e60 (3)]
(52)
6.198910
1.28(4.4093)
(53)
z60 (4)
1.28
[e60 (4)]
(54)
6.198910
1.28(4.5469)
(55)