Econ 582

Forecasting

Eric Zivot

April 15, 2013

Forecasting

Let {} be a covariance stationary are ergodic process, e.g. an ARMA( )

process with Wold representation

X
= + ~ (0 2)
=0
= + + 11 + 22 +
Let = { 1 } denote the information set available at time . Recall,

[] =

X
() = 2 2
=0
Goal: Using produce optimal forecasts of + for = 1 2
 + = + + + 1+1 +
+ 1+1 + + +11 +
Define +| as the forecast of + based on with known parameters. The
forecast error is
+| = + +|
and the mean squared error of the forecast is

(+|) = [2+|]
= [(+ +|)2]
Theorem: The minimum MSE forecast (best forecast) of + based on is

+| = [+|]
Proof: See Hamilton pages 72-73.
Remarks

1. The computation of [+|] depends on the distribution of {} and

may be a very complicated nonlinear function of the history of {} Even if
{} is an uncorrelated process (e.g. white noise) it may be the case that

[+1|] 6= 0

2. If {} is independent white noise, then [+1|] = 0 and [+|] will

be a simple linear function of {}

+| = + + +11 +
Linear Predictors

+ = + + + 1+1 +
+ 1+1 + + +11 +

A linear predictor of +| is a linear function of the variables in

Theorem: The minimum MSE linear forecast (best linear predictor) of +

based on is
+| = + + +11 +

Proof. See Hamilton page 74.

The forecast error of the best linear predictor is

+| = + +|
= + + + 1+1 +
+1+1 + +
( + + +11 + )
= + + 1+1 + + 1+1
and the MSE of the forecast error is

MSE(+|) = 2(1 + 21 + + 21)

Remarks

1. [+|] = 0

2. +| is uncorrelated with any element in

3. The form of +| is closely related to the IRF

4. (+|) = (+|) ()

5. lim +| =

6. lim (+|) = ()
Example: BLP for MA(1) process

= + + 1 WN(0 2)
Here
1 = = 0 for 1
Therefore,

+1| = +
+2| =
+| = for 1
The forecast errors and MSEs are

+1| = +1 MSE(+1|) = 2
+2| = +2 + +1 MSE(+2|) = 2(1 + 2)
Prediction Confidence Intervals

If {} is Gaussian then

+| (+| 2(1 + 21 + + 21))

A 95% confidence interval for the step prediction has the form
q
+| 196 2(1 + 21 + + 21)
Predictions with Estimated Parameters

Let +| denote the BLP with estimated parameters:

+| = + + +11 +
where is the estimated residual from the fitted model. The forecast error
with estimated parameters is

+| = + +|
= ( ) + + + 1+1 + + 1+1

+ + +11 +11
+
Obviously,

MSE(+|) 6= MSE(+|) = 2(1 + 21 + + 21)

Note: Most software computes
d 2 2
2(1 + + +
MSE(+| ) = 1 1)
Computing the Best Linear Predictor

The BLP +| may be computed in many dierent but equivalent ways.

The algorithm for computing +| from an AR(1) model is simple and the
methodology allows for the computation of forecasts for general ARMA models
as well as multivariate models.

Example: AR(1) Model

= (1 ) +
~ (0 2)
2 are known
In the Wold representation = Starting at and iterating forward
periods gives

+ = + ( ) + + + +1 +
+1+1
= + ( ) + + + 1+1 +
+1+1
Based on information at time , the best forecast for +1 + is zero
because (0 2) Hence,

+| = + ( ) = 1 2
The best linear forecasts of +1 +2 + can be recursively computed
using the chain-rule of forecasting (law of iterated projections)

+1| = + ( )
+2| = + (+1| ) = + (( ))
= + 2( )
..
+| = + (+1| ) = + ( )
The corresponding forecast errors are

+1| = +1 +1| = +1
+2| = +2 +2| = +2 + +1
= +2 + 1+1
..
+| = + +| = + + +1 +
+1+1
= + + 1+1 + + 1+1
The forecast error variances are
(+1|) = 2
(+2|) = 2(1 + 2) = 2(1 + 21)
..
1 2
(+|) = 2(1 + 2 + + 2(1)) = 2
1 2
= 2(1 + 21 + + 21)
Clearly,
lim +| = = []

2
lim (+|) =
1 2

X
= 2 2 = ()
=0
AR(p) Models

Consider the AR(p) model

()( ) = (0 2)
() = 1 1
The forecasting algorithm for the AR() models is essentially the same as that
for AR(1) models once we put the AR() model in state space form. Let
= The AR() in state space form is

1 2 1

1 1 0 0 2 0

.. = ... .. .. + ..

+1 0 1 0 0
or
= F1+w
(w) =
Starting at and iterating forward periods gives
+ = F + w+ + Fw+1 + + F1w+1
Then the best linear forecasts of +1 +2 + are computed using the
chain-rule of forecasting are
+1| = F
+2| = F+1|= F2
..
+| = F+1|= F
The forecast for + is given by plus the first row of +| = F :

1 2

1 0 0
1


+| = F = ... .. ..

0 1 0 +1
The forecast errors are given by

w+1| = +1 +1| = w+1

w+2| = +2 +2| = w+2 + Fw+1
..
w+| = + +| = w+ + Fw+1 +
+F1w+1
and the corresponding forecast MSE matrices are

(w+1|) = (w+1) =
(w+2|) = (w+2) + F(w+1)F0
= + F F0
..
1
X 0
(w+|) =
F F
=0
Notice that
(w+|) = + F(w+1|)F0