Vous êtes sur la page 1sur 18

Lecture 7:

Forecasting: Putting it ALL


together
The full model
The model with seasonality, quadratic trend, and ARMA
components can be written:



Ummmm, say what????
The autoregressive components allow us to control for the
fact that data is directly related to itself over time.
The moving average components, which are often less
important, can be used in instances where past errors are
expected to be useful in forecasting.

2

y
t
= b
1
D
1,t
+ ...+ b
s
D
s,t
+a
1
t +a
2
t
2
+ u
t
,
u
t
= |
1
u
t1
+ |
2
u
t2
+ ...+ |
p
u
tp
+ ...+
e
t
+u
1
e
t1
+ ...+u
q
e
tq
Model selection
Autocorrelation (AC) can be used to choose a model.
The autocorrelations measure any correlation or
persistence. For ARMA(p,q) models, autocorrelations
begin behaving like an AR(p) process after lag q.
Partial autocorrelations (PAC) only analyze direct
correlations. For ARMA(p,q) processes, PACs begin
behaving like an MA(q) process after lag p.
For AR(p) process, the autocorrelation is never
theoretically zero, but PAC cuts off after lag p.
For MA(q) process, the PAC is never theoretically zero,
but AC cuts off after lag q.

3
Model selection
An important statistic that can used in choosing a model
is the Schwarz Bayesian Information Criteria. It rewards
models that reduce the sum of squared errors, while
penalizing models with too many regressors.
SIC=log(SSE/T)+(k/T)log(T), where k is the number of
regressors.
The first part is our reward for reducing the sum of
squared errors. The second part is our penalty for
adding regressors. We prefer smaller numbers to larger
number (-17 is smaller than -10).
4
Important commands in EViews
ar(1): Includes a single autoregressive lag
ar(2): Includes a second autoregressive lag
Note, if you include only ar(2), EViews will not include a
first order autoregressive lag
ma(1): Includes a first order moving average term. This
is not the same as forecasting using an average of
recent values
5
Selecting an appropriate time
series model, concluded
Determine if trend/seasonality is important
If it is, include it in your model
Estimate the model with necessary trend/seasonal
components. Look at the correlogram of the residuals:
From the equation dialogue box:
View => Residual Tests => Correlogram Q-statistics
If ACs decay slowly with abrupt cutoff in PAC, this is indicative
of AR components. If the PAC doesnt cutoff, you may need to
include MA components as well.
Re-estimate the full model with trend/seasonality
included with necessary ARMA components. You will
likely have several models to choose from.
6
Selecting an appropriate model,
cont.
After you estimate each model, record SIC/AIC values
Use the SIC/AIC values to select an appropriate model.
Finally, investigate the final set of residuals. There
should be no correlation in your residuals.
Evidenced by individual correlation coefficients within 95%
confidence intervals about zero.
Ljung-Box Q-statistics should be small with probability
values typically in excess of 0.05.
7
Exponential smoothing
Very useful when we have only a handful of observations
Exponential smoothing can be modified to account for
trend and seasonality.
If you suspect your data does not contain
trend/seasonality, simply use single exponential
smoothing:

The forecast h periods into the future is constant and
given by:


8


y
t
=o y
t
+(1o)y
t1
+....+(1o)
t1
y
1
| |


y
T +h|T
=

y
T
Exponential smoothing with trend
Obtain the smoothed level series, L
t
:
L
t
=oy
t
+(1-o)(L
t-1
+T
t-1
)
The trend series, T
t
is formed as:
T
t
=|(L
t
-L
t-1
)+(1-|)(T
t-1
)
The forecasted series:

9


y
T +h|T
= L
T
+ hT
T
Example:
Suppose the first three observations of a series are 500,
350, and 250. Suppose o=0.30, |=0.10. Suppose the
trend and level are initialized such that T
1
=0, L
1
=500.
L
2
=0.30 (350) +(1-0.30)(500-0) = 455
T
2
=0.10(455-500) + 0.90 (0) = -4.5.
The forecasted value for the third period then becomes:

10


y
T +1|T
= L
T
+1T
T
= 455 4.5 = 450.5
Exponential smoothing: Trend
and seasonality
Eviews also allows for exponential smoothing with trend
and seasonality.
With seasonality we use a smoothed series along with
estimates based on trends and seasonality.
Which option should I select?
If you believe your series lacks either seasonality or trend,
single smoothing works perfectly.
From visual inspection of your series, if only trend appears to
be present, you will need either double smoothing or Holt-
Winters with no seasonal.
If seasonality and trend are expected, you will need to use
Holt-Winters with the allowance of multiplicative or additive
seasonality.
11
HUH?
Eviews provides five options when you ask it, no tell it,
to provide exponential smoothing:
Single: (no seasonality/no trend)
Double: (trend value of o=|).
Holt-Winters No seasonal (Trend, o and | are not equal,
but are estimated in the data).
Holt-Winters Additive (Trend and Seasonality. The
seasonal component is estimated with an additive filter).
Holt-Winters - Multiplicative
12
Breaks?
Uh oh? My data appears to have a break.
The developed time series methods assume the
black box generating the data is constant.
Not necessarily true:
Learning curves may cause cost curves to
decrease
Acquisition of companies or new technologies
may alter sales/costs

13
Dealing with breaks?
Solutions:
Limit the sample to the post break period
Sometimes taking logs and/or differencing can help
mitigate the effects of breaks/outliers.
Include variables that help identify the breaks
Model the breaks directly:
The most obvious way is to include a break in mean and/or a
break in trend.
We should make sure that the included break is modeled in a
sensible way
A negative linear trend, for example, will imply the data
may eventually turn negative.
14
Break in mean
15
0 20 40 60 80 100 120 140 160 180 200
0
10
20
30
40
50
60
70
80
90
100
time
y
Break in trend
16
0 20 40 60 80 100 120 140 160 180 200
0
10
20
30
40
50
60
70
80
90
Statistics useful in comparing the
out of sample forecasting accuracy
Mean squared error: For an h-step extrapolation
forecast:

Root mean squared error is the square root of this
number.
Mean absolute error
17

(

Y
T +1|T
Y
T +1
)
2
+ (

Y
T +2|T
Y
T +2
)
2
+ ...+ (

Y
T +h|T
Y
T +h
)
2
h

|

Y
T +1|T
Y
T +1
| + |

Y
T +2|T
Y
T +2
| +...+ |

Y
T +h|T
Y
T +h
|
h
In Eviews:
If you have a forecasted series, say xf, and an original
series x, you can calculate the mean squared error as:
genr mse=@sum((x-xf)^2)/h
To calculate the moving average forecasts:
Suppose you use the most recent four periods
Limit your data set to include only the last four observations
A variable called maf_4 is calculated by:
genr maf_4=@mean(x)
18

Vous aimerez peut-être aussi