Forecasting

Exponential Smoothing
For
Stationary Models

The Last Period method uses only one
period (the last) and the n-Period Moving
Average and Weighted Moving methods
use only the last n periods to make
forecasts the rest of the data is ignored.

Exponential Smoothing uses all the time
series values to generate a forecast with
lesser weights given to the observations
further back in time.

Exponential Smoothing
Basic Concept
Exponential smoothing is actually a way of
smoothing out the data by eliminating
much of the noise (random effects).
At each period t, an exponentially smoothed
level, L
t
, is calculated which updates the
previous level, L
t-1
, as the best current
estimate of the unknown constant level,
0
,
of the time series by the following formula:

L
t
= y
t
+ (1-)L
t-1

Revised Estimate of
the Level at time t
Weight placed on current
time series value
Weight placed on last
estimate for the Level
Current time
series value
Last estimate
for the Level
in Exponential Smoothing
The idea behind smoothing the data is to
get a more realistic idea about what is
really going on.
The value of the smoothing constant, , is
selected by the modeler.
Higher values of allow the time series to be
swayed quickly by the most recent observation.
Lower values keep the smoothed time series
flatter as not that much weight will be given to the
most recent observation.
Usual values of are between about .1 and .7
See graphs for = .1 and = .7 later in this module.
The value (1-) is called the damping factor.
Using Exponential Smoothing to Prepare
Forecasts in Stationary Models
The Level, L
t
, calculated at time period t is
the best estimate at time t for the unknown
constant,
0
.
Since that is the best estimate of
0
, it will
be the forecast for the next data value of
the time series, F
t+1
.


Since the model is stationary, it will be the
forecast for all future time periods until
more time series data is observed.
F
t+1
= L
t

Once a value of has been selected, the Level (or
smoothed value) at time t depends on only two
values --
The current periods actual value (y
t
) with weight of a.
The forecast value for the current period (which is the
level at the previous period, L
t-1
) with weight of 1-a.

Calculations then, for L
t
(and hence for F
t+1
) are
very simple.
Initialization Step
There is no L
0
. So we cannot calculate L
1
by y
1
+ (1- )L
0

Since y
1
is the only value known after period 1, set:


Exponential Smoothing Technique
Initialization Step
L
1
= y
1

Sample Calculations for First Four
Periods of Yoho Data
The first four values of the time series
for the Yoho yoyo time series were:
415, 236, 348, 272
Suppose we have selected to use a
smoothing constant of = .1.
Initialization Period 1
L
1
= y
1
= 415 -- the level for week 1 is 415
F
2
= L
1
= 415 -- the forecast for week 2 is 415

Continued

Week 2
L
2
= .1y
2
+ .9L
1
= .1(236) + .9(415) = 397.1
The smoothed (leveled) value for week 2 is 397.1
F
3
= L
2
= 397.1 The forecast for week 3 is 397.1

Week 3
L
3
= .1y
3
+ .9L
2
= .1(348) + .9(397.1) = 392.19
The smoothed (leveled) value for week 3 is 392.19
F
4
= L
3
= 392.19 The forecast for week 4 is 392.19

Week 4
L
4
= .1y
4
+ .9L
3
= .1(272) + .9(392.19) = 380.171
The smoothed (leveled) value for week 4 is 380.171
F
5
= L
4
= 380.171 The forecast for week 5 is 380.171

Excel Exponential Smoothing

Note:
Rows 8-43
are hidden
=B2
=.1*B3+.9*C2
=D54
Drag C3 down
to C53
Drag D3 down
to D54
Drag D55 down
to D56
=C3
How Exponential Smoothing Uses
All Previous Time Series Values
Recall that the recursive formula used is:
L
t
= y
t
+ (1-)L
t-1
This means:
L
t-1
= y
t-1
+ (1-)L
t-2
L
t-2
= y
t-2
+ (1-)L
t-3
L
t-3
= y
t-3
+ (1-)L
t-4
Etc.
Substituting, L
t
= y
t
+ (1-)L
t-1
= y
t
+ (1-)(y
t-1
+ (1-)L
t-2
) =
= y
t
+ (1-)y
t-1
+ (1-)
2
L
t-2
=
= y
t
+ (1-)y
t-1
+ (1-)
2
y
t-2
+ (1-)
3
L
t-3


= y
t
+ (1-)y
t-1
+ (1-)
2
y
t-2
+ (1-)
3
y
t-3
+ (1-)
4
L
t-4
Etc.
Thus all time series values, y
t
, y
t-1
, y
t-2
, y
t-3
, etc. will be included with
successive weights reduced (dampened) by a factor of (1-).
Exponential Smoothing ( = .1)
How Much Smoothing Is There?
We said the lower the value of , the more
smooth the time series will become.
Actual Data
Smoothed time
series with = .1
A flat smoothed series
What About Larger Values of ?
Here is the smoothed series for = .7:
Exponential Smoothing ( = .7)
Actual Data
Smoothed time
series with = .7
Very sensitive to most recent time
series value not much smoothing
What Value of Should Be Used?
Up to the modeler
If the modeler is considering several
values of , a forecast using each value
could be prepared.
Only consider values of that would give
useful results (not = 0, for instance)
Then a performance measure (MSE, MAD,
MAPE, LAD) could be used to determine
which of the values of that are being
considered have the lowest value of the
selected performance measure.

Review
Exponential smoothing is a way to take some of
the random effects out of the time series by using
all time series values up to the current period.
The smoothed value (Level) at time period t is:
(current value) + (1-)(last smoothed value)
Forecast for period t+1= Smoothed Value at t
Initialization:
First smoothed value = first actual time series value
The smaller the value of , the less movement in
the time series.
Excel approach to exponential smoothing