EXPONENTIAL

SMOOTHING
METHODS
There are two smoothing classification of
forecasting methods, averaging methods and
exponential smoothing methods.

If a time series is generated by a constant process

subject to random error (or noise), then the mean is
a useful statistic and can be used as a forecast for
the next period(s).

However, if the time series involves a trend, or a

seasonal effect, or both a trend and a seasonal
effect, then the simple average is no longer able to
capture the data pattern.
In such situations we have to consider a variety of
smoothing methods that seek to improve upon the
mean as the forecast for the next period(s).

The methods apply an unequal set of weights to

past data, and because the weights typically decay
in an exponential manner from the most recent to
the most distant data point, the methods are known
as exponential smoothing methods. All methods in
this group require that certain parameters be
defined, and these parameter values lie between 0
and 1.
In averaging method we have:
a) Simple average
b) Moving average

In exponential smoothing method we have

a) Single exponential smoothing
b) Holt’s linear model
c) Holt-Winters’ method
d) Pagels’ classification
The forecasting scenario

On the time scale we start from a certain point

called the point of reference. Once a forecasting
model has been selected, we fit the model to the
known data and obtain the fitted values.

For the known observations this allows calculation

of fitted errors. If the new observations become
available we can examine forecasting errors.
Strategy for evaluating any forecasting
methodology

Stage 1: Choose a time-series (data set) to analyze.

Divide this into an initialization set and a test set.

Stage 2: Choose a smoothing method.

Stage 3: Initialize the method. Use the

initialization data set.
Stage 4: Use the smoothing method to forecast
over the test set.
Test measures: MAPE, MSE, etc.
Optimize the values of parameters.

Stage 5: Appraisal decision.

Pros and cons.
Applications potential.
Averaging methods

The mean: The mean forecast is defined as

When a new observation, Yt+1, becomes available,

the forecast for time t + 2 is the new mean
including the previously observed data plus this
new observation:
The recursive form can be written as

This simple averaging process produce good results

if the process underlying the observed Y values has
no noticeable trend, and has no noticeable
seasonality.
Moving averages

A moving average forecast of order k, or MA(k), is

given by

This use of moving averages is slightly different

from the use of moving averages in smoothing
technique.
In smoothing technique we estimated the trend-
cycle in a data series by taking an average of
nearby points.

Here we are forecasting the next observation by

taking an average of the most recent observations.

To avoid confusion, we use MA(k) to denote a

moving average forecast of order k and k MA to
denote a moving average smoother of order k.
Algebraically, the moving average can be written
as follows:
It can be seen that the more observations included
in the moving average, the greater the smoothing
effect.

A forecaster must choose the number of periods (k)

in a moving average. The two extreme cases are k
= 1 and k = n.
MA(1): that is, a moving average of order 1- the
last known data point (Yt) is taken as the forecast
for the next period (Ft+1 = Yt).

MA(n): that is, the mean of all observations is used

as a forecast. So this is equivalent to the mean
forecast method.

In practice, the technique of moving averages as a

forecasting procedure is not used often because the
methods of exponential smoothing are generally
superior.
Exponential smoothing methods

With simple moving average forecasts, the mean of

the past k observations are used as a forecast. This
implies equal weights (equal to 1/k) for all k data
points.

However, with forecasting, the most recent

observations will usually provide the best guide as
to the future. In exponential smoothing procedures,
recent values are given relatively more weight in
forecasting than the older values.
Single exponential smoothing

The forecast for the time t+1 is:

Where α is a constant between 0 and 1. The

forecast for the time t+1 can also be written as:
So Ft+1 represents a weighted moving average of all
past observations. Suppose α = 0.2, 0.4, 0.6, or 0.8.
Then the weights assigned to past observations
would be as follows:
Weights given to past data when a forecast is made
at time t for the next period, using various
forecasting procedures.
One point of concern relates to the initializing
phase of SES. For example, to get the SES
forecasting system started we need F1 because

Since the value for F1 is not known, we can use the

first observed value (Y1) as the first forecast (F1 =
Y1) and then proceed.

Another possibility would be to average the first

four or five values in the data set and use this as the
initial forecast.
The choice of α has considerable impact on the
forecasting. When a small value of α is chosen, the
initial forecast plays a more prominent role than
when a larger α is used.

A large value of α gives very little smoothing in the

forecast, whereas a small value of α gives
considerable smoothing.

When α=1, the exponential smoothing is equivalent

to using the last observation as a forecast.
One way to choose α is to calculate the MSE for a
grid of values (e.g., α = 0.1, 0.2, . . . , 0.9), and
choose the value that yields the smallest MSE
value.

The application of single exponential smoothing

can be illustrated by using the electric can opener
example using α = 0.1, 0.5, and 0.9.

Single exponential smoothing works best for data

which have no trend, no seasonality or other
underlying pattern.
Exercise

The Canadian unemployment rate as a percentage

of the civilian labor force (seasonally adjusted)
between 1974 and the third quarter of 1975 is
shown below.
(a) Estimate unemployment in the fourth quarter of
1975 using a single moving average with k = 3.
(b) Repeat using single exponential smoothing with
α = 0.7.
(c) Compare your two estimates using the accuracy
statistics.
Holt's linear method

Holt (1957) extended single exponential smoothing

to linear exponential smoothing to allow
forecasting of data with trends.

The forecast for Holt's linear exponential

smoothing is found using two smoothing constants,
α and β (with values between 0 and 1), and three
equations:
Here Lt denotes an estimate of the level of the
series at time t and bt denotes an estimate of the
slope of the series at time t and m denotes number
of periods ahead to be forecasted.

The initialization process for Holt's linear

exponential smoothing initialization requires two
estimates.
One to get the first smoothed value for L1 and the
other to get the trend b1. One alternative is to set L1
= Y1 and

As with single exponential smoothing, the weights

α and β can be chosen by minimizing the value of
MSE.

Holt's method is sometimes called double

exponential smoothing.
In the special case where α = β, the method is
equivalent to Brown's double exponential
smoothing.

Using the inventory demand data the following

table shows the application of Holt's linear
smoothing to a series with trend using α = 0.501
and β = 0.072.
Exercise

Using the single randomless series 2, 4, 6, 8, 10,

12, 14, 16, 18 and 20, compute a forecast for period
11 using Holt’s linear method assuming α = 0.5
and β = 0.5.
Holt-Winters' trend and seasonality method

All the methods explained can be used with almost

any type of data that are non-seasonal. When
seasonality exists, these methods are not
appropriate. Holt's method extended by Winters
(1960) can be used to capture seasonality.

The Holt-Winters' method is based on three

smoothing equations - one for the level, one for
trend, and one for seasonality.
There are two different Holt-Winters' methods,
depending on whether seasonality is modeled in an
additive or multiplicative way.

An additive model is appropriate if the magnitude

of the seasonal fluctuations does not vary with the
level of the series. But if the seasonal fluctuations
increase and decrease proportionally with increases
and decreases in the level of the series, then a
multiplicative model is appropriate.
Consider the following data for quarterly exports of
a French company over a six year period
Multiplicative seasonality

The basic equations for Holt-Winters'

multiplicative method are as follows:
Lt represents the level of the series
bt denotes the trend
St is the seasonal component, and
Ft+m is the forecast for m periods ahead

To initialize the Holt-Winters' forecasting method,

we need initial values of the level Lt, the trend bt,
and the seasonal indices St.
To determine initial estimates of the seasonal
indices we need to use at least one complete
season's data (i.e., s periods). Therefore we
initialize trend and level at period s. The level is
initialized by taking the average of the first season:

To initialize trend, it is convenient to use two

complete seasons (i.e., 2s periods) as follows:
The seasonal indices are initialized using the ratio
of the first few data values to the mean of the first
year so that

The parameters α, β, and γ can be chosen to

minimize MSE.

The application of Holt-Winters’ multiplicative

seasonality can be illustrated by using quarterly
sales data considering α = 0.822, β = 0.055 and γ =
.000 chosen by minimizing the MSE.
For this data we obtain the following values

L4 = 380
b4 = 9.75
S1 = .953
S2 = 1.013
S3 = 1.137
S4 = .897
S=4
Application of single exponential smoothing and
Holt's method to quarterly sales data. Smoothing
parameters were chosen by minimizing the MSE
(SES: α = 0.464; Holt's: α = 0.065, β = 0.334).
Additive seasonality

The seasonal component in Holt-Winters' method

may also be treated additively, although this is less
common. The basic equations for Holt-Winters'
additive method are as follows:
The initial values for Ls and bs are identical to
those for the multiplicative method. To initialize
the seasonal indices we use
Exercise

Apply Holt-Winters’ additive seasonality to the

quarterly sales data considering α = 0.353, β =
0.114 and γ = 1. Find the following:

(i) Seasonal indices for the four quarters

(ii) F5, L5, b5, S5 and F6
Exponential smoothing: Pegels' classification

Pegels (1969) provided a simple framework for

discussing exponential smoothing methods having
separate trend and seasonal aspects in a two-way
classification as follows
All nine exponential smoothing models can be
summarized by the formulas:

where P, Q, R, and T vary according to which of

the method is used. Following table shows the
Pegels' classification scheme for calculations and
forecasting using different methods.
Consider cell C-3, which refers to an exponential
smoothing model that allows for multiplicative
trend and multiplicative seasonality. From Pegels’
classification scheme, we obtain the following
formulas:
A comparison of methods

How can a forecaster choose the right model for a

data set?

Human judgment has to be involved, but there are

also some useful suggestions to make. A main
objective is to decide on the nature of trend,
seasonality and the strength of the random
component.
Following table presents the results of 11 different
analyses of the data set for quarterly exports of a
French company over a six year period.

In fitting each model, parameters have been chosen

by minimizing the MSE calculated using errors
from periods 2 through 24 (5 through 24 for
seasonal models).
General aspects of smoothing methods

The major advantages of widely used smoothing

methods are their simplicity and low cost. When
forecasts are needed for thousands of items, as is
the case in many inventory systems, smoothing
methods are often the only methods fast enough for
acceptable implementation.

In implementing smoothing methods, there are

three practical issues which need to be addressed:
initialization, optimization, and prediction
intervals.

Initialization

The reason initial values for the exponential

smoothing methods are needed is that the methods
are recursive equations, and so they need to start
somewhere. The number and type of values depend
upon the particular exponential smoothing
approach being used.
Optimization

All the exponential smoothing methods require

specification of some smoothing parameters.

Not long ago, the computer time needed to

optimize these parameters was sufficiently great
that methods involving more than one or two
parameters were not widely used and values of the
parameters were restricted to a small number of
possibilities (e.g., 0.1, 0.3, 0.5, 0.7, and 0.9).
With the advent of much faster computing, it is
relatively easy to choose optimal values of the
parameters.

All good forecasting packages will give optimal

parameter values automatically by minimizing the
MSE. It is possible to optimize some other
measurement of forecast error (such as MAE or
MAPE) but the MSE tends to be easier to work
with.
Prediction intervals

The forecasts considered so far have been point

forecasts; that is, single numbers that represent
what we think the value of the series will be in the
future. Sometimes, that is all that is required. But it
is often desirable to have a measure of the
uncertainty associated with that forecast.
To measure the uncertainty in forecasts, a
prediction interval is used which is a range in
which the forecaster can be fairly sure that the true
value lies.

Unfortunately, exponential smoothing methods do

not allow the easy calculation of prediction
intervals. One widely-used approach is to find a
statistical model for which a particular exponential
smoothing method is optimal. Then prediction
intervals can be obtained from the statistical model.