NCSS Statistical Software NCSS.

com

Chapter 467

Exponential Smoothing –
Trend & Seasonal
Introduction
This module forecasts seasonal series with upward or downward trends using the Holt-Winters exponential
smoothing algorithm. Two seasonal adjustment techniques are available: additive and multiplicative.

Additive Seasonality
Given observations X 1 , X 2 , , X t of a time series, the Holt-Winters additive seasonality algorithm computes an
evolving trend equation with a seasonal adjustment that is additive. Additive means that the amount of the
adjustment is constant for all levels (average value) of the series.
The forecasting algorithm makes use of the following formulas:
a t = α ( X t − Ft − s ) + (1 − α )(a t −1 + bt −1 )

bt = β (a t − a t −1 ) + (1 − β )bt −1

Ft = γ ( X t − a t ) + (1 − γ ) Ft − s

Here α , β , and γ are smoothing constants which are between zero and one. Again, a t gives the y-intercept (or
level) at time t, while bt is the slope at time t. The letter s represents the number of periods per year, so the
quarterly data is represented by s = 4 and monthly data is represented by s = 12.
The forecast at time T for the value at time T+k is aT + bT k + F[( T + k −1)/ s ]+1 . Here [(T+k-1)/s] is means the
remainder after dividing T+k-1 by s. That is, this function gives the season (month or quarter) that the observation
came from.

Multiplicative Seasonality
Given observations X 1 , X 2 , , X t of a time series, the Holt-Winters multiplicative seasonality algorithm
computes an evolving trend equation with a seasonal adjustment that is multiplicative. Multiplicative means that
the amount of the adjustment is varies with the level (average value) of the series. Note that the nature of most
economic time series make the multiplicative model more popular than the additive model.
The forecasting algorithm makes use of the following formulas:
a t = α ( X t / Ft − s ) + (1 − α )(a t −1 + bt −1 )

bt = β (a t − a t −1 ) + (1 − β )bt −1

Ft = γ ( X t / a t ) + (1 − γ ) Ft − s

467-1
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com
Exponential Smoothing – Trend & Seasonal

Here α , β , and γ are smoothing constants which are between zero and one. Again, a t gives the y-intercept (or
level) at time t, while bt is the slope at time t. The letter s represents the number of periods per year, so the
quarterly data is represented by s = 4 and monthly data is represented by s = 12.
The forecast at time T for the value at time T+k is (aT + bT k )F[( T + k −1) / s ]+1 . Here [(T+k-1)/s] is means the
remainder after dividing T+k-1 by s. That is, this function gives the season (month or quarter) that the observation
came from.

Smoothing Constants
Notice that the smoothing constants determines how fast the weights of the series decays. The values may be
chosen either subjectively or objectively. Values of a smoothing constant near one put almost all weight on the
most recent observations. Values of a smoothing constant near zero allow the distant past observations to have a
large influence.
Note that α is associated with the level of the series, β is associated with the trend, and γ is associated with the
seasonality factors.
When selecting the smoothing constant subjectively, you use your own experience with this, and similar, series.
Also, specifying the smoothing constant yourself lets you tune the forecast to your own beliefs about the future of
the series. If you believe that the mechanism generating the series has recently gone through some fundamental
changes, use a smoothing constant value of 0.9 which will cause distant observations to be ignored. If, however,
you think the series is fairly stable and only going through random fluctuations, use a value of 0.1.
To select the value of the smoothing constants objectively, you search for values that are best in some sense. Our
program searches for that values that minimize the size of the combined forecast errors of the currently available
series. Three methods of summarizing the amount of error in the forecasts are available: the mean square error
(MSE), the mean absolute error (MAE), and the mean absolute percent error (MAPE). The forecast error is the
difference between the forecast of the current period made at the last period and the value of the series at the
current period. This is written as
et = X t − Ft −1
Using this formulation, we can define the three error-size criterion as follows:

∑e
1
MSE = 2
t
n

∑e
1
MAE = t
n

∑X
100 et
MAPE =
n t

To find the value of the smoothing constants objectively, we select one of these criterion and search for those
values of α and β that minimize this function. The program conducts a search for the appropriate values using
an efficient grid-searching algorithm.

Initial Values
Winters method requires initialization since the forecast for period one requires the forecast at period zero, which we
do not, by definition, have. It also requires the seasonal adjustment factors. Several methods have been proposed
for generating starting values. NCSS uses the initialization method described in Bowerman and O’Connell (1993).

467-2
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com
Exponential Smoothing – Trend & Seasonal

Relationship to ARIMA Method

The multiplicative seasonal adjustment model does not have an ARIMA counterpart, while the additive model does.

Assumptions and Limitations

These algorithms are useful for forecasting seasonal time series with (local or global) trend.

Data Structure
The data are entered in a single variable.

Missing Values
When missing values are found in the series, they are either replaced or omitted. The replacement value is the
average of the nearest observation in the future and in the past or the nearest non-missing value in the past.
If you do not feel that this is a valid estimate of the missing value, you should manually enter a more reasonable
estimate before using the algorithm. These missing value replacement methods are particularly poor for seasonal
data. We recommend that you replace missing values manually before using the algorithm.

Procedure Options
This section describes the options available in this procedure.

Variables Tab
Specify the variable(s) on which to run the analysis.

Time Series Variables

Time Series Variable(s)
Specify the variable(s) on which to run the analysis. A separate analysis will be conducted for each variable listed.
Use Logarithms
Specifies that the log (base 10) transformation should be applied to the values of the variable. The forecasts are
converted back to there original metric before display.
Missing Values
Choose how missing (blank) values are processed.
The algorithm used in this procedure cannot tolerate missing values since each row is assumed to represent the
next point in a time sequence. Hence, when missing values are found, they must be removed either by imputation
(filling in with a reasonable value) or by skipping the row and pretending it does not exist.
Whenever possible, we recommend that you replace missing values manually.
Here are the available options.
Average the Adjacent Values
Replace the missing value with the average of the nearest values in the future (below) and in the past (above).

467-3
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com
Exponential Smoothing – Trend & Seasonal

Carry the Previous Value Forward

Replace the missing value with the first non-missing value immediately above (previous) this value.
Omit Row from Calculations
Ignore the row in all calculations. Analyze the data as if the row was not on the database.

Forecasting Options
Number of Forecasts
This option specifies the number of forecasts to be generated.

Seasonal Model Options

Seasonal Adjustment
Select either the Additive or Multiplicative adjustment scheme.

Seasonality Options
Seasons
Specify the number of seasons per year in the series. Use ‘4’ for quarterly data or ‘12’ for monthly data.
First Season
Specify the first season of the series. This value is used to format the reports and plots. For example, if you have
monthly data beginning with March, you would enter a ‘3’ here.
First Year
Specify the first year of the series. This value is used to format the reports and plots.
Smoothing Constant Search Options
Search Method
This option specifies whether a search is conducted for the best values of the smoothing constants and what the
criterion for the search will be.

• Specified Value
No search is conducted. The values of the smoothing constants given in the next options are used.

• Search on MSE
A search is conducted to find the values of the smoothing constants that minimize MSE.

• Search on MAE
A search is conducted to find the values of the smoothing constants that minimize MAE.

• Search on MAPE
A search is conducted to find the values of the smoothing constants that minimize MAPE.

467-4
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com
Exponential Smoothing – Trend & Seasonal

Smoothing Constant Search Options – Pre-

Specified Smoothing Constants
Alpha Smoothing Constant
When the Search Method is set to Specified Value, this option specifies the value of alpha. Alpha is the
smoothing constant for the level of the series. The limits of this value are zero and one. Usually, a value between
0.1 and 0.3 are used. As the value gets closer to one, more and more weight is given to recent observations.
Beta Smoothing Constant
When the Search Method is set to Specified Value, this option specifies the value of beta. Beta is the smoothing
constant for the trend. The limits of this value are zero and one. Usually, a value between 0.1 and 0.3 are used. As
the value gets closer to one, more and more weight is given to recent observations.
Gamma Smoothing Constant
When the Search Method is set to Specified Value, this option specifies the value of gamma. Gamma is the
smoothing constant for the seasonal factors. The limits of this value are zero and one. Usually, a value between
0.1 and 0.3 are used. As the value gets closer to one, more and more weight is given to recent observations.

Reports Tab
The following options control which reports are displayed.

Select Reports
Summary Report
This option specifies whether the indicated report is displayed.
Forecast Report
This option specifies which parts of the series are listed on the numeric reports: the original data and forecasts,
just the forecasts, or neither.

Report Options
Precision
Specify the precision of numbers in the report. Single precision will display seven-place accuracy, while the
double precision will display thirteen-place accuracy. Note that all reports are formatted for single precision only.
Variable Names
Specify whether to use variable names or (the longer) variable labels in report headings.
Page Title
Specify a title to be shown at the top of the reports.

Plots Tab
This section controls the forecast and residual plots.

467-5
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com
Exponential Smoothing – Trend & Seasonal

Select Plots
Forecast Plot - Residual Plot
Each of these options specifies whether the indicated plot is displayed. Click the plot format button to change the
plot settings.

Horizontal Axis Variable

Horizontal Variable
This option controls the spacing on the horizontal axis when missing or filtered values occur.
Your choices are
Actual Row Number
Use the actual row number of each row from the dataset along the horizontal axis.
Constructed Date
Construct a date value from the sequence (relative row) number and the Seasonality Options settings. Any
missing or filtered values are skipped when forming the sequence number.

Storage Tab
The forecasts and residuals may be stored on the current dataset for further analysis. These options let you
designate which statistics (if any) should be stored by designating which columns should receive the statistics.
Note that existing data is replaced. Be careful that you do not specify columns that contain important data.

Data Storage Columns

Forecasts
The forecasts are stored in this column.
Residuals
The residuals are stored in this column.

467-6
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com
Exponential Smoothing – Trend & Seasonal

Example 1 – Trend & Seasonal Exponential Smoothing

This section presents an example of how to generate forecasts of a series using Winters multiplicative seasonal
model. The data in the Sales dataset will be used. We will forecast the values of the Sales variable for the next
twelve months.
You may follow along here by making the appropriate entries or load the completed template Example 1 by
clicking on Open Example Template from the File menu of the Exponential Smoothing – Trend / Seasonal
window.

1 Open the Sales dataset.

• From the File menu of the NCSS Data window, select Open Example Data.
• Click on the file Sales.NCSS.
• Click Open.

2 Open the Exponential Smoothing – Trend / Seasonal window.

• Using the Analysis menu or the Procedure Navigator, find and select the Exponential Smoothing -
Trend/Seasonal procedure.
• On the menus, select File, then New Template. This will fill the procedure with the default template.

3 Specify the variables.

• On the Exponential Smoothing – Trend / Seasonal window, select the Variables tab.
• Double-click in the Time Series Variable box. This will bring up the variable selection window.
• Select Sales from the list of variables and then click Ok.
• Enter 1970 in the First Year box.

4 Specify the reports.

• On the Exponential Smoothing – Trend / Seasonal window, select the Reports tab.
• Select Data and Forecasts in the Forecast Report list box.

5 Run the procedure.

• From the Run menu, select Run Procedure. Alternatively, just click the green Run button.

467-7
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com
Exponential Smoothing – Trend & Seasonal

Forecast Summary Section

Forecast Summary Section

Variable Sales
Number of Rows 144
Missing Values None
Mean 174.2847
Pseudo R-Squared 0.980145
Mean Square Error 16.10279
Mean |Error| 3.114085
Mean |Percent Error| 1.786407

Forecast Method Winter's with multiplicative seasonal adjustment.

Search Iterations 532
Search Criterion Mean Square Error
Alpha 0.3496813
Beta 5.572607E-05
Gamma 6.69852E-11
Intercept (A) 138.3665
Slope (B) 0.5871831
Season 1 Factor 0.9028255
Season 2 Factor 0.8556558
Season 3 Factor 0.9714928
Season 4 Factor 0.997539
Season 5 Factor 1.028944
Season 6 Factor 1.028335
Season 7 Factor 0.9977484
Season 8 Factor 1.005503
Season 9 Factor 0.9752801
Season 10 Factor 1.023642
Season 11 Factor 1.007328
Season 12 Factor 1.205706

This report summarizes the forecast equation.

Variable
The name of the variable for which the forecasts are generated.
Number of Rows
The number of rows that were in the series. This is provided to allow you to double-check that the correct series
was used.
Missing Values
If missing values were found, this option lists the method used to estimate them.
Mean
The mean of the variable across all time periods.
Pseudo R-Squared
This value generates a statistic that acts like the R-Squared value in multiple regression. A value near zero
indicates a poorly fitting model, while a value near one indicates a well fitting model. The statistic is calculated as
follows:
 SSE 
R 2 = 100 1 − 
 SST 
where SSE is the sum of square residuals and SST is the total sum of squares after correcting for the mean.

467-8
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com
Exponential Smoothing – Trend & Seasonal

Mean Square Error

The average squared residual (MSE) is a measure of how closely the forecasts track the actual data. The statistic is
popular because it shows up in analysis of variance tables. However, because of the squaring, it tends to
exaggerate the influence of outliers (points that do not follow the regular pattern).
Mean |Error|
The average absolute residual (MAE) is a measure of how closely the forecasts track the actual data without the
squaring.
Mean |Percent Error|
The average percent absolute residual (MAPE) is a measure of how closely the forecasts track the actual data put
on a percentage basis.
Forecast Method
This line shows which of the two possible seasonal adjustment algorithms was selected.
Search Iterations
This line shows how many iterations were needed to find the best values for the smoothing constants.
Search Criterion
If a search was made to find the best values of the smoothing constants, this row gives the criterion used during
the search.
Alpha
The value of the smoothing constant alpha that was used to generate the forecasts.
Beta
The value of the smoothing constant beta that was used to generate the forecasts.
Gamma
The value of the smoothing constant gamma that was used to generate the forecasts.
Intercept (A)
The value of the y-intercept for time period one!
Slope (B)
The value of the slope.
Season (1-12) Factor
The values of the multiplicative seasonal factors.

467-9
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com
Exponential Smoothing – Trend & Seasonal

Forecast and Residuals Plots

Forecast Plot
The forecast plot lets you analyze how closely the forecasts track the data. The plot also shows the forecasts at the
end of the data series.
Residual Plot
This plot lets you analyze the residuals themselves. You are looking for patterns, outliers, or any other
information that may help you improve the forecasting model. The first thing to compare is the scale of the
Residual Plot versus the scale of the Forecast Plot. If your forecasting algorithm is working well, the vertical scale
of the Residual Plot will be much less than the scale of the Forecast Plot.

467-10
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com
Exponential Smoothing – Trend & Seasonal

Forecasts Section
Forecasts Section

Row Forecast Actual

No. Date Sales Sales Residuals
1 1970 1 124.4976 129 4.502408
2 1970 2 119.9876 122 2.012394
3 1970 3 137.6008 137 -0.600761
4 1970 4 141.66 141 -0.6600129
5 1970 5 146.486 145 -1.486042
6 1970 6 146.4839 144 -2.483881
7 1970 7 141.8699 143 1.130069
8 1970 8 143.9612 140 -3.961171
9 1970 9 138.8632 140 1.136814
10 1970 10 146.7673 148 1.232683
11 1970 11 145.4439 138 -7.443867
12 1970 12 171.6791 163 -8.679084
. . . . .
. . . . .
. . . . .
140 1981 8 220.3259 218 -2.325934
141 1981 9 213.4872 213 -0.4872138
142 1981 10 224.4956 226 1.504373
143 1981 11 222.0268 217 -5.026803
144 1981 12 264.3555 277 12.64447
145 1982 1 201.7888
146 1982 2 191.7484
147 1982 3 218.2773
148 1982 4 224.7152
149 1982 5 232.394
150 1982 6 232.8604
151 1982 7 226.5199
152 1982 8 228.8708
153 1982 9 222.5643
154 1982 10 234.2018
155 1982 11 231.0607
156 1982 12 277.2727

This section shows the values of the forecasts, the dates, the actual values, and the residuals.

467-11
© NCSS, LLC. All Rights Reserved.