27/2/2020 Forecasting Option

Forecasting Option
Contenido
1. Time Series Modeling and Forecasting
2. Bulk Forecasting with the Expert Modeler
2.1. Examining Your Data
2.2. Running the Analysis
2.3. Model Summary Charts
2.4. Model Predictions
2.5. Summary
3. Bulk Reforecasting by Applying Saved Models
3.1. Running the Analysis
3.2. Model Fit Statistics
3.3. Model Predictions
3.4. Summary
4. Using the Expert Modeler to Determine Significant Predictors
4.1. Plotting Your Data
4.2. Running the Analysis
4.3. Series Plot
4.4. Model Description Table
4.5. Model Statistics Table
4.6. ARIMA Model Parameters Table
4.7. Summary
5. Experimenting with Predictors by Applying Saved Models
5.1. Extending the Predictor Series
5.2. Modifying Predictor Values in the Forecast Period
5.3. Running the Analysis
6. Seasonal Decomposition
6.1. Seasonal Decomposition
6.2. Models
6.3. Removing Seasonality from Sales Data
6.3.1. Determining and Setting the Periodicity
6.3.2. Running the Analysis
6.3.3. Understanding the Output
6.3.4. Summary
6.4. Related Procedures
7. Spectral Plots
7.1. Spectral Plots
7.2. Using Spectral Plots to Verify Expectations about Periodicity
7.2.1. Running the Analysis
7.2.2. Understanding the Periodogram and Spectral Density
7.2.3. Summary
7.3. Related Procedures
8. Uncovering causal relationships in business metrics
8.1. Running the analysis
8.2. Overall Model Quality Chart
8.3. Overall Model System
8.4. Impact Diagrams
8.5. Determining root causes of outliers
8.6. Running scenarios
8.7. Forecasting with updated data

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 1/83
27/2/2020 Forecasting Option

1. Time Series Modeling and Forecasting

The Forecasting add-on module provides the following procedures for accomplishing the tasks of creating models
and producing forecasts:

The Time Series Modeler procedure estimates exponential smoothing, univariate Autoregressive Integrated
Moving Average (ARIMA), and multivariate ARIMA (or transfer function models) models for time series,
and produces forecasts. The procedure includes an Expert Modeler that attempts to automatically identify
and estimate the best-fitting ARIMA or exponential smoothing model for one or more dependent variable
series, thus eliminating the need to identify an appropriate model through trial and error. Alternatively, you
can specify a custom ARIMA or exponential smoothing model.

Learn how to use the Expert Modeler

The Apply Time Series Models procedure applies existing time series models--created by the Time Series
Modeler--to the active dataset. You can use this procedure to obtain forecasts for series for which new or
revised data are available, without rebuilding your models.

Learn how to apply saved models

The Temporal Causal Models procedure builds autoregressive time series models for each target and
automatically determines the best inputs that have a causal relationship with the target. The procedure
produces interactive output that you can use to explore the causal relationships. The procedure can also
generate forecasts, detect outliers, and determine the series that most likely causes an outlier.

Learn how to build and analyze temporal causal models

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 2/83
27/2/2020 Forecasting Option

2. Bulk Forecasting with the Expert Modeler

An analyst for a national broadband provider is required to produce forecasts of user subscriptions in order to
predict utilization of bandwidth. Forecasts are needed for each of the 85 local markets that make up the national
subscriber base. Monthly historical data is collected in broadband_1.sav. See the topic Sample Files for more
information.

In this example, you will use the Expert Modeler to produce forecasts for the next three months for each of the 85
local markets, saving the generated models to an external XML file. Once you are finished, you might want to
work through the next example, in, , Bulk Reforecasting by Applying Saved Models, which applies the saved
models to an updated dataset in order to extend the forecasts by another three months without having to rebuild
the models.

Next

Examining Your Data

Running the Analysis

Model Summary Charts

Model Predictions

Summary

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 3/83
27/2/2020 Forecasting Option

2.1. Examining Your Data

It is always a good idea to have a feel for the nature of your data before building a model. Does the data exhibit
seasonal variations? Although the Expert Modeler will automatically find the best seasonal or non-seasonal model
for each series, you can often obtain faster results by limiting the search to non-seasonal models when seasonality
is not present in your data. Without examining the data for each of the 85 local markets, we can get a rough
picture by plotting the total number of subscribers over all markets.

1. From the menus choose:

Analyze > Forecasting > Sequence Charts...

Figure 1. Sequence Charts dialog box

2. Select Total Number of Subscribers and move it into the Variables list.
3. Select Date and move it into the Time Axis Labels box.
4. Click OK.

Figure 2. Total number of broadband subscribers across all markets

The series exhibits a very smooth upward trend with no hint of seasonal variations. There might be individual
series with seasonality, but it appears that seasonality is not a prominent feature of the data in general. Of course
you should inspect each of the series before ruling out seasonal models. You can then separate out series
exhibiting seasonality and model them separately. In the present case, inspection of the 85 series would show that
none exhibit seasonality.
127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 4/83
27/2/2020 Forecasting Option

Next

Parent topic: Bulk Forecasting with the Expert Modeler

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 5/83
27/2/2020 Forecasting Option

2.2. Running the Analysis

To use the Expert Modeler:

1. From the menus choose:

Analyze > Forecasting > Create Traditional Models...

Figure 1. Time Series Modeler dialog box

2. Select Subscribers for Market 1 through Subscribers for Market 85 for dependent variables.
3. Verify that Expert Modeler is selected in the Method drop-down list. The Expert Modeler will
automatically find the best-fitting model for each of the dependent variable series.

The set of cases used to estimate the model is referred to as the estimation period. By default, it includes
all of the cases in the active dataset. You can set the estimation period by selecting Based on time or case
range in the Select Cases dialog box. For this example, we will stick with the default.

Notice also that the default forecast period starts after the end of the estimation period and goes through to
the last case in the active dataset. If you are forecasting beyond the last case, you will need to extend the
forecast period. This is done from the Options tab as you will see later on in this example.

4. Click Criteria.

Figure 2. Expert Modeler Criteria dialog box, Model tab

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 6/83
27/2/2020 Forecasting Option

5. Deselect Expert Modeler considers seasonal models in the Model Type group.

Although the data is monthly and the current periodicity is 12, we have seen that the data does not exhibit
any seasonality, so there is no need to consider seasonal models. This reduces the space of models searched
by the Expert Modeler and can significantly reduce computing time.

6. Click Continue.
7. Click the Options tab on the Time Series Modeler dialog box.

Figure 3. Time Series Modeler, Options tab

8. Select First case after end of estimation period through a specified date in the Forecast Period group.
127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 7/83
27/2/2020 Forecasting Option

9. In the Date grid, enter 2004 for the year and 3 for the month.

The dataset contains data from January 1999 through December 2003. With the current settings, the
forecast period will be January 2004 through March 2004.

10. Click the Save tab.

Figure 4. Time Series Modeler, Save tab

11. Select (check) the entry for Predicted Values in the Save column, and leave the default value Predicted as
the Variable Name Prefix.

The model predictions are saved as new variables in the active dataset, using the prefix Predicted for the
variable names. You can also save the specifications for each of the models to an external XML file. This
will allow you to reuse the models to extend your forecasts as new data becomes available.

12. Click the Browse button on the Save tab.

This will take you to a standard dialog box for saving a file.

13. Navigate to the folder where you would like to save the XML model file, enter a filename, and click Save.
14. Click the Statistics tab.

Figure 5. Time Series Modeler, Statistics tab

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 8/83
27/2/2020 Forecasting Option

15. Select Display forecasts.

This option produces a table of forecasted values for each dependent variable series and provides another
option--other than saving the predictions as new variables--for obtaining these values.

The default selection of Goodness of fit (in the Statistics for Comparing Models group) produces a table
with fit statistics—such as R-squared, mean absolute percentage error, and normalized BIC—calculated
across all of the models. It provides a concise summary of how well the models fit the data.

16. Click the Plots tab.

Figure 6. Time Series Modeler, Plots tab

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 9/83
27/2/2020 Forecasting Option

17. Deselect Series in the Plots for Individual Models group.

This suppresses the generation of series plots for each of the models. In this example, we are more
interested in saving the forecasts as new variables than generating plots of the forecasts.

The Plots for Comparing Models group provides several plots (in the form of histograms) of fit statistics
calculated across all models.

18. Select Mean absolute percentage error and Maximum absolute percentage error in the Plots for
Comparing Models group.

Absolute percentage error is a measure of how much a dependent series varies from its model-predicted
level. By examining the mean and maximum across all models, you can get an indication of the uncertainty
in your predictions. And looking at summary plots of percentage errors, rather than absolute errors, is
advisable since the dependent series represent subscriber numbers for markets of varying sizes.

19. Click OK in the Time Series Modeler dialog box.

Next

Parent topic: Bulk Forecasting with the Expert Modeler

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 10/83
27/2/2020 Forecasting Option

2.3. Model Summary Charts

Figure 1. Histogram of mean absolute percentage error

This histogram displays the mean absolute percentage error (MAPE) across all models. It shows that all models
display a mean uncertainty of roughly 1%.

Figure 2. Histogram of maximum absolute percentage error

This histogram displays the maximum absolute percentage error (MaxAPE) across all models and is useful for
imagining a worst-case scenario for your forecasts. It shows that the largest percentage error for each model falls
in the range of 1 to 5%. Do these values represent an acceptable amount of uncertainty? This is a situation in
which your business sense comes into play because acceptable risk will change from problem to problem.

Next

Parent topic: Bulk Forecasting with the Expert Modeler

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 11/83
27/2/2020 Forecasting Option

2.4. Model Predictions

Figure 1. New variables containing model predictions

The Data Editor shows the new variables containing the model predictions. Although only two are shown here,
there are 85 new variables, one for each of the 85 dependent series. The variable names consist of the default
prefix Predicted, followed by the name of the associated dependent variable (for example, Market_1), followed
by a model identifier (for example, Model_1).

Three new cases, containing the forecasts for January 2004 through March 2004, have been added to the dataset,
along with automatically generated date labels. Each of the new variables contains the model predictions for the
estimation period (January 1999 through December 2003), allowing you to see how well the model fits the known
values.

Figure 2. Forecast table

You also chose to create a table with the forecasted values. The table consists of the predicted values in the
forecast period but—unlike the new variables containing the model predictions—does not include predicted
values in the estimation period. The results are organized by model and identified by the model name, which
consists of the name (or label) of the associated dependent variable followed by a model identifier--just like the
names of the new variables containing the model predictions. The table also includes the upper confidence limits
(UCL) and lower confidence limits (LCL) for the forecasted values (95% by default).

You have now seen two approaches for obtaining the forecasted values: saving the forecasts as new variables in
the active dataset and creating a forecast table. With either approach, you will have a number of options available
for exporting your forecasts (for example, into an Excel spreadsheet).

Next

Parent topic: Bulk Forecasting with the Expert Modeler

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 12/83
27/2/2020 Forecasting Option

2.5. Summary
You have learned how to use the Expert Modeler to produce forecasts for multiple series, and you have saved the
resulting models to an external XML file. In the next example, you will learn how to extend your forecasts as new
data becomes available—without having to rebuild your models—by using the Apply Time Series Models
procedure.

Parent topic: Bulk Forecasting with the Expert Modeler

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 13/83
27/2/2020 Forecasting Option

3. Bulk Reforecasting by Applying Saved Models

You have used the Time Series Modeler to create models for your time series data and to produce initial forecasts
based on available data. You plan to reuse these models to extend your forecasts as more current data becomes
available, so you saved the models to an external file. You are now ready to apply the saved models.

This example is a natural extension of the previous one , Bulk Forecasting with the Expert Modeler, but can also
be used independently. In this scenario, you are an analyst for a national broadband provider who is required to
produce monthly forecasts of user subscriptions for each of 85 local markets. You have already used the Expert
Modeler to create models and to forecast three months into the future. Your data warehouse has been refreshed
with actual data for the original forecast period, so you would like to use that data to extend the forecast horizon
by another three months.

The updated monthly historical data is collected in broadband_2.sav, and the saved models are in
broadband_models.xml. See the topic Sample Files for more information. Of course, if you worked through the
previous example and saved your own model file, you can use that one instead of broadband_models.xml.

Next

Running the Analysis

Model Fit Statistics

Model Predictions

Summary

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 14/83
27/2/2020 Forecasting Option

3.1. Running the Analysis

To apply models:

1. From the menus choose:

Analyze > Forecasting > Apply Traditional Models...

Figure 1. Apply Time Series Models dialog box

2. Click Browse, then navigate to and select broadband_models.xml (or choose your own model file saved
from the previous example). See the topic Sample Files for more information.
3. Select Reestimate from data.

To incorporate new values of your time series into forecasts, the Apply Time Series Models procedure will
have to reestimate the model parameters. The structure of the models remains the same though, so the
computing time to reestimate is much quicker than the original computing time to build the models.

The set of cases used for reestimation needs to include the new data. This will be assured if you use the
default estimation period of First Case to Last Case. If you ever need to set the estimation period to
something other than the default, you can do so by selecting Based on time or case range in the Select
Cases dialog box.

4. Select First case after end of estimation period through a specified date in the Forecast Period group.
5. In the Date grid, enter 2004 for the year and 6 for the month.

The dataset contains data from January 1999 through March 2004. With the current settings, the forecast
period will be April 2004 through June 2004.

6. Click the Save tab.

Figure 2. Apply Time Series Models, Save tab

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 15/83
27/2/2020 Forecasting Option

7. Select (check) the entry for Predicted Values in the Save column and leave the default value Predicted as
the Variable Name Prefix.

The model predictions will be saved as new variables in the active dataset, using the prefix Predicted for
the variable names.

8. Click the Plots tab.

Figure 3. Apply Time Series Models, Plots tab

9. Deselect Series in the Plots for Individual Models group.

This suppresses the generation of series plots for each of the models. In this example, we are more
interested in saving the forecasts as new variables than generating plots of the forecasts.

10. Click OK in the Apply Time Series Models dialog box.

Next

Parent topic: Bulk Reforecasting by Applying Saved Models

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 16/83
27/2/2020 Forecasting Option

3.2. Model Fit Statistics

Figure 1. Model Fit table

The Model Fit table provides fit statistics calculated across all of the models. It provides a concise summary of
how well the models, with reestimated parameters, fit the data. For each statistic, the table provides the mean,
standard error (SE), minimum, and maximum value across all models. It also contains percentile values that
provide information on the distribution of the statistic across models. For each percentile, that percentage of
models have a value of the fit statistic below the stated value. For instance, 95% of the models have a value of
MaxAPE (maximum absolute percentage error) that is less than 3.676.

While a number of statistics are reported, we will focus on two: MAPE (mean absolute percentage error) and
MaxAPE (maximum absolute percentage error). Absolute percentage error is a measure of how much a dependent
series varies from its model-predicted level and provides an indication of the uncertainty in your predictions. The
mean absolute percentage error varies from a minimum of 0.669% to a maximum of 1.026% across all models.
The maximum absolute percentage error varies from 1.742% to 4.373% across all models. So the mean
uncertainty in each model's predictions is about 1% and the maximum uncertainty is around 2.5% (the mean value
of MaxAPE), with a worst case scenario of about 4%. Whether these values represent an acceptable amount of
uncertainty depends on the degree of risk you are willing to accept.

Next

Parent topic: Bulk Reforecasting by Applying Saved Models

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 17/83
27/2/2020 Forecasting Option

3.3. Model Predictions

Figure 1. New variables containing model predictions

The Data Editor shows the new variables containing the model predictions. Although only two are shown here,
there are 85 new variables, one for each of the 85 dependent series. The variable names consist of the default
prefix Predicted, followed by the name of the associated dependent variable (for example, Market_1), followed
by a model identifier (for example, Model_1).

Three new cases, containing the forecasts for April 2004 through June 2004, have been added to the dataset, along
with automatically generated date labels.

Next

Parent topic: Bulk Reforecasting by Applying Saved Models

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 18/83
27/2/2020 Forecasting Option

3.4. Summary
You have learned how to apply saved models to extend your previous forecasts when more current data becomes
available. And you have done this without rebuilding your models. Of course, if there is reason to think that a
model has changed, then you should rebuild it using the Time Series Modeler procedure.

Parent topic: Bulk Reforecasting by Applying Saved Models

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 19/83
27/2/2020 Forecasting Option

4. Using the Expert Modeler to Determine Significant Predictors

A catalog company, interested in developing a forecasting model, has collected data on monthly sales of men's
clothing along with several series that might be used to explain some of the variation in sales. Possible predictors
include the number of catalogs mailed, the number of pages in the catalog, the number of phone lines open for
ordering, the amount spent on print advertising, and the number of customer service representatives. Are any of
these predictors useful for forecasting?

In this example, you will use the Expert Modeler with all of the candidate predictors to find the best model. Since
the Expert Modeler only selects those predictors that have a statistically significant relationship with the
dependent series, you will know which predictors are useful, and you will have a model for forecasting with them.
Once you are finished, you might want to work through the next example, in, , Experimenting with Predictors by
Applying Saved Models, which investigates the effect on sales of different predictor scenarios using the model
built in this example.

The data for the current example is collected in catalog_seasfac.sav. See the topic Sample Files for more
information.

Next

Plotting Your Data

Running the Analysis

Series Plot

Model Description Table

Model Statistics Table

ARIMA Model Parameters Table

Summary

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 20/83
27/2/2020 Forecasting Option

4.1. Plotting Your Data

It is always a good idea to plot your data, especially if you are only working with one series:

1. From the menus choose:

Analyze > Forecasting > Sequence Charts...

Figure 1. Sequence Charts dialog box

2. Select Sales of Men's Clothing and move it into the Variables list.
3. Select Date and move it into the Time Axis Labels box.
4. Click OK.

Figure 2. Sales of men's clothing (in U.S. dollars)

The series exhibits numerous peaks, many of which appear to be equally spaced, as well as a clear upward trend.
The equally spaced peaks suggests the presence of a periodic component to the time series. Given the seasonal
nature of sales, with highs typically occurring during the holiday season, you should not be surprised to find an
annual seasonal component to the data.

There are also peaks that do not appear to be part of the seasonal pattern and which represent significant
deviations from the neighboring data points. These points may be outliers, which can and should be addressed by
the Expert Modeler.

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 21/83
27/2/2020 Forecasting Option

Next

Parent topic: Using the Expert Modeler to Determine Significant Predictors

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 22/83
27/2/2020 Forecasting Option

4.2. Running the Analysis

To use the Expert Modeler:

1. From the menus choose:

Analyze > Forecasting > Create Traditional Models...

Figure 1. Time Series Modeler dialog box

2. Select Sales of Men's Clothing for the dependent variable.

3. Select Number of Catalogs Mailed through Number of Customer Service Representatives for the
independent variables.
4. Verify that Expert Modeler is selected in the Method drop-down list. The Expert Modeler will
automatically find the best-fitting seasonal or non-seasonal model for the dependent variable series.
5. Click Criteria and then click the Outliers tab.

Figure 2. Expert Modeler Criteria dialog box, Outliers tab

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 23/83
27/2/2020 Forecasting Option

6. Select Detect outliers automatically and leave the default selections for the types of outliers to detect.

Our visual inspection of the data suggested that there may be outliers. With the current choices, the Expert
Modeler will search for the most common outlier types and incorporate any outliers into the final model.
Outlier detection can add significantly to the computing time needed by the Expert Modeler, so it is a
feature that should be used with some discretion, particularly when modeling many series at once. By
default, outliers are not detected.

7. Click Continue.
8. Click the Save tab on the Time Series Modeler dialog box.

Figure 3. Time Series Modeler, Save tab

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 24/83
27/2/2020 Forecasting Option

You will want to save the estimated model to an external XML file so that you can experiment with
different values of the predictors—using the Apply Time Series Models procedure—without having to
rebuild the model.

9. Click the Browse button on the Save tab.

This will take you to a standard dialog box for saving a file.

10. Navigate to the folder where you would like to save the XML model file, enter a filename, and click Save.
11. Click the Statistics tab.

Figure 4. Time Series Modeler, Statistics tab

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 25/83
27/2/2020 Forecasting Option

12. Select Parameter estimates.

This option produces a table displaying all of the parameters, including the significant predictors, for the
model chosen by the Expert Modeler.

13. Click the Plots tab.

Figure 5. Time Series Modeler, Plots tab

14. Deselect Forecasts.

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 26/83
27/2/2020 Forecasting Option

In the current example, we are only interested in determining the significant predictors and building a
model. We will not be doing any forecasting.

15. Select Fit values.

This option displays the predicted values in the period used to estimate the model. This period is referred to
as the estimation period, and it includes all cases in the active dataset for this example. These values
provide an indication of how well the model fits the observed values, so they are referred to as fit values.
The resulting plot will consist of both the observed values and the fit values.

16. Click OK in the Time Series Modeler dialog box.

Next

Parent topic: Using the Expert Modeler to Determine Significant Predictors

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 27/83
27/2/2020 Forecasting Option

4.3. Series Plot

Figure 1. Predicted and observed values

The predicted values show good agreement with the observed values, indicating that the model has satisfactory
predictive ability. Notice how well the model predicts the seasonal peaks. And it does a good job of capturing the
upward trend of the data.

Next

Parent topic: Using the Expert Modeler to Determine Significant Predictors

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 28/83
27/2/2020 Forecasting Option

4.4. Model Description Table

Figure 1. Model Description table

The model description table contains an entry for each estimated model and includes both a model identifier and
the model type. The model identifier consists of the name (or label) of the associated dependent variable and a
system-assigned name. In the current example, the dependent variable is Sales of Men's Clothing and the system-
assigned name is Model_1.

The Time Series Modeler supports both exponential smoothing and ARIMA models. Exponential smoothing
model types are listed by their commonly used names such as Holt and Winters' Additive. ARIMA model types
are listed using the standard notation of ARIMA(p,d,q)(P,D,Q), where p is the order of autoregression, d is the
order of differencing (or integration), and q is the order of moving-average, and (P,D,Q) are their seasonal
counterparts.

The Expert Modeler has determined that sales of men's clothing is best described by a seasonal ARIMA model
with one order of differencing. The seasonal nature of the model accounts for the seasonal peaks that we saw in
the series plot, and the single order of differencing reflects the upward trend that was evident in the data.

Next

Parent topic: Using the Expert Modeler to Determine Significant Predictors

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 29/83
27/2/2020 Forecasting Option

4.5. Model Statistics Table

Figure 1. Model Statistics table

The model statistics table provides summary information and goodness-of-fit statistics for each estimated model.
Results for each model are labeled with the model identifier provided in the model description table. First, notice
that the model contains two predictors out of the five candidate predictors that you originally specified. So it
appears that the Expert Modeler has identified two independent variables that may prove useful for forecasting.

Although the Time Series Modeler offers a number of different goodness-of-fit statistics, we opted only for the
stationary R-squared value. This statistic provides an estimate of the proportion of the total variation in the series
that is explained by the model and is preferable to ordinary R-squared when there is a trend or seasonal pattern, as
is the case here. Larger values of stationary R-squared (up to a maximum value of 1) indicate better fit. A value of
0.948 means that the model does an excellent job of explaining the observed variation in the series.

The Ljung-Box statistic, also known as the modified Box-Pierce statistic, provides an indication of whether the
model is correctly specified. A significance value less than 0.05 implies that there is structure in the observed
series which is not accounted for by the model. The value of 0.984 shown here is not significant, so we can be
confident that the model is correctly specified.

The Expert Modeler detected nine points that were considered to be outliers. Each of these points has been
modeled appropriately, so there is no need for you to remove them from the series.

Next

Parent topic: Using the Expert Modeler to Determine Significant Predictors

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 30/83
27/2/2020 Forecasting Option

4.6. ARIMA Model Parameters Table

Figure 1. ARIMA Model Parameters table

The ARIMA model parameters table displays values for all of the parameters in the model, with an entry for each
estimated model labeled by the model identifier. For our purposes, it will list all of the variables in the model,
including the dependent variable and any independent variables that the Expert Modeler determined were
significant. We already know from the model statistics table that there are two significant predictors. The model
parameters table shows us that they are the Number of Catalogs Mailed and the Number of Phone Lines Open for
Ordering.

Next

Parent topic: Using the Expert Modeler to Determine Significant Predictors

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 31/83
27/2/2020 Forecasting Option

4.7. Summary
You have learned how to use the Expert Modeler to build a model and identify significant predictors, and you
have saved the resulting model to an external file. You are now in a position to use the Apply Time Series Models
procedure to experiment with alternative scenarios for the predictor series and see how the alternatives affect the
sales forecasts.

Parent topic: Using the Expert Modeler to Determine Significant Predictors

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 32/83
27/2/2020 Forecasting Option

5. Experimenting with Predictors by Applying Saved Models

You've used the Time Series Modeler to create a model for your data and to identify which predictors may prove
useful for forecasting. The predictors represent factors that are within your control, so you'd like to experiment
with their values in the forecast period to see how forecasts of the dependent variable are affected. This task is
easily accomplished with the Apply Time Series Models procedure, using the model file that is created with the
Time Series Modeler procedure.

This example is a natural extension of the previous example, in, , Using the Expert Modeler to Determine
Significant Predictors, but this example can also be used independently. The scenario involves a catalog company
that has collected data about monthly sales of men's clothing from January 1989 through December 1998, along
with several series that are thought to be potentially useful as predictors of future sales. The Expert Modeler has
determined that only two of the five candidate predictors are significant: the number of catalogs mailed and the
number of phone lines open for ordering.

When planning your sales strategy for the next year, you have limited resources to print catalogs and keep phone
lines open for ordering. Your budget for the first three months of 1999 allows for either 2000 additional catalogs
or 5 additional phone lines over your initial projections. Which choice will generate more sales revenue for this
three-month period?

The data for this example are collected in catalog_seasfac.sav, and catalog_model.xml contains the model of
monthly sales that is built with the Expert Modeler. See the topic Sample Files for more information. Of course, if
you worked through the previous example and saved your own model file, you can use that file instead of
catalog_model.xml.

Next

Extending the Predictor Series

Modifying Predictor Values in the Forecast Period

Running the Analysis

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 33/83
27/2/2020 Forecasting Option

5.1. Extending the Predictor Series

When you're creating forecasts for dependent series with predictors, each predictor series needs to be extended
through the forecast period. Unless you know precisely what the future values of the predictors will be, you'll
need to estimate them. You can then modify the estimates to test different predictor scenarios. The initial
projections are easily created by using the Expert Modeler.

1. From the menus choose:

Analyze > Forecasting > Create Traditional Models...

Figure 1. Time Series Modeler dialog box

2. Select Number of Catalogs Mailed and Number of Phone Lines Open for Ordering for the dependent
variables.
3. Click the Save tab.

Figure 2. Time Series Modeler, Save tab

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 34/83
27/2/2020 Forecasting Option

4. In the Save column, select (check) the entry for Predicted Values, and leave the default value Predicted for
the Variable Name Prefix.
5. Click the Options tab.

Figure 3. Time Series Modeler, Options tab

6. In the Forecast Period group, select First case after end of estimation period through a specified date.
7. In the Date grid, enter 1999 for the year and 3 for the month.

The data set contains data from January 1989 through December 1998, so with the current settings, the
forecast period will be January 1999 through March 1999.
127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 35/83
27/2/2020 Forecasting Option

8. Click OK.

Figure 4. New variables containing forecasts for predictor series

The Data Editor shows the new variables Predicted_mail_Model_1 and Predicted_phone_Model_2,
containing the model predicted values for the number of catalogs mailed and the number of phone lines. To
extend our predictor series, we only need the values for January 1999 through March 1999, which amounts
to cases 121 through 123.

9. Copy the values of these three cases from Predicted_mail_Model_1 and append them to the variable mail.
10. Repeat this process for Predicted_phone_Model_2, copying the last three cases and appending them to the
variable phone.

Figure 5. Predictor series extended through the forecast period

The predictors have now been extended through the forecast period.

Next

Parent topic: Experimenting with Predictors by Applying Saved Models

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 36/83
27/2/2020 Forecasting Option

5.2. Modifying Predictor Values in the Forecast Period

Testing the two scenarios of mailing more catalogs or providing more phone lines requires modifying the
estimates for the predictors mail or phone, respectively. Because we're only modifying the predictor values for
three cases (months), it would be easy to enter the new values directly into the appropriate cells of the Data
Editor. For instructional purposes, we'll use the Compute Variable dialog box. When you have more than a few
values to modify, you'll probably find the Compute Variable dialog box more convenient.

1. From the menus choose:

Transform > Compute Variable...

Figure 1. Compute Variable dialog box

2. Enter mail for the target variable.

3. In the Numeric Expression text box, enter mail + 2000.
4. Click If.

Figure 2. Compute Variable If Cases dialog box

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 37/83
27/2/2020 Forecasting Option

5. Select Include if case satisfies condition.

6. In the text box, enter $CASENUM > 120.

This will limit changes to the variable mail to the cases in the forecast period.

7. Click Continue.
8. Click OK in the Compute Variable dialog box, and click OK when asked whether you want to change the
existing variable.

This results in increasing the values for mail--the number of catalogs mailed--by 2000 for each of the three
months in the forecast period. You've now prepared the data to test the first scenario, and you are ready to run the
analysis.

Next

Parent topic: Experimenting with Predictors by Applying Saved Models

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 38/83
27/2/2020 Forecasting Option

5.3. Running the Analysis

1. From the menus choose:

Analyze > Forecasting > Apply Traditional Models...

Figure 1. Apply Time Series Models dialog box

2. Click Browse, then navigate to and select catalog_model.xml, or choose your own model file (saved from
the previous example). See the topic Sample Files for more information.
3. In the Forecast Period group, select First case after end of estimation period through a specified date.
4. In the Date grid, enter 1999 for the year and 3 for the month.
5. Click the Statistics tab.

Figure 2. Apply Time Series Models, Statistics tab

6. Select Display forecasts.

This results in a table of forecasted values for the dependent variable.

7. Click OK in the Apply Time Series Models dialog box.

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 39/83
27/2/2020 Forecasting Option

Figure 3. Forecast table

The forecast table contains the predicted values of the dependent series, taking into account the values of
the two predictors mail and phone in the forecast period. The table also includes the upper confidence limit
(UCL) and lower confidence limit (LCL) for the predictions.

You've produced the sales forecast for the scenario of mailing 2000 more catalogs each month. You'll now
want to prepare the data for the scenario of increasing the number of phone lines, which means resetting the
variable mail to the original values and increasing the variable phone by 5. You can reset mail by copying
the values of Predicted_mail_Model_1 in the forecast period and pasting them over the current values of
mail in the forecast period. And you can increase the number of phone lines--by 5 for each month in the
forecast period--either directly in the data editor or using the Compute Variable dialog box, like we did for
the number of catalogs.

To run the analysis, reopen the Apply Time Series Models dialog box as follows:

8. Click the Dialog Recall toolbar button.

9. Choose Apply Time Series Models.
10. Click OK in the Apply Time Series Models dialog box.

Figure 4. Forecast tables for the two scenarios

Displaying the forecast tables for both scenarios shows that, in each of the three forecasted months, increasing the
number of mailed catalogs is expected to generate approximately $1500 more in sales than increasing the number
of phone lines that are open for ordering. Based on the analysis, it seems wise to allocate resources to the mailing
of 2000 additional catalogs.

Parent topic: Experimenting with Predictors by Applying Saved Models

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 40/83
27/2/2020 Forecasting Option

6. Seasonal Decomposition
Seasonal Decomposition

Models

Removing Seasonality from Sales Data

Related Procedures

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 41/83
27/2/2020 Forecasting Option

6.1. Seasonal Decomposition

The Seasonal Decomposition procedure removes periodic fluctuations from time series, such as annual or
seasonal highs or lows. It is used primarily as a preliminary tool when attempting to analyze trend in such series.

Next

Parent topic: Seasonal Decomposition

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 42/83
27/2/2020 Forecasting Option

6.2. Models
Two different modeling approaches are available; multiplicative or additive.

Multiplicative. The seasonal component is a factor by which the seasonally adjusted series is multiplied to
yield the original series. In effect, seasonal components that are proportional to the overall level of the
series. Observations without seasonal variation have a seasonal component of 1.
Additive. The seasonal adjustments are added to the seasonally adjusted series to obtain the observed
values. This adjustment attempts to remove the seasonal effect from a series in order to look at other
characteristics of interest that may be "masked" by the seasonal component. In effect, seasonal components
that do not depend on the overall level of the series. Observations without seasonal variation have a
seasonal component of 0.

Next

Parent topic: Seasonal Decomposition

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 43/83
27/2/2020 Forecasting Option

6.3. Removing Seasonality from Sales Data

A catalog company is interested in modeling the upward trend of sales of its men's clothing line on a set of
predictor variables (such as the number of catalogs mailed and the number of phone lines open for ordering). To
this end, the company collected monthly sales of men's clothing for a 10-year period. This information is
collected in catalog.sav. See the topic Sample Files for more information.

To perform a trend analysis, you must remove any seasonal variations present in the data. This task is easily
accomplished with the Seasonal Decomposition procedure.

Next

Determining and Setting the Periodicity

Running the Analysis

Understanding the Output

Summary

Parent topic: Seasonal Decomposition

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 44/83
27/2/2020 Forecasting Option

6.3.1. Determining and Setting the Periodicity

The Seasonal Decomposition procedure requires the presence of a periodic date component in the active dataset—
for example, a yearly periodicity of 12 (months), a weekly periodicity of 7 (days), and so on. It's a good idea to
plot your time series first, because viewing a time series plot often leads to a reasonable guess about the
underlying periodicity.

To obtain a plot of men's clothing sales over time:

1. From the menus choose:

Analyze > Forecasting > Sequence Charts...

Figure 1. Sequence Charts dialog box

2. Select Sales of Men's Clothing and move it into the Variables list.
3. Select Date and move it into the Time Axis Labels list.
4. Click OK.

Figure 2. Sales of men's clothing (in U.S. dollars)

The series exhibits a number of peaks, but they do not appear to be equally spaced. This output suggests
that if the series has a periodic component, it also has fluctuations that are not periodic--the typical case for
real-time series. Aside from the small-scale fluctuations, the significant peaks appear to be separated by
more than a few months. Given the seasonal nature of sales, with typical highs during the December
127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 45/83
27/2/2020 Forecasting Option

holiday season, the time series probably has an annual periodicity. Also notice that the seasonal variations
appear to grow with the upward series trend, suggesting that the seasonal variations may be proportional to
the level of the series, which implies a multiplicative model rather than an additive model.

Examining the autocorrelations and partial autocorrelations of a time series provides a more quantitative
conclusion about the underlying periodicity.

5. From the menus choose:

Analyze > Forecasting > Autocorrelations...

Figure 3. Autocorrelations dialog box

6. Select Sales of Men's Clothing and move it into the Variables list.
7. Click OK.

Figure 4. Autocorrelation plot for men

The autocorrelation function shows a significant peak at a lag of 1 with a long exponential tail—a typical
pattern for time series. The significant peak at a lag of 12 suggests the presence of an annual seasonal
component in the data. Examination of the partial autocorrelation function will allow a more definitive
conclusion.

Figure 5. Partial autocorrelation plot for men

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 46/83
27/2/2020 Forecasting Option

The significant peak at a lag of 12 in the partial autocorrelation function confirms the presence of an annual
seasonal component in the data.

To set an annual periodicity:

8. From the menus choose:

Data > Define Dates...

Figure 6. Define Dates dialog box

9. Select Years, months in the Cases Are list.

10. Enter 1989 for the year and 1 for the month.
11. Click OK.

This sets the periodicity to 12 and creates a set of date variables that are designed to work with Forecasting
procedures.

Next

Parent topic: Removing Seasonality from Sales Data

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 47/83
27/2/2020 Forecasting Option

6.3.2. Running the Analysis

To run the Seasonal Decomposition procedure:

1. From the menus choose:

Analyze > Forecasting > Seasonal Decomposition...

Figure 1. Seasonal Decomposition dialog box

2. Right click anywhere in the source variable list and from the context menu select Display Variable Names.
3. Select men and move it into the Variables list.
4. Select Multiplicative in the Model Type group.
5. Click OK.

Next

Parent topic: Removing Seasonality from Sales Data

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 48/83
27/2/2020 Forecasting Option

6.3.3. Understanding the Output

The Seasonal Decomposition procedure creates four new variables for each of the original variables analyzed by
the procedure. By default, the new variables are added to the active data set. The new series have names
beginning with the following prefixes:

SAF. Seasonal adjustment factors, representing seasonal variation. For the multiplicative model, the value 1
represents the absence of seasonal variation; for the additive model, the value 0 represents the absence of seasonal
variation.

SAS. Seasonally adjusted series, representing the original series with seasonal variations removed. Working with
a seasonally adjusted series, for example, allows a trend component to be isolated and analyzed independent of
any seasonal component.

STC. Smoothed trend-cycle component, which is a smoothed version of the seasonally adjusted series that shows
both trend and cyclic components.

ERR. The residual component of the series for a particular observation.

For the present case, the seasonally adjusted series is the most appropriate, because it represents the original series
with the seasonal variations removed.

Figure 1. Sequence Charts dialog box

To plot the seasonally adjusted series:

1. Open the Sequence Charts dialog box.

2. Click Reset to clear any previous selections.
3. Right click anywhere in the source variable list, and from the context menu select Display Variable
Names.
4. Select SAS_1 and move it into the Variables list.
5. Click OK.

Figure 2. Seasonally adjusted series

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 49/83
27/2/2020 Forecasting Option

The seasonally adjusted series shows a clear upward trend. A number of peaks are evident, but they appear at
random intervals, showing no evidence of an annual pattern.

Next

Parent topic: Removing Seasonality from Sales Data

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 50/83
27/2/2020 Forecasting Option

6.3.4. Summary
Using the Seasonal Decomposition procedure, you have removed the seasonal component of a periodic time
series to produce a series that is more suitable for trend analysis. Examination of the autocorrelations and partial
autocorrelations of the time series was useful in determining the underlying periodicity—in this case, annual.

Parent topic: Removing Seasonality from Sales Data

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 51/83
27/2/2020 Forecasting Option

6.4. Related Procedures

The Seasonal Decomposition procedure is useful for removing a single seasonal component from a periodic time
series.

To perform a more in-depth analysis of the periodicity of a time series than is provided by the partial
correlation function, use the Spectral Plots procedure. To perform a more in-depth analysis of the
periodicity of a time series than is provided by the partial correlation function, use the Spectral Plots
procedure. For more information, see Using Spectral Plots to Verify Expectations about Periodicity.

Parent topic: Seasonal Decomposition

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 52/83
27/2/2020 Forecasting Option

7. Spectral Plots
Spectral Plots

Using Spectral Plots to Verify Expectations about Periodicity

Related Procedures

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 53/83
27/2/2020 Forecasting Option

7.1. Spectral Plots

The Spectral Plots procedure is used to identify periodic behavior in time series. For time series composed of
multiple underlying periodicities or when a considerable amount of random noise is present in the data, Spectral
Plots provides the clearest means of identifying periodic components.

Next

Parent topic: Spectral Plots

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 54/83
27/2/2020 Forecasting Option

7.2. Using Spectral Plots to Verify Expectations about Periodicity

Time series representing retail sales typically have an underlying annual periodicity, due to the usual peak in sales
during the holiday season. Producing sales projections means building a model of the time series, which means
identifying any periodic components. A plot of the time series may not always uncover the annual periodicity
because time series contain random fluctuations that often mask the underlying structure.

Monthly sales data for a catalog company are stored in catalog.sav. See the topic Sample Files for more
information. Before proceeding with sales projections, you want to confirm that the sales data exhibits an annual
periodicity. A plot of the time series shows many peaks with an irregular spacing, so any underlying periodicity is
not evident. Use the Spectral Plots procedure to identify any periodicity in the sales data.

Next

Running the Analysis

Understanding the Periodogram and Spectral Density

Summary

Parent topic: Spectral Plots

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 55/83
27/2/2020 Forecasting Option

7.2.1. Running the Analysis

To run the Spectral Plots procedure:

1. From the menus choose:

Analyze > Forecasting > Spectral Analysis...

Figure 1. Spectral Plots dialog box

2. Select Sales of Men's Clothing and move it into the Variables list.
3. Select Spectral density in the Plot group.
4. Click OK.

Next

Parent topic: Using Spectral Plots to Verify Expectations about Periodicity

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 56/83
27/2/2020 Forecasting Option

7.2.2. Understanding the Periodogram and Spectral Density

Figure 1. Periodogram

The plot of the periodogram shows a sequence of peaks that stand out from the background noise, with the lowest
frequency peak at a frequency of just less than 0.1. You suspect that the data contain an annual periodic
component, so consider the contribution that an annual component would make to the periodogram. Each of the
data points in the time series represents a month, so an annual periodicity corresponds to a period of 12 in the
current data set. Because period and frequency are reciprocals of each other, a period of 12 corresponds to a
frequency of 1/12 (or 0.083). So an annual component implies a peak in the periodogram at 0.083, which seems
consistent with the presence of the peak just below a frequency of 0.1.

Figure 2. Univariate statistics table

The univariate statistics table contains the data points that are used to plot the periodogram. Notice that, for
frequencies of less than 0.1, the largest value in the Periodogram column occurs at a frequency of 0.08333—
precisely what you expect to find if there is an annual periodic component. This information confirms the
identification of the lowest frequency peak with an annual periodic component. But what about the other peaks at
higher frequencies?

Figure 3. Spectral density

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 57/83
27/2/2020 Forecasting Option

The remaining peaks are best analyzed with the spectral density function, which is simply a smoothed version of
the periodogram. Smoothing provides a means of eliminating the background noise from a periodogram, allowing
the underlying structure to be more clearly isolated.

The spectral density consists of five distinct peaks that appear to be equally spaced. The lowest frequency peak
simply represents the smoothed version of the peak at 0.08333. To understand the significance of the four higher
frequency peaks, remember that the periodogram is calculated by modeling the time series as the sum of cosine
and sine functions. Periodic components that have the shape of a sine or cosine function (sinusoidal) show up in
the periodogram as single peaks. Periodic components that are not sinusoidal show up as a series of equally
spaced peaks of different heights, with the lowest frequency peak in the series occurring at the frequency of the
periodic component. So the four higher frequency peaks in the spectral density simply indicate that the annual
periodic component is not sinusoidal.

You have now accounted for all of the discernible structure in the spectral density plot and conclude that the data
contain a single periodic component with a period of 12 months.

Next

Parent topic: Using Spectral Plots to Verify Expectations about Periodicity

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 58/83
27/2/2020 Forecasting Option

7.2.3. Summary
Using the Spectral Plots procedure, you have confirmed the existence of an annual periodic component of a time
series, and you have verified that no other significant periodicities are present. The spectral density was seen to be
more useful than the periodogram for uncovering the underlying structure, because the spectral density smoothes
out the fluctuations that are caused by the nonperiodic component of the data.

Parent topic: Using Spectral Plots to Verify Expectations about Periodicity

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 59/83
27/2/2020 Forecasting Option

7.3. Related Procedures

The Spectral Plots procedure is useful for identifying the periodic components of a time series.

To remove a periodic component from a time series—for example, to perform a trend analysis—use the
Seasonal Decomposition procedure. To remove a periodic component from a time series—for instance, to
perform a trend analysis—use the Seasonal Decomposition procedure. See Removing Seasonality from
Sales Data for details.

Parent topic: Spectral Plots

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 60/83
27/2/2020 Forecasting Option

8. Uncovering causal relationships in business metrics

A business tracks various key performance indicators that describe the financial state of the business over time,
and they also track a number of metrics that they can control. They are interested in using temporal causal
modeling to uncover causal relationships between the controllable metrics and the key performance indicators.
They would also like to know about any causal relationships among the key performance indicators.

The data file tcm_kpi.sav contains weekly data on the key performance indicators and the controllable metrics.
Data for the key performance indicators is stored in fields with the prefix KPI. Data for the controllable metrics is
stored in fields with the prefix Lever. For more information, see the topic Sample Files.

Next

Running the analysis

Overall Model Quality Chart

Overall Model System

Impact Diagrams

Determining root causes of outliers

Running scenarios

Forecasting with updated data

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 61/83
27/2/2020 Forecasting Option

8.1. Running the analysis

1. Open the data file tcm_kpi.sav.
2. From the menus, choose:

Analyze > Forecasting > Create Temporal Causal Models...

Figure 1. Temporal Causal Modeling, introductory dialog box

3. Select date for the Date/Time field and click Continue.

The data for date are scanned to determine the time interval between successive observations.

Figure 2. Temporal Causal Modeling Time Series

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 62/83
27/2/2020 Forecasting Option

4. Select Use predefined roles.

In the sample data set tcm_kpi.sav, the fields Lever1 through Lever5 have the role of Input and KPI_1
through KPI_25 have the role of Both. When Use predefined roles is selected, fields with a role of Input
are treated as candidate inputs and fields with a role of Both are treated as both candidate inputs and targets
for temporal causal modeling.

The temporal causal modeling procedure determines the best inputs for each target from the set of candidate
inputs. In this example, the candidate inputs are the fields Lever1 through Lever5 and the fields KPI_1
through KPI_25.

5. Click Observations in the Select an item list on the Fields tab.

Figure 3. Temporal Causal Modeling Observations

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 63/83
27/2/2020 Forecasting Option

The Observations settings specify the fields that define the observations. Notice that the Time interval
was automatically set to Weeks, as a result of scanning the data.

6. Go to the Model Options tab and click Save in the Select an item list.

Figure 4. Temporal Causal Modeling Save

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 64/83
27/2/2020 Forecasting Option

7. Select Export model system to save the temporal causal model system to a file.

The model system file can be used by the Temporal Causal Model Forecasting procedure to obtain updated
forecasts or to generate any of the available output. It can also be used by the Temporal Causal Model
Scenarios procedure to run scenario analysis.

8. Click Browse to go to where you want to save the model system file and enter a name for the file.
9. Click Run.

Next

Parent topic: Uncovering causal relationships in business metrics

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 65/83
27/2/2020 Forecasting Option

8.2. Overall Model Quality Chart

The Overall Model Quality output item, which is generated by default, displays a bar chart and an associated dot
plot of the model fit for all models. There is a separate model for each target series. The model fit is measured by
the chosen fit statistic. This example uses the default fit statistic, which is R Square.

The Overall Model Quality item contains interactive features. To enable the features, activate the item by double-
clicking the Overall Model Quality chart in the Viewer.

Figure 1. Overall Model Quality

Clicking a bar in the bar chart filters the dot plot so that it displays only the models that are associated with the
selected bar. Hovering over a dot in the dot plot displays a tooltip that contains the name of the associated series
and the value of the fit statistic. You can find the model for a particular target series in the dot plot by specifying
the series name in the Find model for target box.

Next

Parent topic: Uncovering causal relationships in business metrics

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 66/83
27/2/2020 Forecasting Option

8.3. Overall Model System

The Overall Model System output item, which is generated by default, displays a graphical representation of the
causal relations between series in the model system. By default, relations for the top 10 models are shown, as
determined by the value of the R Square fit statistic. The number of top models (also referred to as best-fitting
models) and the fit statistic are specified on the Series to Display settings (on the Build Options tab) of the
Temporal Causal Modeling dialog.

The Overall Model System item contains interactive features. To enable the features, activate the item by double-
clicking the Overall Model System chart in the Viewer. In this example, it is most important to see the relations
between all series in the system. In the interactive output, select All series from the Highlight relations for drop-
down list.

Figure 1. Overall Model System, view for all series

All lines that connect a particular target to its inputs have the same color, and the arrow on each line points from
an input to the target of that input. For example, Lever3 is an input to KPI_19.

The thickness of each line indicates the significance of the causal relation, where thicker lines represent a more
significant relation. By default, causal relations with a significance value greater than 0.05 are hidden. At the 0.05
level, only Lever1, Lever3, Lever4, and Lever5 have significant causal relations with the key performance
indicator fields. You can change the threshold significance level by entering a value in the field that is labeled
Hide links with significance value greater than.

In addition to uncovering causal relations between Lever fields and key performance indicator fields, the analysis
also uncovered relations among the key performance indicator fields. For example, KPI_10 was selected as an
input to the model for KPI_2.

You can filter the view to show only the relations for a single series. For example, to view only the relations for
KPI_19, click the label for KPI_19, right-click, and select Highlight relations for series.

Figure 2. Overall Model System, view for single series

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 67/83
27/2/2020 Forecasting Option

This view shows the inputs to KPI_19 that have a significance value less than or equal to 0.05. It also shows that,
at the 0.05 significance level, KPI_19 was selected as an input to both KPI_18 and KPI_7.

In addition to displaying the relations for the selected series, the output item also contains information about any
outliers that were detected for the series. Click the Series with Outliers tab.

Figure 3. Outliers for KPI_19

Three outliers were detected for KPI_19. Given the model system, which contains all of the discovered
connections, it is possible to go beyond outlier detection and determine the series that most likely causes a
particular outlier. This type of analysis is referred to as outlier root cause analysis and is covered in a later topic in
this case study.

Next

Parent topic: Uncovering causal relationships in business metrics

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 68/83
27/2/2020 Forecasting Option

8.4. Impact Diagrams

You can obtain a complete view of all relations that are associated with a particular series by generating an impact
diagram. Click the label for KPI_19 in the Overall Model System chart, right-click, and select Create Impact
Diagram.

Figure 1. Impact Diagram of effects

When an impact diagram is created from the Overall Model System, as in this example, it initially shows the
series that are affected by the selected series. By default, impact diagrams show three levels of effects, where the
first level is just the series of interest. Each additional level shows more indirect effects of the series of interest.
You can change the value of the Number of levels to display to show more or fewer levels of effects. The impact
diagram for this example shows that KPI_19 is a direct input to both KPI_18 and KPI_7, but it indirectly affects a
number of series through its effect on series KPI_7. As in the overall model system, the thickness of the lines
indicates the significance of the causal relations.

The chart that is displayed in each node of the impact diagram shows the last L+1 values of the associated series
at the end of the estimation period and any forecast values, where L is the number of lag terms that are included in
each model. You can obtain a detailed sequence chart of these values by single-clicking the associated node.

Double-clicking a node sets the associated series as the series of interest, and regenerates the impact diagram
based on that series. You can also specify a series name in the Series of interest box to select a different series of
interest.

Impact diagrams can also show the series that affect the series of interest. These series are referred to as causes.
To see the series that affect KPI_19, select Causes of series from the Show drop-down.

Figure 2. Impact Diagram of causes

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 69/83
27/2/2020 Forecasting Option

This view shows that the model for KPI_19 has four inputs and that Lever3 has the most significant causal
connection with KPI_19. It also shows series that indirectly affect KPI_19 through their effects on KPI_7 and
KPI_17. The same concept of levels that was discussed for effects also applies to causes. Likewise, you can
change the value of the Number of levels to display to show more or fewer levels of causes.

Next

Parent topic: Uncovering causal relationships in business metrics

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 70/83
27/2/2020 Forecasting Option

8.5. Determining root causes of outliers

Given a temporal causal model system, it is possible to go beyond outlier detection and determine the series that
most likely causes a particular outlier. This process is referred to as outlier root cause analysis and must be
requested on a series by series basis. The analysis requires a temporal causal model system and the data that was
used to build the system. In this example, the active dataset is the data that was used to build the model system.

To run outlier root cause analysis:

1. From the menus, choose:

Analyze > Forecasting > Apply Temporal Causal Models...

Figure 1. Apply Temporal Causal Models

2. Browse to the location where you saved the temporal causal model system file and select the file.
3. Click the button that is labeled Reestimate models with updated data, create forecasts or generate
output describing the temporal causal model system.
4. Click Continue.
5. In the Temporal Causal Model Forecasting dialog, click the Options tab and then click Series to display in
the Select an item list.

Figure 2. Temporal Causal Model Forecasting Series to display

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 71/83
27/2/2020 Forecasting Option

6. Move KPI_19 to the Fields to display list.

7. Click Output options in the Select an item list on the Options tab.

Figure 3. Temporal Causal Model Forecasting Output options

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 72/83
27/2/2020 Forecasting Option

8. Deselect Overall model system, Same as for targets, R square, and Series transformations.
9. Select Outlier root cause analysis and keep the existing settings for Output and Causal levels.
10. Click Run.
11. Double-click the Outlier Root Cause Analysis chart for KPI_19 in the Viewer to activate it.

Figure 4. Outlier Root Cause Analysis for KPI_19

The results of the analysis are summarized in the Outliers table. The table shows that root causes were found for
the outliers at 2009-04-05 and 2010-09-19, but no root cause was found for the outlier at 2008-10-12. Clicking a
row in the Outliers table highlights the path to the root cause series, as shown here for the outlier at 2009-04-05.
This action also highlights the selected outlier in the sequence chart. You can also click the icon for an outlier
directly in the sequence chart to highlight the path to the root cause series for that outlier.

For the outlier at 2009-04-05, the root cause is Lever3. The diagram shows that Lever3 is a direct input to
KPI_19, but that it also indirectly influences KPI_19 through its effect on other series that affect KPI_19. One of
the configurable parameters for outlier root cause analysis is the number of causal levels to search for root causes.
By default, three levels are searched. Occurrences of the root cause series are displayed up to the specified
number of causal levels. In this example, Lever3 occurs at both the first causal level and the third causal level.

Each node in the highlighted path for an outlier contains a chart whose time range depends on the level at which
the node occurs. For nodes in the first causal level, the range is T-1 to T-L where T is the time at which the outlier
occurs and L is the number of lag terms that are included in each model. For nodes in the second causal level, the
range is T-2 to T-L-1; and for the third level the range is T-3 to T-L-2. You can obtain a detailed sequence chart of
these values by single-clicking the associated node.

Next

Parent topic: Uncovering causal relationships in business metrics

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 73/83
27/2/2020 Forecasting Option

8.6. Running scenarios

Given a temporal causal model system, you can run user-defined scenarios. A scenario is defined by a time series,
that is referred to as the root series, and a set of user-defined values for that series over a specified time range.
The specified values are then used to generate predictions for the time series that are affected by the root series.
The analysis requires a temporal causal model system and the data that was used to build the system. In this
example, the active dataset is the data that was used to build the model system.

To run scenarios:

1. From the menus, choose:

Analyze > Forecasting > Apply Temporal Causal Models...

Figure 1. Apply Temporal Causal Models

2. Browse to the location where you saved the temporal causal model system file and select the file.
3. Click the button that is labeled Run scenarios to investigate how specific values of inputs affect
predictions.
4. Click Continue.
5. In the Temporal Causal Model Scenarios dialog, click Define Scenario Period.

Figure 2. Scenario Period

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 74/83
27/2/2020 Forecasting Option

6. Select Specify by time intervals relative to end of estimation period.

7. Enter -3 for the starting interval and enter 0 for the ending interval.

These settings specify that each scenario is based on values that are specified for the last four time intervals
in the estimation period. For this example, the last four time intervals means the last four weeks. The time
range over which the scenario values are specified is referred to as the scenario period.

8. Enter 4 for the intervals to predict past the end of the scenarios values.

This setting specifies that predictions are generated for four time intervals beyond the end of the scenario
period.

9. Click Continue.
10. Click Add Scenario on the Scenarios tab.

Figure 3. Scenario Definition

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 75/83
27/2/2020 Forecasting Option

11. Move Lever3 to the Root Field box to examine how specified values of Lever3 in the scenario period affect
predictions of the other series that are causally affected by Lever3.
12. Enter Lever3_25pct for the scenario ID.
13. Select Specify expression for scenario values for root field and enter Lever3*1.25 for the expression.

This setting specifies that the values for Lever3 in the scenario period are 25% larger than the observed
values. For more complex expressions, you can use the Expression Builder by clicking the calculator icon.

14. Click Continue.

15. Repeat steps 10 - 14 to define a scenario that has Lever3 for the root field, Lever3_50pct for the scenario
ID, and Lever3*1.5 for the expression.

Figure 4. Scenarios

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 76/83
27/2/2020 Forecasting Option

16. Click the Options tab and enter 2 for the maximum level for affected targets.
17. Click Run.
18. Double-click the Impact Diagram chart for Lever3_50pct in the Viewer to activate it.

Figure 5. Impact Diagram for Scenario: Lever3_50pct

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 77/83
27/2/2020 Forecasting Option

The Impact Diagram shows the series that are affected by the root series Lever3. Two levels of effects are
shown because you specified 2 for the maximum level for affected targets.

The Forecasted Values table includes the predictions for all of the series that are affected by Lever3, up to
the second level of effects. Predictions for target series in the first level of effects start at the first time
period after the beginning of the scenario period. In this example, predictions for target series in the first
level start at 2010-10-10. Predictions for target series in the second level of effects start at the second time
period after the beginning of the scenario period. In this example, predictions for target series in the second
level start at 2010-10-17. The staggered nature of the predictions reflects the fact that the time series
models are based on lagged values of the inputs.

19. Click the node for KPI_5 to generate a detailed sequence diagram.

Figure 6. Sequence Diagram for KPI_5

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 78/83
27/2/2020 Forecasting Option

The sequence chart shows the predicted values from the scenario, and it also shows the values of the series
in the absence of the scenario. When the scenario period contains times within the estimation period, the
observed values of the series are shown. For times beyond the end of the estimation period, the original
forecasts are shown.

Next

Parent topic: Uncovering causal relationships in business metrics

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 79/83
27/2/2020 Forecasting Option

8.7. Forecasting with updated data

Given a temporal causal model system, you can generate new forecasts when more current data are available
without rebuilding the model system.

1. Open the data file tcm_kpi_upd.sav.

This dataset contains the data for four more weeks beyond the data in tcm_kpi.sav. For more
information, see the topic Sample Files.

2. From the menus, choose:

Analyze > Forecasting > Apply Temporal Causal Models...

Figure 1. Apply Temporal Causal Models

3. Browse to the location where you saved the temporal causal model system file and select the file.
4. Click the button that is labeled Reestimate models with updated data, create forecasts or generate
output describing the temporal causal model system.
5. Click Continue.

Figure 2. Temporal Causal Model Forecasting

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 80/83
27/2/2020 Forecasting Option

6. On the Model tab of the Temporal Causal Model Forecasting dialog, select Reestimate from data, keep
the default selection of All observations, and enter 4 for the value of Extend records into the future.

Specifying Reestimate from data means that the model parameters are reestimated based on the updated
data but that the structure of the models does not change. Specifying All observations ensures that the
estimation period includes the four extra time periods in the updated data. Specifying 4 for the value of
Extend records into the future generates forecasts for four weeks beyond the end of the estimation period.

7. Click the Options tab and then click Output options in the Select an item list on the Options tab.

Figure 3. Temporal Causal Model Forecasting Output options

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 81/83
27/2/2020 Forecasting Option

8. Deselect all selected items on Output options.

9. Click the Save tab.

Figure 4. Temporal Causal Model Forecasting Save

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 82/83
27/2/2020 Forecasting Option

10. Select Predicted Values and keep the default value of Predicted for the root name. Select All targets for
Targets to save and enter Predictions for the dataset name.
11. Click Run and then select the new Predictions dataset.

Figure 5. New dataset with predictions

The Data Editor for the new Predictions dataset shows the variables that contain the model predictions.
Although only two are shown here, there are 25 new variables, one for each of the 25 targets. The variable names
consist of the default prefix Predicted, followed by the name of the associated target. The dataset contains four
extra cases for the forecasts, along with automatically generated values of the date field in the forecast period.
Parent topic: Uncovering causal relationships in business metrics

127.0.0.1:49194/help/advanced/print.jsp?topic=/../nav/3_13&pageBreak=true&breakConfirmed=true 83/83