Taking Control of Models in SAS Visual Forecasting

[#1]

It has been a while since I have discussed the topic of time series. I think it is about time that we revisited this analysis using SAS Visual Forecasting functionality in the Model Studio area of SAS Viya. In this time of big data, the amount of time series that may need to be analyzed is growing. SAS Visual Forecasting is a great tool for assisting with quickly modeling many series with minimal interaction from you. However, there will always be high value series that warrant extra time and focus to obtain the best forecasting model possible. In this post, we will discuss how you can leverage your knowledge of the business and data to improve forecasts for high value series via the Interactive Modeling Node within SAS Visual Forecasting.

Adding the Node to the Pipeline

We begin with a pipeline that already contains an executed modeling node. On the left pane, click the Nodes icon and then expand the Postprocessing section. Click and drag the Interactive Modeling node onto the connecting segment under the current modeling node. This will add our Interactive Modeling node to the pipeline.

Select any image to see a larger version.
Mobile users: To view the images, select the "Full" version at the bottom of the page.

Before we can utilize the Interactive Modeling node, we need to run the pipeline. Once it is completed, we can then right-click on the Interactive Modeling node and select Open.

Know the Display

Let's get to know the parts of this display. On the left pane, we see items that will look familiar if you've used the Time Series Viewer or the Forecast Viewer before. These are the primary and secondary attributes that we provided to our project from either BY variables or from an external data set. With this, we can subset the data down to collections of series on which we may want to focus.

On the right pane, we see the list of series that are in the collection dictated by the selections on the left pane. When we first enter the node, there are no attributes selected, thus the right pane shows all the series at the base level of the data hierarchy. Selecting a single series from the right pane will display it. Diagnostics are available that allow us to explore and assess the series. We can also edit existing models or create new models from scratch for the selected series.

The center pane contains the tools and visualizations that we will need to interact with the selected series. This pane is divided into three tabs: Forecast, Modeling, and Series Analysis. The Forecast area will look very familiar as it is presenting the information from the Forecast Viewer.

When multiple series are selected, the image shown is an envelope plot. When one series of a collection is selected, information about that single series is overlaid on the envelope plot. If selected attributes reduce the selection to a single series only, then we will see the plot of that series in our viewer.

The Series Analysis area lets you explore the selected time series in more detail. You can request to see Basic analysis, Autocorrelation analysis, Decomposition analysis, Seasonal Adjustment analysis, and Stationarity analysis. Remember that all these diagnostics will be derived from the selected series prior to any modeling being performed. You can also drag model inputs from the left into the bottom pane and obtain the described diagnostic plots for them.

Now, we will switch to the Modeling tab. The Modeling area displays a list of all the models that were tried for the selected series from connected modeling nodes in the pipeline. The flagged model named PREDECESSOR is the model that was declared the champion model from the connected, predecessor modeling nodes. After selecting one of these models, the bottom pane provides summary measures and diagnostics about the generated residuals, parameter estimates and model fit.

A point of confusion is that we see some of the same plots and summary measures here that we encountered in the Series Analysis tab. Remember that diagnostics and plots generated here are from the post modeling perspective. For example, the White Noise test generated in the Modeling tab tests the selected model's residuals, and not the underlying series.

The Power of Customization

For a single series, you may also want to leverage tools outside of this application. Procedures such as PROC TIMESERIES, PROC ARIMA in SAS/ETS provide a comprehensive toolbox for exploring, assaying and identifying custom models for a single time series. Once you have assessed a high value series and identified a model using these methods, you can then return to SAS Visual Forecasting to build your custom model in the Interactive Modeling node.

One approach for building custom models is to start with one of the generated models and modify it. With the candidate model selected, click on the Copy button on the top right of the middle pane. This opens a dialog box where you can make changes to the transfer function and ARIMA parts of the selected model.

When done, click on the SAVE button. Your model will be executed on the selected time series and results from it will be added to the middle pane. Note that the Type of your modified model is Custom. It will not be modified because of data updates or other changes to the forecasting project.

You will have the option to make this new model the champion for this selected series if you wish. Right-click on your new model and select Set as Champion.

But what if you want to start a new model from scratch with the information that you found during your exploration? Not to worry. Use the Create Model button and select which model type you would like to try for the selected series.

The dialog box that appears will resemble the one that we saw during the model editing process; however, this model will completely be under your control, with no initial settings.

SAS Visual Forecasting, within Model Studio, is an excellent resource for automated, large-scale forecasting It is great to see that you do have the power to customize and develop models for high value timeseries where your knowledge of the data and business can add precision to the generated forecasts With the use of the Interactive Modeling node, you have this power at your fingertips. Give this node a try in your next SAS Visual Forecasting project.

Find more articles from SAS Global Enablement and Learning here.