How to use Time Series Forecasting in SAP Analytics Cloud

Late 2018, within SAP Analytics Cloud, the first set of ‘Predictive Modelling’ was introduced. This set includes ‘Automatic Predictive Modelling’ for:

• Classification;

• Regression;

• Segmented Time Series Analysis.

For these three Automatic Predictive Models a ‘Predictive Scenario’ can be created in the following way:

In this blog, I will explore the Segmented Time Series Analysis for an existing Planning Model. This Planning Model contains the Financial Statement from 2017 up until March 2019, were a forecast for some of the remaining months of 2019 will be created:

Predict, store, plan

You might wonder why you would want to use the Segmented Time Series Analysis for this purpose, as there are already some Forecast options made available for Graphs. As is already done in this story for the ‘Net Revenue Trend’:

However, one of the benefits of the Predictive Scenario’s is that the output can be stored, something which the ad-hoc forecast within the Net Revenue Trend graph can’t. In this case this means that the output of the Forecast can be used as input for decisions made by the planner.

Exclude irrelevant variables

Going back to the Predictive Scenario within SAP Analytics Cloud: after having selected the ‘Time Series’-model, I can now create a new ‘Predictive Model’ and give it a generic description along with defining the input Dataset:

Having selected a Dataset as input, the following input-options become available:

• Signal Variable (i.e. which measure to Forecast);

• Date Variable (i.e. Forecasting over which Time Dimension);

• Segmented By (i.e. considering which object to use for segmentation), which is optional.

Besides these generic ‘Variable Roles’, irrelevant Variables can be excluded from the training of the algorithm.

For the actual Training of the Forecasting model, I can select the:

• Process: either all observations or a user-defined ‘Window of Observations’;

• Range: until which period the Forecast should run;

• Forecast: the ‘Number of Forecasts’, and whether only ‘Positive Forecasts’ should be included.

Please note that besides the above options, some output settings are quite well ‘hidden’. E.g. the ‘Selected Output Columns’ can only be selected after clicking open your Predictive Model:

Create the first model

After all parameters are set, the model can be Trained for the first time. When the model is successfully trained, overview statistics become available. Additionally, the model can now be ‘Applied’ and ‘Published’:

If you Apply the Trained Model, the Model’s output can be stored as a Dataset:

Instead, you could also opt to Create and Train a second Predictive Model, look at the Model Overview, and, compare the results between the models.

Being satisfied with the first model, the stored Dataset shows a Forecasted value, for example for April 2019 (kts_1), along with its 95% lower- and upper Confidence interval value (kts_1_lower- and kts_1_upperlimit_95% respectively):

Looking at the roadmap of SAP with regards to SAP Analytics Cloud and Predictive capabilities, (luckily) there seems to be a lot of focus on additional Predictive Modelling Scenario’s, (better) Data Connectivity (e.g. for exports of the output datasets) and Tasks/Scheduling.