Forecasting tutorial: outlier detection and adjustment

An outlier is any point in the last 20% of historical data that is outside of the confidence range for that point in history. The Watson Studio library automatically identifies outliers in your historical data. When forecasting in Planning Analytics Workspace you can choose to adjust for outliers in your forecast. The forecast model is corrected by using a normalized data point for the outlier to increase the forecast accuracy. The actual data point that is identified as an outlier is not modified.

Before you begin

This tutorial is meant to be followed in sequential order. Each part assumes that the previous part was completed.

About this task

When forecasting, many models are evaluated to determine which one proves to be the most accurate overall, then the one with highest overall accuracy is selected. Although not displayed in Planning Analytics Workspace, the points along the model history have a confidence interval similar to that displayed with the forecasted values. Where many points along the model history may map nearly exactly to the actual historical points received, there may be a subset that sit outside of the confidence level, and it is these points that are identified as outliers. in a forecast

Procedure

Configure the Exploration view of the Tutorial_<your initials> cube as in the previous section:
- Years on the columns
- Members on the rows
- Version on the context area, with the Actual member displayed
Select all row members, then click the Forecast icon on the Exploration toolbar.
Click Preview.
Use the Forecast selector to cycle through the previews for all five row members.
The forecast previews for Member1, Member2, and Member3 have varying degrees of predictive accuracy, but none of these previews contain outliers. Conversely, the previews for Member4 and Member 5 contain outliers, as indicated by the Outliers detected button on the preview.
On the preview for Member4, click the Outliers detected button to open the Outliers detected panel.
Click Adjust outliers to adjust the outlier values in the forecast preview.

When outliers are adjusted in this way, an Outliers adjusted toggle is available on the forecast preview. You can toggle this option on or off to see the impact of outlier adjustment on your forecast preview.

Alternatively, you can enable the Advanced option to automatically adjust outliers in a forecast. Click the Advanced tab on the Forecast page, then toggle the Adjust outliers option to the On position.

When Adjust outliers is enabled, outliers are automatically corrected before a preview is generated. You don't see an Outliers adjusted option when a preview is generated with Adjust outliers enabled, but the preview does show where the outliers exists in the historical data and how the preview was adjusted to compensate for the outliers.