Is the accuracy of your forecast being skewed but a few outliers in your historical data?
New in PA Workspace 75 is the ability to automatically detect outlier values and then optionally remove those outliers from forecast calculations.
Generally, an outlier is a historical value that deviates outside the typical value range. Technically, it's a value which is outside the confidence interval (CI) of the forecast model after the CI is propagated backward into history.
For example in the chart below, the outlier value for Nov 2021, shown in red, is slightly outside the confidence interval envelope, shown is light blue, when the envelope is propagated backwards into history.
A new icon “Outliers Detected” is displayed (PAW 75) to indicate that outliers were detected during a forecast preview.
It should be noted that outliers are detected in the most recent 20% of historical data, this is because the early part of history is used the ‘train’ the forecast model, and the later part (20%) is used to test the forecast model – outliers are detected during the later test stage.
After detecting an outlier, the user is presented with the option to correct the outlier (or substitute the outlier value with a more appropriate value from the forecast model). This corrected value, shown in purple, is used for calculating future forecast values.
As shown in the example below, this correction has resulted in a 29% accuracy improvement, and more confidence (smaller confidence envelope).
Despite correction, the corrected value DOES NOT replace the historical actual value. In this example, the value for Nov 2021 does not change in PA/TM1, the corrected outlier value is simply used in the forecast calculation.
As shown, correction of outlier typically results in improved accuracy and confidence and a better forecast.
If you have any questions, please post them here.
#PlanningAnalyticswithWatson