FAQ: Predictive Forecaster

Select a different plan version for the forecast. You might consider creating a designated plan version for forecasted data. You can use reports and dashboard charts to compare the differences between versions.

Go to
Modeling
>
Predictive Forecaster
.
From the list, hover over the forecast until the 3 dots display.
Click the 3 dots and select
View History
.
In the left panel is a list of runs for that forecast.
Click a failed run.
In the
Status
section, read the list of errors.

Try AutoFit, which applies the best algorithm based on the data.
Reference a large number of actuals data points.
Check your actuals for blanks, which the algorithm treats as zeroes.
Use the Croston algorithm when your actuals data has a lot of blanks.
Forecast only as far into the future as necessary.
Limited availability: Enable the Accuracy Metric, which generates a score that you can review in the Confidence Metrics tab.

See Reference: Machine Learning Algorithms for Adaptive Planning.

Use AutoFit when you:

Use Predictive Forecaster for the first time.
Don’t have a good grasp on the internal workings of the available algorithms.
Don’t have a clear picture of the features and patterns in your data.
Aren’t getting the results that you anticipated when you selected other algorithms.
Limited Availability: Received a low score for the
Accuracy Metric
in the
Confidence Metrics
subsection.

AutoFit uses an additional cycle that might add additional time for the predictions.

Use the filters to limit the accounts, levels, and custom dimensions.
Select a faster algorithm, like Holt-Winters or Kalman Filter. Always check the forecasted data against your expectations.
Limited Availability: Disable the
Forecast Range
check box. Disabling this option decreases the run time of the forecast because you stop generating the upper limit and lower limit of the forecast. You can enable it again at a later date.
Limited Availability: Disable the
Accuracy Metric
check box. Disabling this option decreases the run time because it eliminates the back-testing. Consider disabling after you've run the forecast and checked the score. You can enable it again at a later date.

The algorithm calculates the forecasted data pertaining to the missing level, account, or dimension without using the levers.

Don’t add custom dimensions to the lever sheet or leave the column of the custom dimension blank for a particular row.

Daily cadence is a calculation based on the calendar days and time periods within the time range. To calculate:

Calculate how many days are in the defined time range of the forecast. Example: Start Period is January 2023. End Period is February 2023. There are 59 calendar days on a typical calendar between January 1 and February 28.
Calculate how many time periods are in the time range. Example: Use the same Start and End period as above. The smallest time strata of your calendar is weeks. There are 9 weeks between January 1 and February 28.
Divide the calendar days by the time periods. Example: 59 divided by 9 equals 6.55.
The quotient must be equal to, or less than, 2 to qualify for the weekly seasonality option. Example: The cadence of 6.55 doesn't qualify.

Go to
Modeling
>
Predictive Forecaster
.
From the list of forecasts, click the name of the forecast.
Click the prompt for
Accounts
,
Levels
, or
Dimensions
.
Click the
Selected
tab. The list displays all selected values.
To remove a value, click the x next to a value.

You didn't enable the options that generate the charts for the forecast. To enable the metrics, edit the forecast and click either

Forecast Range

or

Accuracy Metric

. Each option runs a different task and generates a different chart. See Concept: Forecast Range and Accuracy with Predictive Forecaster.

Rolling forecasts align important dates of the forecast with your actuals version. When used properly, it creates a dynamic forecast definition that rolls forward as you load actuals data. The forecast references your freshest, most recently added actuals data and continuously populates the forecast version with the most accurate predictions.