Skip to main content
Documentation
Documentation
Adaptive Planning
Last Updated: 2023-06-23
Create Machine Learning Forecasts

Create Machine Learning Forecasts

Requirements for actuals data:
  • Have enough data points in your historical data to support the algorithm. See Reference: Machine Learning Algorithms for Adaptive Planning.
  • Have a ratio of at least 3:2 actuals data points to forecasted data points. Example: To forecast 2 years, you need 3 years of actuals data.
  • Regardless of your calendar structure, to use the custom seasonality options, you need:
    • 3 years of actuals for yearly seasonality.
    • 18 months of actuals for biannual seasonality.
    • 9 months of actuals for quarterly seasonality.
Requirements for levers:
Security:
  • Access to the levels, versions, and accounts that you want to include in the forecast.
  • Predictive Forecaster
     permission.
You can improve efficiency and accuracy in your plans by creating machine learning (ML) forecasts. Predictive Forecaster uses ML algorithms of your choice to populate sheets in plan versions. The algorithms predict data based on historical references that you define in your actuals. You can also designate additional plan versions for the algorithm to reference as historical data. You can create forecasts and save them, then come back and run the forecast later.
You can also check the history of your runs, and review
Confidence Metrics
 when your instance has them available.
The forecasts replace existing data on the sheets in the plan version that you select.  You can't revert the forecasted data after you run a successful forecast.
  1. Select
    Modeling
     from the main menu.
  2. Click
    Predictive Forecaster
    .
  3. Click
    Add New
    .
  4. Enter a unique
    Name
     with fewer than 129 characters.
  5. As you complete the
    Forecast
     section, consider:
    Option Description
    Sheet
    Select the standard or cube sheet that you want to populate with the forecasted data.
    Forecast Version
    Select the plan version that you want to populate with the forecasted data. Don’t select locked versions because your forecast will fail. After the forecast fails, you can check the history and review the errors, which further direct you.
    Forecast Start Period
    Select when the machine learning data should start.
    Forecast End Period
    Select a period after the
    Forecast Start Period
     to indicate when the machine learning data should stop.
    Probability Range
    Limited availability: Contact us to see if you’re eligible for this capability.
    Click the check box to generate 2 additional forecasts. 1 includes the upper limit of the predictions. The other includes the lower limit of the predictions. This option also generates a fan chart in the Confidence Metrics tab when you view the history of the forecast. The chart illustrates the range over time. See Concept: Probability and Accuracy with Predictive Forecaster.
    Probability Level
    Limited availability: Contact us to see if you’re eligible for this capability.
    Displays when you select Probability Range.
    Select a percent between 55 and 95. A higher percent results in a broader range between the lower and upper limit.
    Upper Limit Version
    Limited availability: Contact us to see if you’re eligible for this capability.
    Displays when you select Probability Range.
    Select the plan version that you want to populate with the highest probable values of the forecast. The version must be within range of the forecast.
    Lower Limit Version
    Limited availability: Contact us to see if you’re eligible for this capability.
    Displays when you select Probability Range.
    Select the plan version that you want to populate with the lowest probable values of the forecast. The version must be within range of the forecast.
    Accuracy Metric
    Limited availability: Contact us to see if you’re eligible for this capability.
    Select the check box to run a backtest of the forecast against existing actuals. The degree of accuracy displays as a percent in the Confidence Metrics tab of when you view the history of the forecast.
  6. As you complete the
    Filters
     section, consider:
    Option Description
    Accounts
    Required. Select the accounts that you want to populate with machine learning data.
    Levels and Dimensions
    Optional. Select the levels and custom dimensions that you want to populate with machine learning data.
    We recommend using level and dimensions filters to limit the number of time series in the forecast.
    When you don't specify a filter, the forecast runs on all levels and dimensions. Example: Don't select anything from
    Levels
    . The forecast runs on all the levels included on the sheet. For larger sheets, including all levels and dimensions can extend the time it takes to complete the forecast.
    The filters include only the accounts, levels, and dimensions already on the sheet that you selected from in the
    Forecast
     section.
    For each filter:
    • You can select 1 or more accounts, levels, or dimensions.
    • Selecting parents automatically selects the descendants.
    • On the
      Selected
       tab, you can review and remove selections.
  7. As you complete the
    Reference Data
     section, consider:
    Option Description
    Actuals Version
    Select the actuals version that you want the algorithms to reference as historical data.
    Actuals Start Period
    Select a period that’s after the
    Forecast End Period
    .
    Actuals End Period
    Select a period that's after the
    Actuals Start Period
    . You need at least a 3:2 ratio of actuals to forecast. You also need the required number of data points for the algorithm that you select.
    Add Plan Version
    Click to add a plan version for more reference data. Ensure the time range of the plan reference doesn't overlap with the forecast or actuals time range.
    Example: The forecast starts January 2023 and ends March 2023. The actuals reference starts January 2022 and ends December 2022. The plan reference must not overlap periods between January 2022 and March 2023.
  8. As you complete the
    Algorithm
     section, consider:
    Option Description
    Algorithm
    Select an algorithm. You might need to go through a trial and error process to find the best algorithm for your data.
    Seasonality
    Available for these algorithms:
    • AutoFit
    • Holt-Winters
    • Kalman Filter
    • Orbit DLT
    • Prophet
    AutoFit only applies the seasonality that you define for algorithms that support seasonality.
    When you select:
    • Yearly
      , you must have 3 years of actuals data.
    • Biannual
      , you must have 18 months of actuals data.
    • Quarterly
      , you must have 9 months of actuals data.
    • None
      , your data has no seasonality.
    For the Kalman Filter algorithm, select the
    Data varies seasonally
     check box to indicate that you want to model with seasonality. Clear the check box when you don’t want to account for seasonality.
    Weekly Seasonality
    Available for:
    • AutoFit.
    • Prophet.
    Select the check box when you also want the algorithm to consider weekly patterns. This option requires that you have a daily cadence in the time ranges of the forecast and reference versions.
    When you select AutoFit, the weekly seasonality only applies to the data that uses Prophet.
    Multiplicative Seasonality
    Available for:
    • AutoFit
      .
    • Holt-Winters
      .
    • Prophet
      .
    Select the check box to indicate that the seasonality is proportional to the variations in trend.
    Round Values
    Select the check box when you don't want negative numbers in the forecast.
    Lever Sheet
    Available for:
    • AutoFit.
    • Orbit DLT.
    • Prophet.
    Select up to 3 sheets that contains regressor data for the algorithm to reference. Before selecting this option:
    For AutoFit, the regressor data applies to the data that uses Prophet or Orbit DLT.
  9. Click
    Run
    .to start the forecast.
    You can also click
    Save
     to run it later.
The forecast displays in the list with a status:
Status
Description
Blank
Saved the parameters of the forecast but didn't run it yet.
In Queue
Ran the forecast, but it can't start until pending forecasts complete.
In Progress
The forecast is running. We provide the percent of completion and a status bar in the
Status
 section on the
Forecast History
 page.
Success
The forecasted data has populated the plan version. From the forecast list, click the sheet name to open the sheet. Check that you're in the version you selected for the forecasted data. Review the predicted data.
Failed
The forecast had errors. To help you troubleshoot and try again, we display a list of errors in the
Status
 section on the
Forecast History
 page.