跳至主要內容
文件
文件
Adaptive Planning
上次更新時間 ：2025-09-19
Setup Considerations: Predictive Forecaster

Setup Considerations: Predictive Forecaster

You can use this topic to help make decisions when planning your configuration and use of Predictive Forecaster. It explains:
  • Why to set it up.
  • How it fits into the rest of Workday.
  • Downstream impacts and cross-product interactions.
  • Security requirements and business process configurations.
  • Questions and limitations to consider before implementation.
Refer to detailed task instructions for full configuration details.

What It Is

Predictive Forecaster is an Intelligent Planning capability powered by machine learning (ML). Predictive Forecaster seeds plan versions with ML-generated data, which serves as a reliable starting point for budget managers. With Predictive Forecast, you select and test various algorithms to generate time series forecasts based on the historical and regressor data that you define within Adaptive Planning.

Business Benefits

  • Use advanced ML technology within Adaptive Planning.
  • Improve accuracy by enabling ML to process large amounts of data, patterns, and drivers.
  • Decrease the time that it takes to complete a planning cycle with pretrained ML algorithms.
  • Generate seed data for large sets of intersections with little manual effort.
  • Use schedules to automate the runs of forecasts and rolling forecasts to generate forecasts on a rolling basis as you import actuals data.
  • Preview tenants only: Use schedules to automate the runs of forecasts.

Use Cases

As a retail company, you need to plan how many product units you’re expecting to sell from each location. You also want to analyze a prospective product. You have a cube sheet that holds sales details from previous years. You have a P&L standard sheet that holds your revenue accounts.
You can:
  • Define the historical reference. Your historical data must include actuals, but you can also add plan versions to the reference and levers. Example: You can define your actuals reference to start in 2018 and end at the end of 2019. You can skip the fiscal period for 2020 in your actuals because it was an unusual year. Instead, you can reference the historical plan version for 2020.
  • Choose an algorithm that supports seasonality when your data requires it. Example: When sales increase at the end of year for the holiday season, you can specify that you want the algorithm to account for annual seasonality patterns.
  • Generate different sets of forecasted data into different sheets, or the same sheet for different versions. Example: You can generate a forecast in the 2025 Budget Plan for the P&L standard sheet for all revenue accounts in the general ledger. You can also generate a forecast for the 2025 What If Analysis in your Revenue Cube for a specific product dimension.
  • Use level and dimension filters to pinpoint the forecasted data. Example: You can run a forecast for the Northwest Sales level and specific products. Run a separate forecast for the Southwest Sales and different products.
  • Use different algorithms for different sets of data or different runs of the same data. Example: Select the Prophet algorithm for the forecast for P&L accounts. After you review the forecasted data, you can try N-BEATS. Compare the forecasts with reports to see which algorithm best suits that particular set of data.
  • Set up the forecast at the beginning of the year as you start your planning cycle. Use the scheduling and rolling forecast capability to automatically generate forecasts as you collect more actuals data for the year.
  • Save time by duplicating existing forecasts, which enables you to make quick changes from a base definition. Example: You can duplicate the start and end periods, while changing only the accounts.

Questions to Consider

Question
Considerations
How accurate have your forecasts been in the past and how much time does it take your team to plan forecasts?
We provide ML algorithms that are a science- and math-based way to get more accurate forecasts and budgets. Using ML forecasts provides your team seed data as a starting point, which can save your team valuable time and improve accuracy.
Does your actuals data work with Predictive Forecaster?
  • You must have a ratio of 3:2 actuals data points to forecasted data points. Example: A 2-year forecast requires 3 years of actuals data.
  • In addition, some algorithms have a minimum requirement for data points in the historical data. See Reference: Machine Learning Algorithms for Adaptive Planning .
  • Each forecast runs on a specific standard or cube sheet. You can only forecast data from and to cube and standard sheets.
  • You can only forecast and reference general ledger accounts, custom accounts, and cube standard accounts.
Do your actuals have many blanks or zeros?
The more complete your actuals data, the more accurate the forecast. The algorithms treat historical data that’s missing or blank as zeros, which can affect the forecast. When your actuals data is sparse, use the algorithms best suited for sparse data. Reference: Machine Learning Algorithms for Adaptive Planning
How frequently do you load new actuals?
You can use the rolling forecasts to align the forecast start and end periods with the
Completed Values Through
 date of the reference data. Selecting this option shifts the forecast periods forward and shifts the reference data forward to ensure fresh and more accurate predictions.
How far out do you want to forecast?
Set up your forecast start and end periods so you forecast out only as far as necessary. The further you forecast, the less accurate the forecast becomes.
Besides actuals data in Adaptive Planning, what other data do you want to use for your forecasts?
You can use external data in the forecast by importing into a lever modeled sheet. To use levers, you must use 1 of these algorithms:
  • ARIMA
  • AutoFit
  • LightGBM
  • Orbit DLT
  • Prophet
AutoFit only uses the lever data when it applies an algorithm that supports levers.
You can also use the seasonality options, except when you select these algorithms:
  • Croston
  • LightGBM
  • N-BEATS
How comfortable are you defining seasonality?
Seasonality is the interval of time in which values have a repetitive pattern. Use these algorithms when you want to define the seasonality patterns:
  • AutoFit
  • Holt-Winters
  • Orbit DLT
  • Prophet
AutoFit only uses the seasonality that you define when it selects an algorithm that supports seasonality.
Use these algorithms to let the algorithm capture the seasonality on its own:
  • ARIMA
  • Kalman Filter
  • LightGBM
  • N-BEATS

Recommendations

  • Filter forecasts by fewer accounts, levels, and custom dimensions. Limiting your forecast definition generates the forecasted data faster.
  • Run or schedule forecasts at least 1 day before you expect the forecasted data to populate in your plan. Forecasts with many accounts and levels might take longer to complete.
  • When your historical data has many blanks due to sparsity, we recommend that you use Croston or AutoFit, which selects an algorithm for each data series based on the reference data.
  • Schedule forecasts to run automatically. Or, designate a time and a person to run the forecasts so that you don't forget.
  • Use intuitive names for the forecast so that you can run the forecast again without having to redefine it.
  • Sync the forecast with the latest loads of actuals data by enabling the
    Rolling Forecast
     check box. This check box shifts the forecast and reference forward as you load actuals data to include the most recent actuals reference.

Requirements

  • You must have actuals data present in the standard or cube sheet that you select for historical and forecasted data.
  • To use levers, you must import the external data into a modeled sheet, and you must select the ARIMA, AutoFit, LightGBM, Orbit DLT, or Prophet algorithm.
  • Regardless of your calendar structure, to use the seasonality options when they’re available, you need the equivalent of:
    • 3 years of actuals for yearly seasonality.
    • 18 months of actuals for biannual seasonality.
    • 9 months of actuals for quarterly seasonality.
  • Many algorithms have required  minimums for data points in the historical reference. In addition, you must have a ratio of at least 3:2 actuals data points to forecasted data points. Example: A 2-year forecast requires 3 years of actuals data.
  • For the weekly seasonality option, the time range for all versions referenced in the forecast must have a daily cadence. See FAQ: Predictive Forecaster.

Limitations

  • You can’t forecast modeled sheets or modeled accounts.
  • Sheets, Explore Cell, and audit trail don't currently indicate when ML has generated the data.
  • You can't forecast metric accounts or any accounts with default formulas. However, you can reference ML-predicted data in formulas and links.
  • You can only use up to 3 regressors.
  • The forecast doesn't automatically include updated reference data in your actuals. You must manually update the forecast period to include the updated reference data in your actuals.
  • When you use a non-standard calendar, results are less accurate. Example: 4-4-5 calendars for retail use cases.
  • You can't create forecasts for sheets with custom dimensions that have enabled this setting:
    Data import automatically creates dimension values
    .

Tenant Setup

No impact.

Security

  • Predictive Forecaster
     permission.
  • Access to the levels, accounts, and versions that you want to include in the forecast.

Business Processes

No impact.

Reporting

You can report on forecasts generated by Predictive Forecaster.

Integrations

No impact.

Connections and Touchpoints

Workday offers a Touchpoints Kit with resources to help you understand configuration relationships in your tenant. Learn more about the Workday Touchpoints Kit on Workday Community.

Other Impacts

Running forecasts impacts the overall performance of your model in much the same way that any integration import process does.