The Predictions tool generates customized forecasts automatically for inclusion in your model. It leverages advanced statistical and Machine Learning algorithms to analyze your historical data and make predictions about future values. You can use it for a wide range of planning needs, including demand planning, financial forecasting, or revenue projections. Use it with any time-dependent Metric in your Pigment Workspace.

What are the benefits of using Predictions?

Predictions allows you to go beyond manual inputs or traditional statistical methods with the following benefits:

• Improved accuracy: Machine Learning models can detect more complex patterns in your data compared to manual or statistical methods.
• Use of external factors: You can incorporate external data—such as pricing, campaigns, or market indicators—and measure their actual impact on the forecast.
• No data science expertise required: The feature is designed to be simple and intuitive, so business users can run and understand predictions without needing technical knowledge.
• Fully integrated with Pigment: Predictions run natively within your Pigment Applications and can be seamlessly connected to your existing Metrics, and workflows.

Before you begin

How does Predictions work?

Predictions generates time series forecasts at the most granular level available. For example, if your Metric is Sales with Dimensions Month, Product and Region, Predictions generates one data point per Product, per Region for each Month of the forecast period. Each time series is processed independently to capture its specific trends and seasonality.

The output comprises a baseline forecast along with the upper and lower bounds of a 95% prediction interval (see image below). This is an estimate of the range in which future individual values are expected to fall and can be taken as: “we are 95% confident that the actual future value will fall within this range."

Predictions creates a new Metric for the results, containing the Dimensions of the input Metric and an additional Prediction Type Dimension, which includes the Items Actuals, Baseline, Upper, and Lower.

How do forecasting models work behind the scenes?

Most forecasting models—even though they use different mathematical techniques—tend to follow the same logic under the hood:

1. Detect trend: They identify the underlying direction of the data over time. “Are sales generally increasing? Flat? Declining?”

2. Detect seasonality: They look for repeating patterns over time (daily, weekly, monthly, yearly). “Do sales always spike in December? Are Mondays usually slower than Fridays?”

3. Detect impact of external drivers: If you provide external factors (price, weather, promotions), they model how they influence the forecast. “How much does a price drop affect sales? Does a holiday increase demand?”

4. Use patterns to predict the future: When trends, seasonality, and external factors are understood, the model projects them forward to make future predictions.

Prerequisites

Access rights

Access rights for predictions are based on the permissions set on the source Metric:

• Read access determines the data used for prediction. Forecasts are based only on the data you’re allowed to view, and any prediction results you save are limited to that same visible scope.
• Write access is required to save predicted results. To copy predictions into the source Metric, you must have write access on the target Metric. If not, the operation fails with an error.

How to run Predictions

You can access Predictions from two places:

1. From the Predict button:

   • Find and select the button shown below in the top toolbar of eligible Metrics:

     

   • From the panel that opens on the right, select + New.

2. From the Application's Settings:

   • In the sidebar, select the three dots at the bottom with the More options tooltip.
   • Select All settings from the dialog that opens.
   • Select Predictions, then Start a Prediction.

After accessing Predictions, a dialog opens with the below sections.

Define Data Source

This Metric provides the historical data you want to forecast into the future. Click in this field to bring up a dropdown of eligible Metrics. Metrics must have the following to be eligible:

• a data type of Number or Integer
• a native Calendar Dimension (Year, Quarter, Month, Week, Day)
• empty values at the end of the Calendar—the period to predict for

   Note

  Metrics with the Prediction Type Dimension are ineligible.

Configure Predictions

1. Define the following Settings for your Predictions:
   • Predict until: the time period for which you want to predict.
   • Model: the choice of algorithm used to generate predictions. See this section for more information.
   • (Optional) External factors: data outside the source Metric that may influence its future values. See Use External Factors in Your Predictions for more information.

2. Set the following Advanced options to optimize your Predictions:
   • Fill blank values with: replace missing values in the source Metric with one of the following:
     • Previous non-blank: 'forward filling' for when the last known value is a reasonable estimate for the missing period, such as stable pricing or inventory levels.
     • Zero: for when missing values represent no activity, such as where there are no sales during a period.
   • Predict only positive values: for when the predicted Metric shouldn’t be negative, such as sales, revenue, demand. This ensures that the prediction output will only contain positive values.
   • Predict only integer values: for when the predicted Metric should be whole numbers, such as units sold, headcount. This ensures that the prediction output is rounded to the nearest integer.

What forecasting models does Predictions support?

Prophet

Prophet is an open-source forecasting model developed by Meta, designed to handle seasonality, trends, and holiday effects. For more information, see the official Meta documentation. By default, we recommend using Prophet, as it handles a wide range of seasonal patterns and trend shifts effectively, making it a strong choice for most business forecasting scenarios.

Auto ETS

Auto ETS is an open-source model based on the same logic as the native FORECAST_ETS function. It automatically selects the best-fitting ETS model for your data. It optimizes the smoothing parameters—alpha (level), beta (trend), and gamma (seasonality)—to minimize forecasting error, without requiring any manual tuning.

Seasonal Differencing

Seasonal Differencing is a built-in model that removes repeating seasonal patterns by calculating the difference between each value and its equivalent from the previous season (e.g. this month versus the same month last year), then applies a moving average to smooth the result.

Review and save Predictions

Predictions creates a new Metric for the results, containing the Dimensions of the input Metric and an additional Prediction Type Dimension, which includes the Items Actuals, Baseline, Upper, and Lower.

It also populates a panel on the right of the pane that summarizes the inputs and settings chosen. If you added external factors, the panel includes a report describing their contribution. See Use External Factors in your Predictions for more information.

The temporary Metric disappears after one day. Before it expires, you have the following options:

1. Do nothing. It disappears after one day.
2. Save it in a Metric for inclusion in your model. It becomes permanent. You can either create a new Metric or save Predictions results in an existing Metric with the same Dimensions.
3. Copy and paste the results into the source Metric. The temporary prediction Metric still disappears after one day if it isn’t saved.

Best practices

To get the most accurate and reliable results from Predictions, follow these best practices:

• Keep your data clean and consistent: Predictions accuracy depends heavily on input quality. Check for missing values, outliers, and formatting issues before running Predictions.

• You need sufficient historical data: Ideally you have four years of consistent, monthly historical data at granular level for each time series. However you may be able to improve forecast accuracy with as little as two years’ data.

• External data must be high-quality: Predictions perform best when external factors are included. This data needs to be relevant, complete and predictable. See full eligibility information here.

• Use backtesting to evaluate models and external factors: During testing, compare the performance of different models and external variables using backtesting Metrics. In production, keep only the configurations that consistently perform best.

   Note

  You can explore the “Statistical Forecasting” templates in the Template Library to see a practical example of how to run backtesting and automatically select the best-performing model.

• Predict at the right level of granularity: Make sure the Dimensions used in your source Metric reflect the level at which decisions are made (such as Product x Region).

• Regularly review model performance: Forecasting is not one-and-done. Reassess accuracy over time and re-tune model settings or inputs as your business conditions evolve.

• Review baseline forecasts and adjust when needed: Use Predictions as a start point and apply manual adjustments where business context or recent events aren’t reflected in the data.

Manage Predictions

Open the Predictions tool through Settings (see (2) above for guidance) to manage all your Application’s Predictions. Here you can sort and filter temporary predictions by:

• creation date
• prediction model used
• source Metric used

Use the results to re-open predictions and save, copy or delete, as described above.

Limitations

• Not all time series are predictable: Statistical and machine learning models might fail to accurately capture highly volatile or irregular patterns. Pigment's other forecast tools might produce better results in these cases.
• Predictions are based on historical trends: Sudden structural changes (such as new product launches, pricing model changes, market disruptions) are not automatically captured unless reflected in the data.
• Forecasting at high granularity increases compute time: Running predictions on a large number of Dimension combinations (such as tens of thousands of product x region x channel intersections) can impact performance.

Learn more

• Use External Factors in your Predictions 
• Forecasting Approaches in Pigment