Forecasting Future Results

Applicable to Sisense on Linux and Microsoft Windows

Sisense enables you, from a widget's menu, to easily forecast future values based on historical data. The flexible forecast settings allow you to easily change a parameter and see how it affects future values, so you can understand which scenario will help you meet your targets and take the right business decisions.

Forecasting can be applied to the following widget types:

After you have applied forecasts to your widgets, you and your Viewers can view the results and interact with the widgets.

This article includes the following sections:

How it Works

Sisense’s AI engine uses an AutoML-like approach. By default, the AI engine runs an ensemble of algorithms, using AI to assign different weights for each of them to create the most accurate forecasting. If you do not select an Explaining variable, you can also choose which algorithms Sisense will use to calculate the forecast. See below for more detailed explanations of how and when to select the different types of algorithms.

Forecasting Model Types

In Sisense, you can forecast your data based on your previous data and the type of information you want to forecast. There are two types of forecasting models:

  1. Select one forecasted variable over a time dimension (also known as “univariate”). For example, you could forecast future revenues over a certain time period. Sisense creates the forecast based on past data.

    The AI engine driving this type of forecast uses the following models:
  2. Select one forecasted variable and one “explaining” variable over a time dimension (also known as “multivariate”). For example, let’s say you suspect that your buyer age range has a significant effect on your total revenue. You can check that hypothesis by setting "age range" as the Explaining variable, and watching how it affects your predicted revenue. In this way, you can test different factors to see how they affect the predicted behavior of your metric.

    The AI engine driving this type of forecast uses the following models:

    • Random Forest
    • CatBoost
    • ElasticNet

Sisense Cloud and Security

Sisense Forecasting relies on the Sisense Cloud Service for data processing. The historical data that you select is sent to the Sisense Cloud Service where it is processed by machine-learning algorithms and returned as a forecast.

Note: Since data is sent to the cloud service, Forecasting requires that your Sisense server has Internet access. Contact your Customer Service Manager for more information.

For Forecasting to work, make sure that the following addresses are whitelisted on your network:

Sisense views your security very seriously and takes the following steps to limit your data exposure:

Implementing Forecasting

To use forecasting in Sisense, do the following:

  1. Activate forecasting : This is a one-time-only procedure that must be done by your Sisense Administrator.
  2. Add Forecasting to widgets: If the widget is one of the supported types (Line chart, Area chart, Column chart), you can add Forecasting to it.
  3. Customize the time periods: Select how many time periods ahead to forecast.
  4. (Optional) Change the confidence measure: Set the probability of the forecasted value to be located within the error bands, see how it affects the forecast, and find the level that’s right for you.

These steps are discussed in more detail, below.

Activating Forecasting

By default, Forecasting is deactivated in Sisense and a Sisense Administrator must activate it.

To do so:

  1. Click the Admin tab.
  2. From the menu, select Feature Management.
  3. Toggle on Forecast.
  4. Click Save.

This triggers the HTTPS communication between Sisense and the AI engine in the Sisense cloud service.

Adding Forecasts to Widgets

Your Sisense user role determines how you interact with the Forecast feature. Designers add forecasting to widgets that Viewers can then use and customize. Viewers can customize the forecasting definitions to ensure that forecasting is done according to the preferred business logic.

To add forecasting to your widgets:

  1. Create a widget with one of the supported widget types (Line chart, Area chart, Column chart):
  2. In the Data panel, place a time variable on the X-axis and the time-dependent variable that you want to forecast on the Y-axis.
  3. Click the icon under the forecasted variable and select Forecast.

  4. Toggle on the Visible control at the top of the Forecast Settings window to enable the forecast.
  5. The read-only forecasted variable is the time-dependent variable that you selected. If you want to add an additional variable, add one in the Explaining Variable field.
  6. In theEvaluation Period field, select whether you want to use all your past data for the forecast or only a recent period, and select the period.
    Note: The minimum number of data points required for a reliable forecast is 30 data points for univariate and 60 data points for multivariate forecast.
  7. In Forecast Period, indicate how many data points into the future to predict.
  8. Select Advanced for additional options:
    1. In Confidence, select the desired confidence level to set the upper and lower prediction bands in accordance with your business requirements. You can choose between confidence coefficients ranging from 80% to 99%. This percentage indicates the probability of the forecasted value to be located within the error bands.

      In general, the higher the coefficient, the larger the error boundaries, which increases the chance of the predicted value to be included within them. The lower the coefficient, the better you will be able to see the forecasted trend.
    2. Set the Boundaries Limitation to define limits for your forecast so only realistic values are displayed. You can define the following limitations:
      • Value : Select the checkbox and enter a minimum value below which forecasted values are not considered.
      • Value : Select the checkbox and enter a maximum value above which forecasted values are not considered.
      • Numeric values cannot contain a fraction: Select to contain only integers in your forecast.
    3. If an Explaining variable was not selected, in the Type field, Sisense recommends that you leave the default ‘“Ensemble”’ value for the most accurate forecast. Alternatively, you can select the desired model from the list.

      See Forecasting Model Types for more information.

  9. Click Apply. The forecasted period with values and their prediction bands is added to the chart. You can also see a summary of the forecast, see Viewing the Forecast Summary for more information.
    Note: Click the forecast line of the widget to quickly adjust the forecast settings.

“Ignore” Time Periods When Calculating a Forecast

From Sisense for Windows V8.2.5, if you experienced irregular data behavior (such as unusually low sales due to, say, Holidays or COVID-19), you can choose to exclude those time periods from the historical data on which the forecast is based. These excluded time periods are ignored when the forecast is calculated.

Note: You can only exclude time periods from the end of the time series (last x months). You cannot exclude time periods from the beginning or middle of the time series.

  1. To exclude time periods from the data for forecasting:
  2. Edit a widget that displays data over time and that has a forecast applied.
  3. In the Evaluation Period dropdown, select Recent Period.
  4. Set the number of months you want to include in the forecast.
  5. Set the number of “last” months you want to ignore (exclude).
    Note: Your historical data must include at least 30 data points to calculate a forecast. Therefore, if you exclude data points from the evaluation period, make sure you still have at least 30 points overall (as shown in the example, below).

Ensuring Accurate Results

One of the primary concerns of Viewers who leverage forecasting is the accuracy of the forecasted results. Sisense enables you to evaluate the accuracy of the forecast by comparing it to existing data.

By default, the forecast ensembles receive only 85% of the data displayed in the widget. When evaluating forecast accuracy, the remaining 15% (the most recent 15% of the values) are displayed alongside the predicted values for the same period. This way you can compare the performance predicted by the forecasting model against actual performance, see the size of the error, and modify your forecasting settings accordingly.

To evaluate forecast accuracy:

  1. Once the forecasted period is displayed in the widget, click Analyze It.
  2. Select Forecast > Actual vs. Forecast.

    Note: This option appears only after you’ have run the forecast.

    The comparison graph appears.

  3. To close the comparison window, click anywhere outside this window.

Viewing the Forecast Summary

Once the forecast is built, you can see the settings it was built with.

If the selected forecasting model was “Ensemble”, this will show you the statistical outputs and the respective weights of the models in the ensemble to create the most accurate forecasting. If a specific model was selected, its statistical output will be shown.

When an Explaining variable is selected, the relationship score between the target and the Explaining variable will be shown to indicate the accuracy of the choice of the Explaining variable (the closer to 1, the more related those variables are).

To view the forecast summary:

  1. Once the forecasted period is displayed in the widget, click Analyze It.
  2. Select Forecast Forecast Information.

    Note: This option appears only after you’ve run the forecast.
    The Forecast Information panel appears.
    In the example below you can see that the model that was selected was “Ensemble” with 60% of the weight on Holt-Winters and 40% of the weight on Prophet.