One AI Forecasting in Storyboards

Embedded Forecasting allows you to extend time series trend charts into the future on Storyboards

Forecasting is a powerful way to guide decision making by providing a visualization of how metrics are trending into the future in real time. One Model includes a robust forecasting framework powered by One AI that makes highly configurable forecasts available directly in Storyboards.

Overview

Available in line and spline charts where time is on the pivoted dimensions of the query, forecasting can be run at yearly, quarterly, monthly or daily levels of granularity.  When a forecast is run, it extends an existing trend beyond the latest time period.  A confidence range is also displayed in a lighter shade of the trend line color.  The confidence range is the range in which the forecasting algorithm is confident the number will fall based on the defined confidence level, for which the default setting is 95%.  

Forecasting can be run on demand when viewing a Storyboard or can be configured to run every time a Storyboard is loaded, provided the user has the appropriate application access.  More advanced users can also apply various configuration options to customize the forecast.

Permissioning Forecasting

Any user, regardless of their permissions, can view a forecast configured to run automatically by a user possessing the permission to perform this configuration. All users are also able to disable a forecast in Storyboard view mode using the Turn off Forecast option in the tile menu and re-enable it using the Forecast option in the same menu.

The ability to run or configure new forecasts, either on demand or in Storyboard design mode, requires that the user is assigned to an Application Access Role with the CanEnableOneAIForecast permission checked.

Forecasting on Demand

The images below highlight how easily you can run a forecast for a chart and see the results projected onto the chart in real time.  The options to run or turn off a forecast are found by clicking the light bulb icon .  If the light bulb icon is not present on the tile, either you don’t have the application access necessary for forecasts or the data contained in the tile does not support a forecast.


1.png

 
2.png


In the spirit of providing full transparency whenever possible, explanations and supporting data are available when forecasts are run.  

Clicking on the information icon in the tile header dynamically generates a description of the forecast type(s) run.  If the forecast was run with the default "Let One AI Decide" configuration, both the Curve Fit and ARIMA are attempted and thus both are explained.  If a specific forecast type was selected in design mode (more on this later in this article), only that forecast type is explained.
 
3.png

Clicking on a forecasted data point in a chart displays supporting data related to the forecast configuration and the specific point selected.

13.png

 

5.PNG

In the image above, “confidence interval” refers to the shaded area on the chart and the "chosen forecaster" is automatically selected by One AI based on the lowest (best) AIC score of each forecaster tested.

Automatic Forecasting on Storyboard Loading

Storyboard designers can enable forecasting on charts as the default behavior.  This means that whenever a particular Storyboard is opened the chart will automatically include a forecast.  The option to enable this feature is available in the Tile Settings on the Discover tab.

6.png

Forecast Configuration

A variety of configuration options for forecasting are also available in Storyboard design mode in the Tile Settings on the Discover tab.  There are several levels of configurability available, progressing in order from more basic (letting One AI choose) to advanced (manual configuration).  These options will be explained in the following sections.

Note that tooltips are available for all forecaster settings and can be found by hovering over the question icons .  For this reason, the individual settings will not be explained in this article.

10.png

General Settings

General Settings apply to any forecast, regardless of the forecaster type selected.  The default values are auto-populated where applicable.

7.png

Forecaster Type

The default behavior when a forecast is run is to Let One AI Decide.  This means One AI tests an Autoregressive Integrated Moving Average (ARIMA) model as well as a Curve Fit.  It scores both using Akaike information criterion (AIC) scores and chooses the method with the best (lowest) score.  The AIC measurement was chosen to better balance model complexity and fit.

A specific forecaster type can be selected rather than allowing One AI to automatically select the best fit.

8.png 


A forecaster type can be manually selected while still allowing One AI to choose the best settings for that forecaster based on the data in the trend.

Forecaster Custom Settings

You can manually configure forecaster settings if you do not wish to rely on One AI to automatically choose settings based on the data.  This is done by toggling on the Use Custom Settings option for any forecaster other than "Let One AI Decide".  The custom settings available differ based on the forecaster selected.  As with general settings, the default values are auto-populated where applicable.

9.png

Important Notes About Forecasting

  • Incomplete time periods: If the latest time period in your chart is incomplete (e.g. if your chart is at the yearly level and you are part way through the way), the forecast algorithm will automatically exclude the latest data point and generate a new forecasted data point using the complete historical trend data to improve forecast performance.
  • Data structure: Forecasting requires more than 3 data points and the time periods need to be even and consistent. For example, a trend containing a mix of months and quarters will not allow forecasting. A continuous time selection is also highly recommended. Gaps in the trend will negatively impact the accuracy of the forecast. If you include future-dated time periods, these will be excluded by the forecasting algorithm.
  • Data considered: Only data presented in the tile is considered for the forecast. Forecasting does not look outside of the bounds of the query.
  • Performance: The Prophet forecaster takes significantly longer (up to minutes) to run than the other options, especially with the default configuration selected. Applying manual configuration selections or expanding the amount of data contained in the trend can improve performance of Prophet forecasts.

Forecaster Explanations

Curve Fit Forecast

A Curve Fit forecast entails fitting a line or a curve to the plotted points in the trend and extending the trend forward. This is the simplest form of forecasting and does not reflect seasonality. The advantage however is that it can be applied to a minimal number of data points.

By default, One AI uses a linear fit. A custom setting of quadratic or decaying exponential can also be configured in the tile settings in Storyboard design mode.

  • The linear equation is best for data that increases or decreases steadily with time.
  • The quadratic equation is most useful for data that changes direction over time, either decreasing then increasing or the inverse.
  • The decaying exponential curve works best for data that rapidly decreases with time.

Autoregressive Integrated Moving Average (ARIMA) Forecast

An ARIMA forecast predicts future values based on past values. It makes use of lagged moving averages to smooth time series data. An ARIMA forecast works best with more (12+) data points and is able to reflect seasonality to a certain degree.

The "AR" (autoregressive) part of ARIMA indicates that the metric is regressed on its prior values. "I" (integrated) indicates that the values have been replaced with the difference between themselves and the previous values. This process is referred to as differencing, and may have been performed more than once. Finally, "MA" (moving average) uses the dependency between a value and a residual error from a moving average model applied to past values. The combination of each of these components attempts to make the model fit the data as well as possible.

Prophet Forecast

According to Meta (Prophet's developer) "Prophet is a procedure for forecasting time series data based on an additive model where non-linear trends are fit with yearly, weekly, and daily seasonality, plus holiday effects. It works best with time series that have strong seasonal effects and several seasons of historical data. Prophet is robust to missing data and shifts in the trend, and typically handles outliers well."

Please note that Prophet can take a long time to return results, especially when forecasting on small amounts of data.

Learn more about Prophet here: Prophet

 

Was this article helpful?

0 out of 0 found this helpful

Comments

0 comments

Please sign in to leave a comment.