Forecast Explainability
Forecast Explainability is only available for Forecasts generated from an AutoPredictor. You can upgrade existing legacy predictors to AutoPredictor. See Upgrading to AutoPredictor.
Forecast Explainability helps you better understand how the attributes in your datasets impact forecasts for specific time series (item and dimension combinations) and time points. Forecast uses a metric called Impact scores to quantify the relative impact of each attribute and determine whether they increase or decrease forecast values.
For example, consider a forecasting scenario where the target is sales
and
there are two related attributes: price
and color
. Forecast may
find that the item’s color has a high impact on sales for certain items, but a negligible
effect for other items. It may also find that a promotion in the summer has a high impact on
sales, but a promotion in the winter has little effect.
To enable Forecast Explainability, your predictor must include at least one of the following: related time series, item metadata, or additional datasets like Holidays and the Weather Index.
To view aggregated Impact scores for all time series and time points in your datasets, use Predictor Explainability instead of Forecast Explainability. See Predictor Explainability.
For a step-by-step guide on Forecast Explainability, see Item-Level Explainability
Topics
Interpreting Impact Scores
Impact scores measure the relative impact attributes have on forecast values. For example, if the ‘price’ attribute has an impact score that is twice as large as the ‘store location’ attribute, you can conclude that the price of an item has twice the impact on forecast values than the store location.
Impact scores also provide information on whether attributes increase or decrease forecast values. In the console, this is denoted by the two graphs. Attributes with blue bars increase forecast values, while attributes with red bars decrease forecast values.

It is important to note that Impact scores measure the relative impact of attributes, not the absolute impact. Therefore, Impact scores cannot be used to determine whether particular attributes improve model accuracy. If an attribute has a low Impact score, that does not necessarily mean that it has a low impact on forecast values; it means that it has a lower impact on forecast values than other attributes used by the predictor.
It is possible for all or some impact scores to be zero. This can occur if the features have no impact on forecast values, the AutoPredictor used only a non-ML algorithm, or you did not provide related time series or item metadata.
For Forecast Explainability, Impact scores come in two forms: Normalized impact scores and Raw impact scores. Raw impact scores are based on Shapley values and are not scaled or bounded. Normalized impact scores scale the raw scores to a value between -1 and 1.
Raw impact scores are useful for combining and comparing scores across different Explainability resources. For example, if your predictor contains over 50 time series or over 500 time points, you can create multiple Forecast Explainability resources to cover a greater combined number of time series or time points, and directly compare raw impact scores for attributes. However, raw impact scores for Forecast Explainability resources from different forecasts are not directly comparable.
When viewing Impact scores in the console, you will only see Normalized impact scores. Exporting Explainability will provide you with both raw and normalized scores.
Creating Forecast Explainability
With Forecast Explainability, you can explore how attributes are impacting forecast values for specific time series at specific time points. After specifying time series and time points, Amazon Forecast calculates Impact scores for only those specific time series and time points.
You can enable Forecast Explainability for a predictor using the Software Development Kit (SDK) or the Amazon Forecast console. When using the SDK, use the CreateExplainability operation.
Specifying time series
A time series is a combination of the item (item_id) and all dimensions in your datasets
When you specify time series (item and dimension combinations) for Forecast Explainability, Amazon Forecast calculates Impact scores for attributes for only those specific time series.
To specify a list of time series, upload a CSV file identifying the time series by their item_id and dimension values to an S3 bucket. You can specify up to 50 time series. You must also define the attributes and attribute types of the time series in a schema.
For example, a retailer may want to know how a promotion impacts sales for a
specific item (item_id
) at a specific store location
(store_location
). In this use case, you would specify the time
series that is the combination of item_id and store_location.
The following CSV file selects the following five time series:
-
Item_id: 001, store_location: Seattle
-
Item_id: 001, store_location: New York
-
Item_id: 002, store_location: Seattle
-
Item_id: 002, store_location: New York
-
Item_id: 003, store_location: Denver
001, Seattle 001, New York 002, Seattle 002, New York 003, Denver
The schema defines the first column as item_id
and the second column
as store_location
.
You can specify time series using the Forecast console or the Forecast Software Development Kit (SDK).
Specifying time points
If you do not specify time points ("TimePointGranularity":
"ALL"
), Amazon Forecast will consider the entire forecast horizon when
calculating Impact scores.
When you specify time points for Forecast Explainability, Amazon Forecast calculates Impact scores for attributes for that specific time range. You can specify up to 500 consecutive time points within the forecast horizon.
For example, a retailer may want to know how their attributes impact sales during the winter. In this use case, they would specify the time points to span only the winter period in the forecast horizon.
You can specify time points using the Forecast console or the Forecast Software Development Kit (SDK).
Visualizing Forecast Explainability
When creating Forecast Explainability in the console, Forecast automatically
visualizes your Impact scores. When creating Forecast Explainability with the CreateExplainability operation, set
EnableVisualization
to "true" and impact scores for that Explainability
resource will be visualized within the console.
Impact score visualizations last for 30 days from the date of Explainability creation. To recreate the visualization, create a new Forecast Explainability.
Exporting Forecast Explainability
Export files can directly return information from the Dataset Import. This makes the files vulnerable to CSV injection if the imported data contains formulas or commands. For this reason, exported files can prompt security warnings. To avoid malicious activity, disable links and macros when reading exported files.
Forecast enables you to export a CSV file of Impact scores to an S3 location.
The export contains raw and normalized impact scores for the specified time series, as well as normalized aggregated impact scores for all specified time series and all specified time points. If you didn’t specify time points, the impact scores are already aggregated for all time points in your forecast horizon.

You can export Forecast Explainability using the Amazon Forecast Software Development Kit (SDK) and the Amazon Forecast console.
Restrictions and best practices
Consider the following restrictions and best practices when working with Forecast Explainability.
-
Forecast Explainability is only available to Forecasts generated from AutoPredictor - You cannot enable Forecast Explainability for Forecasts generated from legacy predictors (AutoML or manual selection). See Upgrading to AutoPredictor.
Explainability requires attributes - Your predictor must include at least one of the following: related time series, item metadata, Holidays, or the Weather Index.
-
Impact scores of zero denote no impact - If one or more attribute has an impact score of zero, then these attributes have no significant impact on forecast values. Scores can also be zero if the AutoPredictor used only a non-ML algorithm, or you did not provide related time series or item metadata.
-
Specify a maximum of 50 time series - You can specify up to 50 time series per Forecast Explainability.
-
Specify a maximum of 500 time points - You can specify up to 500 consecutive time points per Forecast Explainability.
-
Forecast also calculates some aggregated Impact scores - Forecast will also provide aggregated impact scores for the specified time series and time points.
-
Create multiple Forecast Explainability resources for a single Forecast - If you want impact scores for more than 50 time series or 500 time points, you can create Explainability resources in batches to span a larger range.
-
Compare Raw impact scores across different Forecast Explainability resources - Raw impact scores can be directly compared across Explainability resources from the same forecast.
-
Forecast Explainability visualizations are available for 30 days after creation - To view the visualization after 30 days, create a new Forecast Explainability with the same configuration.