Forecaster Advanced Explainability — time-phased writeback into a Time-dimensioned module: expected

Farouk

Hi all,

I'm using Forecaster's Advanced Explainability export ("Send to Anaplan") and I'd like to confirm the intended behavior and the correct interpretation of the results.

Context

The Advanced Explainability data form requires a Time dimension, and it writes feature values per period into a target module dimensioned as Item ID × Features × Forecast version × Time. This differs from the explainability module documented in Anapedia, which is built without Time (Item ID + Forecast Feature Names in rows, single Feature Contribution line item in %, Summary = None).

The features written per item are: base_value_trend, Historical: linear_trend, Historical: Seasonality_Yearly, Anaplan Postprocessing. The Feature Contribution values are signed and non-normalized (ranging into the thousands of %, with negatives), so they look like a signed decomposition of the predicted level rather than the 0–100% relative-importance scores of the classic (PlanIQ-era) explainability.

Questions

1. Is time-phased explainability writeback the intended behavior of Advanced Explainability, and how should the per-period values be read versus the time-less Feature Contribution % in the standard module?
2. Are these components additive — i.e. should base_value_trend + Historical: linear_trend + Historical: Seasonality_Yearly + Anaplan Postprocessing reconcile back to the Predicted Value for a given item/period?
3. base_value_trend and Historical: linear_trend return identical values across all periods and items. Is this by design (base value = linear-trend component), or does it point to a duplicated source→target mapping on my side?
4. What exactly does the Anaplan Postprocessing feature represent as a decomposition component?

Any confirmation from the product team or from anyone who has set up the Advanced Explainability writeback would be appreciated. Screenshot of the results module attached.

Thanks!

NitzanP

Hi @Farouk.

1. Yes. This is the intended behavior of advanced explainability. The per-period advanced explainability values reflect the approximate contribution of each feature to the predicted value of a given item at a given time. These values should be in a standard numeric format. The standard (time-less) explainability values reflect the impact of each feature relative to other features in regard to a given item across all time periods. These values should be in %.
   
2. Yes. All values combined should add up to the predicted value.
   
3. This behavior is expected in MVLR, but doesn't apply to all algorithms. Note that the advanced explainability functionality will be expanded to more algorithms in the future.
   
4. The postprocessing feature accounts for special scenarios in which Forecaster's logic changes the raw predicted value (i.e. the value generated by the forecast model). For example, Forecaster converts negative predicted values to zero if no negative values are present in the historical time series.

Also, please note:

• We plan to standardize the feature names to make them more user-friendly.
• Anapedia will be updated with detailed advanced explainability content in the coming days.

Best,
Nitzan Paz