Forecast Accelerator 1.1.0

1.1.0

May 12, 2025

Inputs, calculations, and parameters associated with the Store feature, which originates from the Multi Factor Elasticity concept, have been removed. The Store feature has also been removed from the transaction mapping in the definition step.

PFPCS-9043

The export to a Data Source has been made optional, allowing users to disable this step when necessary, such as for testing purposes. This enables scheduling the model to run periodically without performing data export each time.

PFPCS-9137

The learning rate has been incorporated into the hyperparameter tuning process, allowing advanced users to set its value via the TrainingParameters parameter table. This addition aims to improve forecasting accuracy at the product and customer levels, especially when lower learning rates yield better results.

PFPCS-9180

Customer fields for additional sources are now supported, enabling data mapping and computation using customer-related information. These sources can be joined to the main source via product or customer features as join fields.

PFPCS-9199

Lags, differences, and rolling input fields can now be set to zero in the Model Configuration, effectively disabling their creation. This allows users to exclude specific features from the forecast model training without impacting the overall process.

PFPCS-9341

During elasticity computation, the price_var_rate is now recalculated for each price point, along with the differences on price_var_rate. Additionally, during forecast calculation, lags and price rates are adjusted at each step using future values to improve the accuracy of the elasticity analysis.

PFPCS-9376

The periods_since_last_sale value is now correctly updated for forecasted rows, resetting to zero if the previous target exceeded a threshold, or incrementing otherwise. This ensures accurate computation of the metric for future dates.

PFPCS-9173

The calculation fails when a categorical feature, stored as an integer, contains null values; nulls are replaced with the string null, causing a backend error. To avoid this issue, categorical features are stored as text in queries.

PFPCS-9198

The model training process for the M5 forecast encountered an out-of-memory (OOM) error, despite reducing the scope. To address this issue, a set of the Advanced Configuration Options parameters has been introduced.

PFPCS-9203

The restriction limiting feature fields to text or integer types in additional sources has been removed, allowing for broader applicability of feature fields across all data types.

PFPCS-9253