PlanIQ - New product introduction: all you ever wondered about starting your forecast from scratch

New Product Introduction (NPI) or Cold Start in time series forecasting is a scenario where forecasts should be produced for new items (e.g. new SKU) for which there is limited historical data, or no historical data at all.

The advanced neural networks machine learning algorithms supported by PlanIQ can generate predictions for new items by leveraging metadata attributes. These attributes represent static (non-time related) categorical data that describes the items. The algorithms use this data to tie together items with shared attributes, which in turn allows to leverage similar items’ historical data to produce forecasts (e.g., a clothing company introduces a new pair of blue jeans, and the historical data of other jeans or items with a blue color can support the forecast). Other algorithms could be used in instances where the new items have some (limited) history.

Items with no historical data

In order to perform cold start forecasting with PlanIQ for items with no historical data, use an existing  forecast model configured with following settings:

  • Data collection containing an attributes module and at least 300 historical values (observations) across all time series in the historical data module to meet the neural networks machine learning algorithms data requirements.  
  • Algorithm – either DeepAR+ or CNN-QR.

Once the forecast model has been configured, take the following steps:

  1. Add the new items to the item ID list (click here for more details)
  2. Populate the attributes module with data for the new items. Note that for cold start forecasting to work, the metadata attributes must share some characteristics with preexisting items, as the engine will use those attributes to make inferences about the new items. In our example above, the model would not be able to predict demand for blue jeans if there are no other items with shared attributes and historical data.
  3. Run a forecast action based on the forecast model. The results will include forecasts for the new items.

Items with limited historical data

In order to perform cold start forecasting with PlanIQ for items with limited historical data, we recommend using one of the neural networks machine learning algorithms and relevant metadata attributes as described above. Alternatively, other algorithms could be used as well (ETS, ARIMA or Prophet), but they will not be able to leverage metadata attributes and benefit from the historical data of other similar items.

Once the forecast model has been configured, take the following steps:

  1. Add the new items to the item ID list.
  2. If the neural networks machine learning algorithms are used, populate the attributes module with data for the new items. The metadata attributes must share some characteristics with preexisting items, as the engine will use those attributes to make inferences about the new items. If other algorithms are used, proceed to the next step.
  3. Import the historical data (actuals) of the limited-history items into the historical data module.
  4. Run a forecast action based on the forecast model. The results will include forecasts for the new items.

Example:

The image of the historical data module below shows the last four periods in the time series data. The column ‘Units_Sold’ holds the historical values. The items Blue Shirt, Orange Sweater and Green Jeans are established items present across all time periods, while the items Blue Jeans and Orange Jeans are “cold start” items. Please consider the following:

  • Blue Jeans is a new item with limited historical data. Note that it has non-zero values only for the last two periods.
  • Orange Jeans is a new item with no historical data, and has zero values for all periods.

_image_1_.png

The image of the attributes module below shows that the two new items share attributes with the three already-existing items, and therefore PlanIQ would be able to generate a forecast for those items.

annejulie_1-1622760697844.png

Got feedback on this content? Let us know in the thread below.

Contributing authors: Nitzan Paz, Christophe Keomanivong, Frankie Wolf, Timothy Brennan, Andrew Martin, and Evgenya Kontorovich.

Tagged: