Using tags to improve inventory forecasting accuracy - Optimization Software

Using tags to improve inventory forecasting accuracy












Home » Resources » Here

Update May 2016: The notion of tags has been replaced by categories and hierarchies which are more powerful and yield better accuracy in practice. See forecasting with categories and a hierarchy.

By leveraging correlations that exist within the historical data, Lokad can deliver more accurate forecasts. However, when the history is short, it is difficult to detect such correlations just by comparing time-series, because there is little information available. Thus, Lokad introduces the notion of tags. Those tags are used to represent attributes such as families, sub-families or categories, and they are leveraged to strengthen the analysis of correlations.

Better correlations with tags

From the business viewpoint, Lokad delivers optimized inventory decisions; however, from a more technical viewpoint, Lokad actually delivers time-series forecasts. In order to improve those forecasts, Lokad extensively leverage correlations between time-series. Intuitively, taken individually, many time-series don’t contain much information (because the time-series is too short or too erratic); however, when looking at many time-series, it becomes possible to extract more robust indicators concerning seasonality for example.

However, when time-series are short, for example when looking at a product launched less than 3 months ago, the correlation process itself is complicated because there is little data available to reliably correlate the new product to any older product.

In such situations, it is interesting to leverage a priori knowledge about products such as their families, categories or other similar properties that reflect a proximity perceived by customers. Tags are a feature of Lokad precisely intended to offer the possibility to reflect such information in the input dataset.

Intuitively, if Lokad detects that all items associated to a given tag shares the same seasonality, then any new item associated with this tag will be forecast following the same seasonal pattern. The seasonal pattern comes immediately into effect, without having to wait many months to become able to correlate the seasonality of the item with other known seasonalities.

Item descriptors represented as tags

Concretely, a tag is nothing but a special column within the Lokad_Items file (see the file format of Lokad) whose name starts with the prefix Tag. The tag column can contain any value. All items associated with the same tag value are treated by Salescast as belonging to an implicit group defined by the tag value itself.

The tag should reflect some kind of categorization of the items (*) known a priori. Most companies organize the items being sold or produced in hierarchies, and those hierarchies are excellent candidates for tags.

(*) In the Lokad terminology, an "item" is associated to a single time-series. Depending on your business context, it can represent a product, a SKU, a barcode, etc.

For example, let’s consider a retailer that has organized its catalog with 3 nested hierarchical levels: categories, families and sub-families. Then, this hierarchy can be converted into 3 tags respectively assigned to categories, families and sub-families.

Then, tags can also be used to represent domain-specific attributes, for example:
  • Color, size and fabric represent possible tags for clothing retailer.
  • Author, collection and format represent possible tags for book retailer.

Tips for tags

In this section, we cover a couple of tips to more efficiently leverage tags in Lokad.

Don’t insert a tag to reflect the expected seasonality

Tags should reflect some kind of non-statistical information available about the item. While it might be tempting to add a tag that reflects the supposed seasonality - or any similar statistical pattern such as trend for example - we suggest to avoid those type of tags as they don’t deliver true extra information to Lokad. Lokad is capable of seeking seasonal correlations by itself.

Each tag value must be occurring multiple times

A tag is only useful for Lokad if it covers multiple items. If a tag value applies only to a single item, then there is no correlation possible based on this tag. In particular, item identifiers are typically poor candidate for tags, precisely because the identifier is unique of each item.

Don’t include more than half a dozen of tags

In theory, Lokad can process up to 100 tags (*), however, in practice we strongly recommend not using more than half a dozen tags. It’s better to have 3 tags that carry some truly valuable information as opposed to have 10 tags but with a lot of noise involved.

(*) The limit should be understood as if there are more than 100 tags, Lokad will fail at processing your data.