Skip to main content

Inputs

Attribute Stats​

Attribute statistics identify an attribute by attribute_type and attribute_value. They also identify the aggregation level:

  • Dealer, regional, or national.
  • Model or category.

These statistics generally change slowly. Product statistics may include contribution margin and incremental price. Sales statistics may include abnormal days on lot and demand take rates.

Separate scorers could account for contribution margin, turn, and recent dealer behavior. This would allow the selection configuration to control those tradeoffs without boosting demand take rates before ingestion.

Candidate Pruning​

The selection step could prune candidates. This would provide more flexibility but require scoring a larger candidate set. Separate scorers for demand-rate adjustments may provide more benefit.

Inventory Pipeline Modeling Parameters​

The supply model estimates the probability that an inventory vehicle will be available on a given date. Its inputs may include:

  • Current status and days in that status.
  • Production date.
  • Model and dealer identifiers.

A supply matrix consolidates the inventory projection for use by different scoring approaches. The expected production date of the recommended order also affects its projected availability.

Inventory Candidates​

Inventory candidates represent the latest inventory state for a dealer in the candidate format. Scorers use them with the supply matrix.

For OEMs without regular inventory snapshots, candidate selection may need to fetch inventory from the ordering API at runtime. This would require either passing current inventory to candidate selection or providing API credentials. It would also require a library that converts inventory vehicles into candidate attributes.

Attribute Weights​

Attribute weights are model parameters for take-rate scoring. Other scoring approaches, such as choice set or reinforcement learning, would need different parameters.

Outputs

Candidate​

The output is a normalized candidate with its metadata preserved.

The BMW and Mercedes ordering API implementations need elective codes. Some implementations may also need each code's feature type. The caller could join that information outside candidate selection.