Proof of concept
The first steps consisted of data exploration and collaborative definition of aspects to be considered, such as business objectives, ethical issues in handling demographic data, modeling trade-offs between explainability and predictability.
During the modeling, we identified a stacked 2-component solution as the ideal solution to recommend prices and discounts. First, sales probabilities for each car were predicted using a gradient boosted tree model. In the second step, we identified price-elasticity curves as a suitable and efficient way to display how the probability of a sale depends on the discount offered. The ideal discounts are selected based on these insights and optimization rules.
We included model output explanations (SHAP) in our price recommendations. This was the key to increased confidence in the model and allowed business users to review the recommendations for specific cars when needed.
Roll out
For the initial roll-out, we provided discount recommendations in daily batches, with data exchanged via a secure file-sharing system. After a successful initial roll-out, we helped the customer to host the pipeline in their infrastructure thanks to our flexible docker-based framework.
Later, the customer carried out a complete redesign and fundamental renewal of its data environment. The pipeline was transferred from the legacy environment to the new cloud platform accordingly. Again, a Docker-based solution enabled a flexible and easy transfer to the newly created DevOps pipeline.
Improvement Phase
Our constant monitoring helped to detect data integrity issues in the daily data deliveries as well as edge cases in the model. We further assisted the client in redefining the strategy for rebate components in their systems, which impacted both the input and output of the model.
.png)
.png){kind=small}
.png)
.png)