Discount Optimization

Challenge

The machine learning solution was developed for a premium automotive manufacturer in a build-to-stock market. Discounts serve as one of the most effective purchase incentives, especially since buyers cannot individually configure the vehicles and their features. At the same time, suboptimal use of discounts is associated with significant direct profit losses. We faced the challenge of optimally distributing already fixed budgets across individual vehicles within specific vehicle groups using machine learning. Functionalities for targeted inventory management needed to be developed. The initial goal was to maximize sales and prevent inventory from exceeding a certain maximum duration.

Approach

To achieve this, we implemented a pipeline that automates the entire data science process—from data preparation and feature building through backtesting, including feature selection and model tuning, to the deployment of the final model as a service. The general approach consists of two interlocking algorithms: First, the sales probability of a vehicle within a freely selectable forecast horizon is predicted. The forecast model provides the input for the optimization step: The sales probability for the entire range of possible discount rates is predicted for each inventory vehicle. The allocation problem can then be solved by translating it into an integer optimization, maximizing the sales probability of all inventory vehicles according to the fixed discount budgets. Individual car groups (e.g., overaged) can be weighted higher in the optimization, allowing them to bind higher discounts.

Result

Our client was enabled to optimally distribute discount budgets based on modeled sales probability patterns to individual vehicles. Additionally, the solution offers the ability to simulate the impact of varying budget levels on the total predicted sales (also in specific product segments). Currently, the pipeline serves automated data-driven inventory management and thus maximizes sales. Looking ahead, the solution can be adapted with minimal effort to minimize or maximize any other quantifiable metric (e.g., margin or revenue).