The article discusses the transition from machine learning model metrics to economic outcomes in Customer Value Management (CVM). It highlights that a high-ranking model does not automatically translate to profit, as the cost of incentives and communication must be accounted for.
The author provides a synthetic example where 10,000 customers are selected. With a baseline purchase probability of 0.10 and a probability of 0.18 with an offer, the absolute uplift is 8 percentage points. Given a margin of 300 ₽, a discount of 50 ₽, and a contact cost of 2 ₽, the additional profit is calculated as 130,000 ₽.
The text emphasizes that discounts reduce margins for all customers who make a purchase with the offer, including those who would have bought anyway. This can lead to negative results even with positive uplift, such as in a stress scenario where a discount of 150 ₽ results in a loss of 50,000 ₽.
To measure the true value of a model, the author suggests comparing the new policy's profit against the current policy's profit, minus additional model usage costs. This requires a randomized design where customers are assigned to different policies and results are compared across the entire assigned population, not just the recipients of offers.