Price reduction campaign optimization: product recommendations validated on historical performance

In food retail, price reduction campaigns commit significant budgets and are immediately visible to customers. The pricing department chooses which products to reduce, at national or local level, after supplier negotiations or to reposition against competitors.

It still needs to know what a reduction brings in. Three effects blur the reading of sales.

• Seasonality. A product's sales vary from one week to the next, with or without a price reduction.
• Promotions. An ongoing promotion can inflate or mask the effect of a reduction.
• Cannibalization. What one product gains, a neighboring product can lose.

Without reliable measurement, no one can say which reductions paid off, or choose the next ones based on facts. The need: recommend, product by product, whether to cut the price, to maximize the return on investment of each wave.

We treated the topic as a question of proof before making it a question of models. Three principles structure the approach.

• A common impact measure. Before comparing strategies, you need to know what a reduction actually brought in. This measure arbitrates everything else.
• Back-testing before production. We replay each recommendation strategy on past waves and score it with the same measure. Only the best one goes into production.
• A recommendation that states its level of certainty. Each product receives a confidence score, so the pricing team knows where to focus its decisions.

• Impact measurement. Using loyalty card sales, a measure isolates the effect of each price reduction by neutralizing seasonality, promotions and cannibalization.
• Test bench. We replay several recommendation strategies, including the elasticity-based approach already used internally, on past reduction waves and compare them on this measure.
• Recommendation. The selected strategy applies to ongoing waves, with a confidence score per product.
• Industrialization. The pipeline runs on Google Cloud: Dataform pipelines on BigQuery, orchestrated by Airflow.

A six-month project.

On local waves, the recommendations now apply automatically. They cover nearly a quarter of revenue. Several dozen waves have already been processed.

Elasticity shows the impact: at each wave, we measure the gap in average elasticity between recommended and non-recommended products, which lets the pricing team adjust its decisions based on facts.

• A measure that filters out the noise. We isolate seasonality, promotions and cannibalization to keep only the effect of the reduction.
• Strategies settled on historical data. The selected recommendation proved itself on real waves before being applied.
• A confidence score per product. The pricing team knows when to follow the recommendation and when to arbitrate.
• A pipeline in production. From sales ingestion to recommendation, the pipeline runs without manual intervention.

This approach addresses four situations.

• Your price reduction or promotion campaigns commit significant budgets without a reliable measure of their profitability.
• Your pricing teams make product-by-product decisions under time pressure, after supplier negotiations.
• You have underused loyalty or receipt data.
• You want to compare several approaches, including your current models, before industrializing one.

We start from your sales data, your business rules and your past campaigns. The experts who design the recommendations also put them into production and evolve them with you.