Optimizing Regular Prices: Combining Experimental Approaches, ML and Economic Theory

Elizaveta Chefanova, Andrey Shevtsov24 May 2025

Problem

Yandex Lavka (quick commerce, 15,000+ SKUs, 700+ dark stores) wants to reprice automatically, in step with the market, without scaring customers or missing GMV and margin targets. Retail prices are conservative: a product's regular price changes at most about once a month, and often stays put for half a year. A demand model trained on such history sees almost no price variation, so it cannot isolate the price effect. The number of candidate price sets is also huge: 100 SKUs with two options each give 2100.

Example from the article: a product that stood at 100 rubles for half a year says nothing about demand at 95, 99 or 105 rubles.

Solution in brief

Run a deliberate price-randomization experiment so that every product spends equal time at several price steps around the business price. Train a demand model on the enriched data and read local elasticity from it for each concrete situation. Replace the boosting forecast by a linear approximation to compute metrics for many price sets quickly.

Questions this case answers
  • How to estimate price elasticity for many items when observed prices almost never change and the demand model sees no price variation
  • How to run a price-randomization experiment across many items without a net loss in revenue or margin, then choose a price set from a huge combinatorial space
  • How to estimate price elasticity when regular prices change rarely and historical data has almost no price variation
  • How to run a price-randomization experiment in retail without hurting GMV, margin or customer experience

Tools and methods

  • Price randomization across store clusters — per-product switchback with steps from -2 to +2 around the business price
  • CatBoost demand model — target is log of sales relative to the previous period
  • Local elasticity — linear regression over model forecasts at perturbed prices
  • Linear approximation of the model — 12x faster metric computation, leads to linear programming
  • Lagrange multiplier method — maximize GMV plus λ times front margin
Continue reading — it's free

Google or a one-time email code. No password.