Pricing in Yandex Lavka

Vsevolod Paramonov, Andrey Shevtsov7 Nov 2025

Problem

Lavka is a grocery delivery service. Its prices used to be set in a Google spreadsheet by one person from commerce. The analytics team now has to set prices per product so that a combined metric of revenue (GMV) and front margin (FM, roughly profit before logistics and marketing) is as high as possible.

The hard part is measuring how demand reacts to price. Real prices rarely change, so a model trained on history cannot separate the effect of price from seasonality or weather.

Example from the article: one product cost 129 rubles in the business-controlled cluster. In the experiment the other three store clusters sold it at 119, 125, 135 or 139 across four weeks.

Solution in brief

A CatBoost model forecasts demand. Its forecasts at a grid of prices give "all else equal" points, and a linear regression on them yields each product's elasticity. A 10-week experiment shifts prices by -2 to +2 grid steps across store clusters to give the model price variation. The next version optimizes prices by gradient steps through a neural network.

Questions this case answers
  • How to estimate price elasticity of demand for thousands of items when historical prices barely vary, and set prices to maximize a combined revenue and margin metric
  • How to design a price experiment that creates enough price variation without hurting commercial metrics when user-level randomization is not used
  • How to estimate price elasticity of demand for thousands of SKUs when historical prices barely change
  • How to automate price setting to maximize a combined GMV and front margin metric

Tools and methods

  • CatBoost demand model — forecasts the log ratio of demand to the previous period.
  • Grid-based elasticity regression — slope of relative demand against relative price.
  • PMI score — finds complements and substitutes.
  • Gradient-based price optimization — neural demand model with frozen weights.
  • Cluster price-shift experiment — balanced weekly price steps.
Continue reading — it's free

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