Optimizing Markdown in Fashion E-Commerce with Machine Learning

Eleanor Loh15 Aug 2022

Problem

A fashion retailer buys stock months ahead and later clears unsold items with discounts ("markdown"). It must pick which products to discount and by how much. Sales are recorded only at the discounts chosen in the past, and deeper discounts usually went to poor sellers. A model trained on this data can conclude that lower prices cause lower sales, and offline accuracy will not show it because the test set has the same blind spot.

Example from the article: ASOS has more than 26.7 million active customers and over 100K products at any time, and its operations team has relied on rule-based markdown.

Solution in brief

Ithax is a supply-side algorithm that needs no demand model: it selects products and sets depths to hit financial targets. Promotheus then adjusts depths with a price elasticity model, limited to depths where offline validation shows low error. A randomized online test against the operations team confirmed the gain.

Questions this case answers
  • How to choose a price or discount per item when outcomes are observed only at the prices the previous policy chose and no price-variation history exists
  • How to validate a demand model offline and restrict its decisions to the region where it is accurate when counterfactual actions are unobservable
  • How to choose which products to discount and how deeply when demand is observed only at past prices?
  • How to start a markdown optimization system when there is no history of price changes?

Tools and methods

  • Ithax — product selection and depths by binary search on stock value and stock depth targets
  • LightGBM — demand forecast with a monotonic constraint on discount depth
  • Time-series cross-validation — offline check that keeps the partial-information problem
  • Feasible region — limit on candidate depths by out-of-sample error per product group
  • Randomized online test with holdout — profit against manual pricing and Ithax
Continue reading — it's free

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