Zalando
high-frequency markdown pricing (predict-then-optimize)
Zalando drives the markdowns of several million items through forecasting and multi-objective optimization; validated by 23 A/B tests across 12 markets, the system delivered about 6% more profit at equivalent sales and revenue.
Key points
- High-frequency markdown pricing on millions of items, several times a week.
- In-house gradient-boosted trees forecasting and multi-objective optimization (predict-then-optimize).
- About 6 percent more profit, decision time cut from hours to minutes.
- Evidence level B, validated by 23 A/B tests across 12 markets, confirmed active status.
Objective
Set the right discount level several times a week on millions of items with a limited shelf life, to maximize profit without sacrificing sales or revenue.
The deployment
Zalando, one of Europe's leading fashion e-commerce players, has to re-optimize markdowns on a catalog of several million items, several times a week, with a decision window of a few hours. The system described combines a daily-resolution demand forecast by gradient-boosted trees and a multi-objective optimization framework, following the predict-then-optimize paradigm. It manages the price of around 600,000 items at any given moment across all markets. The new system was validated by 23 A/B tests across 12 markets during the 2023-2024 sales campaigns, then deployed to production.
Results Proof B
Two technical papers by Zalando's teams describing the system in production, with quantified results validated by 23 A/B tests across 12 markets. Figures from internal research and not consolidated into financial results, hence a B level.
How it works
Documented architectureThe stack in detail
- outil Modeles de forecasting gradient-boosted trees (in-house) Daily-resolution demand forecast per item and per price, the foundation of the predict-then-optimize system.
- outil Solveur d'optimisation multi-objectif (in-house) Chooses the markdown levels that trade off profit, sales, and revenue, with Lagrangian decomposition to hold the scale (around 600,000 items managed at any given moment).
- infra Pipeline predict-then-optimize Zalando Production chain that re-optimizes the catalog several times a week within a window of a few hours, validated by 23 A/B tests across 12 markets.
How it runs, concretely
For ops teams-
1Demand forecast AI
Gradient-boosted trees models estimate daily demand per item as a function of price.
-
2Multi-objective optimization AI
An optimization framework chooses the markdown levels that trade off profit, sales, and revenue across millions of items.
-
3Price application AI
The computed markdowns are applied on the site and the app for all relevant markets.
-
4Oversight and A/B testing data team
The pricing team validates changes via A/B tests and monitors the profit impact.
The daily demand forecast per item and per market, which feeds the markdown optimization. Without a reliable forecast, the trade-off between profit, sales, and revenue degrades.
How your customers perceive this type of use
Sourced studiesLe pricing algorithmique est le terrain le plus inflammable : 68% des consommateurs disent se sentir leses quand les marques utilisent le pricing dynamique et 80% jugent plus dignes de confiance les marques aux prix constants (Gartner, 2024). L'equite percue varie selon le secteur : le pricing dynamique n'est juge juste que par 33% a 40% des repondants selon qu'il s'agit de concerts ou de cinemas (YouGov, 17 marches). Le prix personnalise par les donnees individuelles est le plus rejete : 47% des Americains s'y opposent fermement (Consumer Reports, 2024).
Acceptance conditions
- La constance des prix comme signal de confiance : 80% jugent plus fiables les marques aux prix stables (Gartner 2024)
- Le secteur conditionne l'equite percue : le pricing dynamique est mieux tolere pour les cinemas (40% le jugent juste) que pour les concerts (33%) (YouGov 2024)
Red lines
- Le pricing dynamique percu comme abus : 68% se sentent leses (Gartner 2024)
- Le prix individualise a partir des donnees personnelles : 47% d'opposition ferme (Consumer Reports 2024)
- Les frais caches et hausses imprevues, vecus par 79% des consommateurs sur un an et associes a la perte de confiance (Gartner 2024)
Sources: Gartner 2024 · YouGov 2024 · Consumer Reports 2024
How to replicate
Inference, not sourcedData prerequisites
- sales and price history per item
- catalog with product life cycle
- cost and margin structure per item
Org prerequisites
- pricing and data science team
- compute infrastructure to re-optimize the whole catalog in hours
- A/B testing process
Possible stack
- gradient-boosted trees forecasting
- multi-objective optimization solver
- predict-then-optimize pipeline
The plan, step by step
- Step 1Gather the sales/price history per item and the margin structure.Deliverable: Item x day x price dataset with costs and product life cycle.
- Step 2Build the demand forecast per item as a function of price.Deliverable: Forecast backtested on history, errors documented by category.
- Step 3Attach markdown optimization under constraints (profit, sales, revenue) on a sub-catalog.Deliverable: Markdown engine operational on a limited scope.
- Step 4Validate via A/B tests against the current pricing rules, market by market.Deliverable: Profit readout at equivalent sales and revenue vs previous approach.
- Step 5Industrialize: short re-optimization window, monitoring, extension to the full catalog.Deliverable: Production system re-optimizing the catalog several times a week.
First step: Build a demand forecast per item and per price before attaching the markdown optimization to it.
The same pattern, proven in other industries
Airbnb
dynamic price suggestion to sellers (host pricing)
Delta Air Lines
AI dynamic pricing (yield management)
full-funnel AI ad suite (automated ROAS bidding, new-customer prioritization, budgets and targeting managed by the algorithm)
Sources
- S1 High-Frequency Pricing at Scale for E-Commerce Primary archive pending
- S2 Tricks from the Trade for Large-Scale Markdown Pricing: Heuristic Cut Generation for Lagrangian Decomposition Primary archive pending
An error, newer info, a source?
This page lives on its accuracy. If a figure has moved, if the deployment has changed, or if you have a higher-quality source, tell us. Every sourced correction is verified before publication.