A markdown optimization retail program decides what a flat discount never does: which SKUs to cut, when, and how deep each step goes, against a fixed end date and the stock in each store. Walmart rolled one, built on deep reinforcement learning, out to all U.S. stores from 2019, and the Walmart team reported a 21% higher sell-through rate and 7% lower costs (Chen et al., Interfaces, 2021, DOI 10.1287/inte.2020.1065). At Zara, a controlled field experiment in Belgian and Irish stores (fall-winter 2008) found the new process increased clearance revenues by approximately 6% (Caro and Gallien, Operations Research, 2012, DOI 10.1287/opre.1120.1102). Each model decided per item when and how much, and each was measured against people, not last year.
What a Markdown Optimization Retail Program Actually Decides
A markdown optimization retail program is a model that, for each SKU in each store, chooses when to cut the price, by how much, and in how many steps, so that the stock on hand is sold by a fixed date at the highest margin the demand allows, under rules the merchant sets in advance.
The price moves one way: in the formal setup used by Jia, Li and Ravi (2023), "the seller monotonically decreases the price." There is a clock, the season end or a sell-by date, and a stock count per store that has to reach zero by then. Leftover units have a salvage value, and every price change has a cost; Walmart set out to "reduce the substantial cost to relabel merchandise." A regular price model carries none of this. A promotion is temporary and lifts volume before the price returns; a markdown is permanent and clears stock.
Smith and Achabal (Management Science, 1998, DOI 10.1287/mnsc.44.3.285) derived clearance price policies that "have been tested and implemented at three major retail chains." By Caro and Gallien's account, Zara used "a manual and informal decision-making process for determining price markdowns" until 2007. Walmart moved to deep reinforcement learning in 2019.
This guide is about the price of stock that has to be gone by a date: which items to mark down, when, how deep, and how to know it worked. How the regular shelf price is set in the first place is a different model with a different objective, and we cover it in how SKU-level base pricing is built.
Why a Flat End-of-Season Percentage Burns Margin Twice
An end-of-season markdown applied as one percentage across a category makes two mistakes at once. It is too deep on SKUs that would have sold at full price, which gives margin away, and too shallow or too late on SKUs still on the rail after the date, which leaves stock for a liquidator. That is our reading of the mechanics; Smith and Achabal describe the same trap: "There is little time to react to observed sales, and pricing errors result in either loss of potential revenue or excess inventory to be liquidated."
Manual planning does not scale: decisions multiply as SKUs times stores times weeks. The Zalando authors of a paper on large-scale markdown pricing (Streeck et al., 2024) note that "the 'predict-then-optimize' paradigm is frequently applied, particularly for markdown pricing strategies," and explain why off-the-shelf solvers struggle at their scale.
The Census Bureau's August 2026 release (published October 15, 2026) puts the total business inventories/sales ratio at 1.33, down from 1.40 in August 2025. That figure pools manufacturers, wholesalers and retailers, so it says nothing about your apparel overhang. What markdown optimization retail models see, and a flat percentage does not, is the stock count behind each SKU in each store.
Three Decisions: Which SKUs, When, and How Deep
The first decision is the target. In Oracle's Lifecycle Pricing Optimization documentation, the merchant enters "The sell-through target that must be achieved by the end of the clearance season." As of Oracle's 25.2 documentation, a Hard target that cannot be met under the rules returns no solution. Dataiku's markdown solution has the user pick the metric: "profit, volume, or custom." Someone has to write these down before any percentage exists.
The second decision is timing and step count. Oracle's price ladder "can be a percentage ladder (off full price or current ticket price) or a price-point ladder," and fields such as Min Time Between Markdowns stop the model from cutting every week. The case for starting early comes from a 2026 paper by Bhatia, Gülpınar and Aydın: in numerical experiments on perishable goods, "early and well-timed markdowns based on available inventory levels" in several steps "can simultaneously increase revenue and reduce waste." Model results on food, not a fashion field test, but they argue against waiting for the official sale date.
The third decision is depth, per SKU and store. Walmart describes a policy "tailored to each store," with "demographics-based store personalization" among the benefits.
Margin floors from regular pricing still apply. Markdowns add their own constraints: no price increases, a hard or soft target, a minimum gap between steps, relabeling cost in the objective. Oracle's what-if mode is "used to evaluate potential outcomes and make informed decisions before finalizing the actual optimization run." In practice, markdown optimization retail teams argue about the target date far more than about the discount.
What Published Markdown Systems Reported, and Against What
The four markdown optimization retail systems below share one trait: each was measured against the decisions people were already making. Caro and Gallien write that "Zara is currently using this process worldwide for its markdown decisions during clearance sales." At ASOS, Loh et al. (KDD 2022) report improvements in profitability of 86% for Promotheus and 79% for Ithax, relative to an experienced team's manual strategy in a controlled online test; that is not company profit growth. The Zalando authors, who run their system in production, give no exact figure for their weekly gains.
| Retailer and source | What the model decided | Method | What was reported, and against what |
|---|---|---|---|
| Zara (Caro and Gallien, 2012) | Clearance prices per article | Forecasting model feeding a price optimization model | About 6% higher clearance revenue; field experiment, Belgian and Irish stores, fall-winter 2008 |
| Walmart (Chen et al., 2021) | Markdown price and timing per store, with a clear-by date | Deep reinforcement learning, simulation and optimization | 21% higher sell-through, 7% lower costs; all U.S. stores from 2019, after six months of testing |
| ASOS (Loh et al., 2022) | Markdown depth per product, with a cold-start rule | Supply-side rule (Ithax), then a price response model (Promotheus) | 79% and 86% better profitability than the manual strategy; controlled online test |
| Zalando (Streeck et al., 2024) | Markdown prices across millions of products weekly | Mixed-integer optimization with Lagrangian decomposition and heuristics | Weekly profit gains in the millions of euros; reported by the authors, in production |
Four methods, one test design: the model against the people already deciding.
The Data and Model Layers a Markdown Run Needs
Beyond regular pricing inputs, a markdown model needs stock on hand per store and day, an end date per SKU, unit cost (to maximize margin, not revenue), salvage value, the cost of a price change, and the delivery calendar, since top-ups change starting stock. Price history needs promotion flags.
The model usually has three layers. First, a demand forecast at the current price. Second, the response to further steps: for Freshippo, Hua et al. (KDD 2021) propose "a novel data-driven and interpretable pricing approach for markdowns" that estimates counterfactual demand from observational data. Dataiku trains "a demand forecast model on the historical data by product to test the discounts and estimate price elasticity." Third, the optimizer: mixed-integer programming at Zalando, a multi-period decision process at Freshippo, reinforcement learning at Walmart and, in a field experiment from July 2018 across thousands of items in a markdown scenario, on Tmall (Liu et al.). The choice is engineering, driven by catalog size and data.
Without price history there is no response to estimate, so ASOS ran Ithax first and moved to Promotheus once data existed. A warning of our own: response learned in promotional weeks overstates how shoppers react to a permanent cut, and markdown optimization retail data without promotion flags pushes a first run toward cuts that are too deep.
Measuring a Markdown Pilot: Sell-Through, Margin Kept, Stock Left
Use Walmart's definition of sell-through: units sold during the markdown period divided by the inventory at its start. Next to it, report margin kept against a no-model forecast, stock left after the date, the number of price steps and time to target.
Our rule is a control group. Zara's experiment covered all Belgian and Irish stores; ASOS used a controlled online test. Before-and-after comparisons fail for markdowns, because weather, range and season never repeat. A chain with too few stores to split can split categories that share an end date.
Mixed end dates, online returns, shoppers crossing between test and control stores, and mid-test rule changes all break the measurement. A markdown optimization retail pilot that reports margin without the stock left at the end is reporting half the result.
Where to Start, and How We Approach It
The order we use:
- One category with a hard end date and one channel.
- Sell-through target, margin floor and metric agreed in writing before the first run.
- No price history? A supply-side rule that collects it comes first.
- Control stores or categories from the first season.
- Store-level price response only once the data supports it.
- A what-if before every approval, and no manual override without a log entry.
A markdown optimization retail program fails the same way a clearance sale fails: no date, no stock count, and nobody measuring against anything. The stock you clear now is last season's answer to which products belong on the shelf, and next season starts from that answer, which is where assortment decisions get made.
At DS Stream we build the markdown model on the client's data, in the client's cloud, with a control group planned from the first season. The merchant sets the target and rules; the model works out the steps. It is part of our analytics and machine learning work for retailers. To test it on real stock, bring us one category with a season end date.
FAQ
What does markdown optimization decide, and what stays with the merchant?
A markdown optimization retail program decides, per SKU and store, when the first cut happens, how deep each step is and how many steps there are, so stock is gone by the target date at the highest margin demand allows. The merchant sets the date, target, margin floor and price ladder.
Is markdown optimization the same as regular price optimization?
No. Regular prices serve a goal for the whole assortment and can rise. A markdown moves one way toward a date and a stock count; in Jia, Li and Ravi's words, "the seller monotonically decreases the price." Dynamic pricing moves prices both ways.
When should the first markdown happen?
Usually earlier than the team would like. Smith and Achabal noted there is little time to react to observed sales, and the 2026 perishables paper found in numerical experiments that early, inventory-based steps raised revenue and cut waste together. Derive the date from the target and weeks of stock left.
Can markdown optimization work for products with no price history?
Yes, though not with a price response model on day one. ASOS ran Ithax, a supply-side rule with no demand estimate, to gather data, then moved to Promotheus. In a controlled online test, both beat the manual strategy on profitability, by 79% and 86% relative to it.
