Europe's largest online fashion retailer rebuilt its sale campaign pricing around a two stage pipeline, and measured about six percent more profit at equivalent sales across 23 A/B tests in 12 markets.
When Zalando's pricing team needs to re-price more than five million fashion items for a sales campaign, the work now takes minutes instead of hours. The shift came from splitting the problem in two: a tabular machine-learning model forecasts what shoppers will want, and a separate optimizer decides what each item should cost, balancing short-term revenue against long-term profitability. In a preprint on arXiv, ten authors describe a system that runs daily across the retailer's twelve European markets and that the authors say delivered roughly six percent more profit at equivalent sales and revenue during the 2023 and 2024 sales seasons.
Zalando, Europe's largest online fashion retailer, framed the work as a response to a pattern familiar to anyone running promotional events for physical goods. Sales campaigns concentrate demand into a few days, demand curves shift hour by hour, and the same shirt may need to be priced differently in Berlin and Paris. The previous approach updated prices on a weekly cadence and relied on a hybrid of algorithmic recommendations and human overrides, a loop that the authors say stretched decision time out to hours and constrained how aggressively the catalog could respond.
The new system replaces that loop with a forecast-then-optimize pipeline. A gradient-boosted ensemble of decision trees, a standard tabular machine-learning method rather than a generative model, produces a daily demand forecast for every article in every market. A separate multi-objective optimizer then sets prices with two goals in mind: long-term profit, the headline financial measure, and net merchandise value, a proxy for the longer-run health of the catalog. The split is deliberate. The authors argue that training one end-to-end model to both predict demand and prescribe prices conflates a prediction problem with a prescription problem, and that separating the two makes each step easier to audit and to retune.
The scale is what makes the architecture interesting. The system prices more than five million articles, runs in twelve markets, and was validated, according to the authors, in 23 A/B tests spread across two sales seasons. The headline outcome, about a six percent lift in profit at flat sales and revenue, is large enough to matter at Zalando's size and small enough to be plausible as an engineering win rather than a market shift. The operational win is the second-order claim: pricing decisions that used to take a team hours now resolve in minutes, freeing the same engineers to run more experiments.
The optimization framing is the part that practitioners will want to interrogate. Maximizing profit and net merchandise value at the same time is not the same thing as maximizing revenue or conversion, and the choice of weights between the two objectives shapes every downstream decision. A reader building on this work would need to know how those weights were set, how they were held constant across markets, and what would happen if a competitor ran a more aggressive sale. The preprint reports the headline result but leaves most of those tuning choices to future work.
A few caveats frame the result. The paper is an arXiv preprint, not peer-reviewed, and the 5M-article scale, the profit uplift, and the superiority over weekly-granularity baselines are author-reported rather than independently audited. Author affiliations are not fully visible in the abstract excerpt, though Tim Januschowski is publicly associated with Zalando. There is no third-party industry analyst or customer quote in the packet to corroborate adoption or production deployment, so the design should be read as a reference architecture, not as a market signal.
What generalizes beyond fashion e-commerce is the pattern rather than the numbers. Daily-granularity demand forecasting, a clean split between a demand model and a price optimizer, and a multi-objective framing that puts long-term catalog health on equal footing with short-term profit are choices that any retailer running frequent promotions could stress-test. The specific six percent figure belongs to Zalando's 2023 and 2024 sales seasons. The lesson is that weekly cadence and human-in-the-loop overrides are a design choice, not a constraint of retail pricing, and that a retailer with enough engineering depth can move the cadence from days to minutes without changing the unit economics of the sale.