Pre-order optimisation:
Ordering with data instead of gut feeling
How a two-stage model (LightGBM + LSTM) uses trend signals, trade-fair orders and weather data to improve pre-order accuracy by 42% - and then cuts the forecast error by around another 40% after 2 in-season weeks.
This module is part of a six-part model calculation about a fictional mid-sized fashion company (1,800 SKUs, 18 markets). All figures, datasets, model metrics and euro amounts are entirely fictional - they illustrate the methodology and the kind of effect achievable. They are not results of real clients and not a performance promise. What values are achievable on your real data is established only in a pilot project.
01 The problem - ordering blind, 6 months before delivery
Why 35% of jeans end up in the sale
A fashion company with 1,800 SKUs (jeans, t-shirts, shirts, hoodies) must place its pre-orders at production sites 6-8 months before delivery. At that point there are no weather data, no sell-through figures for the coming season - only samples, buyers' intuition and prior-year comparisons.
The result: 35% of merchandise ends up sold at markdown prices (average discount: 42%). At the same time, the stockout rate stands at 8% (published: ~8% revenue loss from out-of-stock, IHL) - and re-production is not an option with 8-12 week lead times. Between overstock and stockout lies the sweet spot that only data can hit.
02 Data foundation - 4 seasons, 1,800 SKUs, 18 markets
What your merchandise management system already knows about your customers
Buying teams must order summer stock in January for August delivery - without knowing whether the summer will be warm, whether wide-leg jeans will continue to boom, or whether a TikTok trend in March will turn everything upside down. Classic pre-ordering is based on prior year ±10%, seasoned with buyers' opinion. That is no longer sufficient.
03 Exploratory analysis - where the pre-order goes wrong
The patterns your buying team recognises - but cannot quantify
Wide-leg jeans have a pre-order deviation of 48% - nearly every other unit is planned incorrectly. The reason: trend-driven items have no stable prior-year baseline. Basic t-shirts, by contrast, deviate by only 12% - the prior-year logic works there. The model must therefore work category-specifically, not with a one-size-fits-all method.
04 Feature engineering - what makes the pre-order better
External signals that already exist 6 months before season start
The key: the model operates in two stages. Stage 1 (pre-order, 6 months out) uses the 12 long-lead signals. Stage 2 (in-season adjustment, from week 2) corrects the pre-order based on the first real sell-through data - recommending redistributions between markets, flash re-orders or early markdowns.
05 Model - two-stage demand prediction
Pre-order recommendation + in-season correction in a single system
The pre-order alone (Stage 1) is already 42% more accurate than the current buying method. But the real leverage comes from Stage 2: after just 2 weeks of sell-through data, the LSTM corrects the forecast to 11.2% WAPE - enabling timely redistribution. A bestseller in Munich that is underperforming in Hamburg? The system detects it in week 2 and reallocates 200 units. The actual improvement depends on data quality and adoption rate within the buying team. The model calculation assumes: first-year results typically land at the lower end of the range - with increasing effect in subsequent years as the data foundation matures and the model is fine-tuned. The 31.8% baseline and the in-season value are assumptions of the model calculation; the relative improvement sits within the published band of 20-50% (McKinsey).
06 Business impact - the pre-order calculation
What better pre-orders mean in euros
| Category | Units/year | Overstock current | Overstock with model | Savings | Lever |
|---|---|---|---|---|---|
| Jeans Wide Leg (trend) | 320,000 | 48% | 38% | €0.95M | Trend signals via social |
| T-Shirt Graphic | 256,000 | 42% | 33% | €0.60M | Pinterest save rate |
| Hoodie | 384,000 | 31% | 24% | €0.55M | Long-range weather forecast |
| Jeans Slim Fit (NOS) | 704,000 | 18% | 14% | €0.90M | Carry-over stability |
| T-Shirt Basic (NOS) | 576,000 | 12% | 9% | €0.45M | In-season redistribution |
Top 5 categories = €3.45M of €3.84M; the rest is spread across the remaining categories. Savings per category include both the overstock and the stockout share.
The result shown of €3.84M is a model calculation: 224,000 avoided markdown units × €12 net + 96,000 recovered stockout units × €48 contribution margin × 25%. The assumption: in the first year, typically around half of the potential is achievable, with increasing effect in subsequent years. The greatest leverage lies in trend categories (wide leg, graphic tees): planning accuracy is worst there today, and external signals deliver the most added value.
07 Next steps for your brand
From analysis to implementation
Run your last 4 seasons through the model. Would the model have delivered a more accurate pre-order? Backtesting with real figures.
Next season: pre-order for 100 focus SKUs with model recommendation vs. 100 SKUs without. A/B comparison at season end.
From week 2: live comparison of pre-order vs. sell-through. Automated recommendations for redistributions, flash re-orders and early markdowns.