Markdown optimisation:
The right discount at the right time
How a reinforcement-learning agent computes SKU-specific price paths - instead of a blanket 30% for all - and increases overall margin by 4–7%.
01 The problem - blanket 30% from January destroys margin
Why identical discounts for bestsellers and slow movers make no sense
The classic markdown strategy: Full price → 30% mid-season → 50% end-of-season → 70% outlet. Every SKU follows the same cycle, regardless of how well or poorly it sells. A slim-fit jean with 92% sell-through needs no discount - a wide-leg in the wrong colour needs one from week 3.
02 Model - reinforcement learning for SKU-specific price paths
Bestseller (Slim Fit Dark Blue): No discount until end of season, then max. 15%. Trend item (Wide Leg Light Wash): 10% early, then stepped reductions. NOS basics (T-Shirt White): Never more than 20% - loyal customers buy them without a discount anyway. The agent learns that targeted small discounts early preserve more margin than blanket large discounts late. In practice, we recommend a phased rollout: initially the model delivers recommendations that category management reviews. After validation over 1–2 seasons, the degree of automation can be incrementally increased.
03 The full picture - all 6 modules
| Modul | Thema | Technologie | Jährl. Impact | Time-to-Value |
|---|---|---|---|---|
| 1 | Vororder-Optimierung | LightGBM + LSTM (Two-Stage) | €9,8M | 4–6 Wochen |
| 2 | Trend-Radar | CLIP + Social Listening | €3,2M | 3–4 Wochen |
| 3 | Size & Fit Prediction | Collaborative Filtering | €4,8M | 2–3 Wochen |
| 4 | Retouren-Analyse | NLP Topic Modelling | €4,6M | 2–4 Wochen |
| 5 | Kollektionsplanung | Graph-Analyse + Optimization | €4,1M | 3–4 Wochen |
| 6 | Markdown-Optimierung | Reinforcement Learning | €4,2M | 6–8 Wochen |
Even at 50% realisation, that is ~€15M in additional margin - at a company with €230M revenue (3.2M units × €72 RRP), a meaningful margin improvement. The total investment for all 6 modules is €800k–1.5M. ROI: 10–20× in the first year.
04 Next steps
Size recommender (Module 3) + returns quick fixes (Module 4). Deployable immediately with existing data. Impact: €5–8M.
Pre-order model (Module 1) + trend radar (Module 2) + assortment optimisation (Module 5). Impact: +€15–20M.
Markdown optimisation (Module 6) + all modules at full capacity + in-season dashboard. Impact: €25–50M.