TL;DR
Zalando published a Nature Scientific Reports paper detailing ZEOS, a discrete event simulation-based replenishment system combining LightGBM quantile forecasts with extended (R,s,Q) policies, achieving 22% GMV gains and 91% fill rates.
Key Points
- Monte Carlo simulation-driven optimization handles demand, returns, and lead-time uncertainty across 2M articles from 800 merchants over 12 months
- Extended (R,s,Q) policy with lifecycle-aware parameters (Q₀, t_limit) outperforms static industry baselines including tuned (s,S) policies
- Probabilistic forecasting alone drives largest gains; optimizing for 75th percentile cost distribution adds critical tail-risk protection against demand spikes
- 22% GMV improvement, 34% availability gain, 70-80% merchant adoption rate with stable seasonal performance
Why It Matters
This demonstrates how probabilistic modeling and discrete event simulation can bridge inventory theory and production scale. For ML engineers and optimization researchers, it's a concrete case study in building decision-support systems that explicitly handle uncertainty rather than point estimates—directly applicable to supply chain, logistics, and resource allocation problems.
Source: engineering.zalando.com