论文公告：使用扩展 (R, s, Q) 策略和概率模型的补货优化实用方法

In the world of e-commerce, inventory management is a high-stakes balancing act often described as the Inventory Paradox. Carry too much stock, and your capital is locked in storage and liquidation; carry too little, and you face the "silent killer" of retail—stock-outs, where customer intent meets an empty shelf.

在电商世界中，库存管理是一场高风险的平衡行为，常被描述为Inventory Paradox。库存过多，资金被锁定在存储和清仓中；库存过少，则面临零售业的「无声杀手」——缺货，客户意图遇到空荡荡的货架。

Following our previous discussion on the high-level architecture of our inventory optimization system, we are excited to dive into the applied science that powers the engine.

继我们之前关于库存优化系统高层架构的讨论之后，我们很兴奋地深入探讨驱动该引擎的应用科学。

In our recent publication in Nature Scientific Reports, A practical approach to replenishment optimization with extended (R, s, Q) policy and probabilistic models, we describe how Zalando moved beyond traditional "point" forecasts to build a simulation-driven replenishment engine that explicitly optimizes under uncertainty.

在Nature Scientific Reports最近发表的论文中，使用扩展 (R, s, Q) 策略和概率模型的补货优化实用方法，我们描述了 Zalando 如何超越传统的「点」预测，构建了一个模拟驱动的不确定性下明确优化的补货引擎。

The Design: A Unified Optimization Architecture

设计：A Unified Optimization Architecture

The ZEOS Inventory Optimization Tool isn't just a prediction model; it’s a central replenishment engine supported by a suite of probabilistic forecasting components. We combined Discrete Event Simulation (DES) with stochastic optimization to determine replenishment policies that maximize value across an article’s entire lifecycle.

ZEOS Inventory Optimization Tool 不仅仅是一个预测模型；它是一个由一系列概率预测组件支持的中央补货引擎。我们将Discrete Event Simulation (DES)与stochastic optimization结合，以确定在商品整个生命周期中最大化价值的补货策略。

Figure 1: Component view of the replenishment engine. The ZEOS Inventory Optimization Tool consists of the core optimization engine and supporting probabilistic components.

图 1：补货引擎的组件视图。ZEOS Inventory Optimization Tool 由核心优化引擎和支持的概率组件组成。

The system is built around three core pillars:

该系统围绕三个核心支柱构建：

1. The Forecaster (LightGBM): The future is rarely a single number. Instead of predicting a si...