Global Forecasting Model Secures First Place in VN2 Inventory Planning Challenge

Global: Winning Solution in VN2 Inventory Planning Challenge

The VN2 Inventory Planning Challenge, a competition that simulates weekly inventory decisions for retail chains with a two‑week product delivery lead time, was won by a team that employed a two‑stage predict‑then‑optimize pipeline. The approach paired a global multi‑horizon forecasting model built with CatBoost gradient‑boosted decision trees and a cost‑aware ordering policy that explicitly trades off shortage and holding costs.

Forecasting Model Architecture

The forecasting component uses a single global model that jointly learns from all available time‑series data. By treating each product’s demand history as part of a larger dataset, the model captures cross‑series patterns that improve accuracy across the board.

Feature Engineering for Stockouts

To handle censored demand during out‑of‑stock periods, the pipeline incorporates stockout‑aware features that signal when observed sales reflect inventory shortages rather than true demand. This helps the model distinguish between genuine low demand and artificial zeros caused by stockouts.

Training Strategy and Global Paradigm

The model is trained using CatBoost’s gradient‑boosted decision trees, with per‑series scaling applied so that learning focuses on temporal patterns instead of absolute sales volumes. Additionally, time‑based observation weights are assigned to give greater importance to recent data, reflecting shifts in consumer behavior.

Cost‑Aware Ordering Policy

In the second stage, projected inventory levels are rolled forward to the start of the delivery week. A target stock level is then calculated by balancing the expected shortage cost against the holding cost, producing an ordering decision that minimizes total expected cost.

Evaluation and Results

When evaluated through the official competition simulation across six rounds, the combined solution achieved first place, outperforming alternatives that used either a standalone forecasting model or a naïve ordering rule.

Potential Extensions

Although designed for the VN2 competition, the authors note that the methodology can be adapted to real‑world retail environments and extended to accommodate additional operational constraints such as supplier capacity limits or dynamic pricing.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.