[Paper Review] One Global Model, Many Behaviors: Stockout-Aware Feature Engineering and Dynamic Scaling for Multi-Horizon Retail Demand Forecasting with a Cost-Aware Ordering Policy (VN2 Winner Report)

Bartosz Szabłowski|arXiv (Cornell University)|Jan 26, 2026

Forecasting Techniques and Applications0 citations

TL;DR

The VN2 winning solution uses a two-stage predict-then-optimize pipeline: a global multi-horizon CatBoost forecast with stockout-aware features, followed by a cost-aware ordering policy.

ABSTRACT

Inventory planning for retail chains requires translating demand forecasts into ordering decisions, including asymmetric shortages and holding costs. The VN2 Inventory Planning Challenge formalizes this setting as a weekly decision-making cycle with a two-week product delivery lead time, where the total cost is defined as the shortage cost plus the holding cost. This report presents the winning VN2 solution: a two-stage predict-then-optimize pipeline that combines a single global multi-horizon forecasting model with a cost-aware ordering policy. The forecasting model is trained in a global paradigm, jointly using all available time series. A gradient-boosted decision tree (GBDT) model implemented in CatBoost is used as the base learner. The model incorporates stockout-aware feature engineering to address censored demand during out-of-stock periods, per-series scaling to focus learning on time-series patterns rather than absolute levels, and time-based observation weights to reflect shifts in demand patterns. In the decision stage, inventory is projected to the start of the delivery week, and a target stock level is calculated that explicitly trades off shortage and holding costs. Evaluated by the official competition simulation in six rounds, the solution achieved first place by combining a strong global forecasting model with a lightweight cost-aware policy. Although developed for the VN2 setting, the proposed approach can be extended to real-world applications and additional operational constraints.

Motivation & Objective

• Motivate inventory planning as translating demand forecasts into cost-aware ordering decisions under stockouts and lead times.
• Propose a scalable, global forecasting approach to handle heterogeneous, intermittent retail time series.
• Translate forecasts into robust, interpretable ordering decisions that trade off shortage and holding costs.
• Show that a global model plus a lightweight policy can outperform per-series tuning in VN2 and be extensible to real-world constraints.

Proposed method

• Stage 1: train a global multi-horizon demand forecaster (three-week horizon) using CatBoost on a tabular representation of all store–product time series.
• Stage 2: apply a cost-aware ordering policy that projects inventory to the delivery week and targets stock levels balancing shortage and holding costs.
• In-stock censored demand is handled via stockout-aware feature engineering and a per-series scaling strategy to emphasize patterns rather than levels.
• Incorporate time-decayed observation weights to adapt to behavioral shifts and recent demand changes.
• Use a direct multi-horizon forecasting approach for h in {1,2,3} to avoid error accumulation and maintain horizon-specific modeling.
• Exhibit a modular, two-stage pipeline that separates forecasting from ordering decisions for scalability and interpretability.

Experimental results

Research questions

• RQ1 How can a single global model learn diverse demand patterns across many time series in a VN2-like setting?
• RQ2 Can stockout censoring be effectively mitigated through feature engineering to improve forecast-driven inventory decisions?
• RQ3 Does a lightweight, cost-aware ordering policy using three-week ahead forecasts yield better total cost performance under lead time and asymmetric costs?
• RQ4 Is the proposed two-stage framework scalable and extensible to additional operational constraints beyond VN2?

Key findings

• The solution won the VN2 competition by combining a strong global forecasting model with a lightweight cost-aware policy.
• Stockout-aware features, per-series scaling, and time-decayed observation weights improve forecast robustness across heterogeneous series.
• The forecasting model uses a direct three-horizon strategy with CatBoost to produce t+1 to t+3 forecasts for all series.
• Per-series scaling reduces dominance of high-volume series and emphasizes temporal patterns.
• The approach is modular and extensible to real-world settings with additional constraints.