Efficient inventory management is one of the most critical challenges for e-commerce professionals, especially when working with platforms like Shopify or managing supply chains with limited resources. Stockouts, surplus inventory, and inefficient forecasting can lead to costly financial consequences, impacting profitability and customer satisfaction.
Recently, the VN2 Inventory Planning Competition brought together industry experts and data enthusiasts from around the globe to tackle this very problem. This transformative event revealed cutting-edge approaches to using probabilistic forecasts, machine learning, and optimization techniques to streamline inventory ordering processes. In this article, we’ll break down the lessons from VN2’s top-performing solutions, providing actionable insights for e-commerce professionals, supply chain managers, and anyone grappling with inventory challenges.
Understanding the Problem: Inventory Planning in a Real-World Context
At its core, inventory planning is about balancing two opposing goals: minimizing holding costs while avoiding costly stockouts. In VN2, participants were tasked with managing weekly inventory orders for 599 products over six weeks. The challenge required competitors to simulate a real-world scenario by factoring in:
- Lead Times: A two-week delay before orders were received.
- Cost Dynamics: Holding costs were significantly lower than shortage costs, incentivizing careful balancing of inventory.
- Seasonality and Demand Variability: The dataset included seasonal trends and unpredictable demand spikes.
The competition aimed to highlight innovative techniques for inventory management, encouraging participants to not only forecast demand but also optimize their ordering strategies over time.
sbb-itb-499c055
Key Strategies from the VN2 Competition
Below, we explore the standout approaches from the competition’s top solutions, focusing on their methodologies, tools, and insights.
1. Probabilistic Forecasting for Better Decision-Making
Several top teams embraced probabilistic forecasting, a method that predicts the probability distribution of future demand rather than just a single point estimate. This approach allows businesses to account for uncertainty and make more informed decisions about stock levels.
Example Approaches:
- The Frugal Five Team: This group used LightGBM models to estimate quantiles of demand, creating a probabilistic distribution. Their inventory ordering policy considered not just the mean forecast but also the likelihood of extreme demand scenarios.
- Carlo Cavalieri’s Solution: Carlo used a neural network-based model (DeepAR) to generate probabilistic forecasts. By simulating 1,000 potential demand paths, his model minimized both holding and shortage costs, providing a robust ordering strategy.
Key Insight: Probabilistic forecasting shines when dealing with seasonal or uncertain demand, empowering businesses to strike a balance between overstocking and stockouts.
2. Combining Machine Learning with Business Rules
Teams that combined advanced machine learning models with cost-aware policies often outperformed others. The emphasis was on creating adaptable, smart solutions that aligned with real-world cost considerations.
Notable Examples:
- Reinforcement Learning by Matias: Matias deployed deep reinforcement learning to directly optimize long-term profit. Instead of forecasting demand first, his model learned optimal ordering policies by simulating thousands of scenarios within a neural network framework.
- Bartosz Szabowski’s Winning Solution: Bartosz combined CatBoost (a powerful gradient boosting model) with a straightforward, cost-aware inventory policy. His method translated forecasting uncertainty into optimized order quantities, using smart feature engineering to improve model performance.
Key Insight: While machine learning models can handle demand forecasting, integrating business costs and constraints (e.g., shortage penalties) into ordering policies is critical for success.
3. Feature Engineering: The Secret Sauce
Feature engineering emerged as a pivotal factor in the competition. The best-performing solutions didn’t rely solely on the raw data; instead, they crafted tailored features to capture underlying demand patterns.
Techniques Used:
- Behavioral Features: Teams created features to account for trends, seasonality, and product-specific behaviors. For instance, Bartosz scaled features dynamically to account for products with different sales magnitudes.
- Time Decay Weights: Recent demand data was weighted more heavily than older data to reflect changing trends and seasonality.
- Global Seasonality: Instead of relying on product-level seasonality (which can be noisy), teams like Philip and Jakob applied global patterns across all products, achieving more stable results.
Key Insight: Thoughtful feature engineering amplifies the predictive power of machine learning models, especially in diverse datasets with varying demand patterns.
4. Optimization Beyond Forecasting
Forecast accuracy alone didn’t guarantee success. The key was translating forecasts into effective inventory policies. This required balancing short-term and long-term outcomes.
Innovative Approaches:
- Simulation Models: Philip and Jakob simulated inventory scenarios using a combination of LightGBM forecasts and cost-aware policies to find the most effective order quantities.
- Cost-Aware Target Stock Levels: Bartosz’s solution used a formula that dynamically adjusted safety stock based on forecast uncertainty and cost ratios. His model optimized a single parameter, ensuring scalability for larger datasets.
Key Insight: Inventory planning is a holistic problem. Success requires integrating forecasting with cost-aware optimization strategies for robust decision-making.
Lessons for E-Commerce and Supply Chain Professionals
The VN2 competition isn’t just an academic exercise - it offers practical lessons for real-world inventory management. Here are the takeaways from the top-performing approaches:
Key Takeaways:
- Embrace Probabilistic Forecasting: Move beyond point forecasts to models that account for uncertainty. Probabilistic methods provide more actionable insights for inventory decisions.
- Leverage Feature Engineering: Thoughtfully crafted features can significantly enhance model performance. Focus on capturing seasonality, trends, and demand variability.
- Prioritize Cost-Aware Policies: Align your inventory strategy with business objectives by incorporating holding and shortage costs into decision-making.
- Combine Simplicity with Sophistication: Advanced models like reinforcement learning are promising, but even straightforward approaches (e.g., scaling and cost ratios) can achieve outstanding results.
- Test, Validate, Repeat: Rigorous validation and simulation are key to building reliable solutions. Use historical data to stress-test your models before deploying them in production.
- Adapt to Your Context: Solutions like global seasonality or time decay weights might not work universally. Tailor your approach to your business’s unique dynamics.
Final Thoughts
The VN2 Inventory Planning Competition highlighted the immense potential of data-driven decision-making in inventory management. Whether you’re managing a small e-commerce store or a complex supply chain, these lessons provide actionable strategies to improve efficiency and profitability.
As the landscape of supply chain optimization continues to evolve, the integration of machine learning, probabilistic forecasting, and cost-aware policies will define the next generation of inventory management tools. By adopting the principles showcased in VN2, businesses can stay ahead of the competition, reduce waste, and deliver superior customer experiences.
The future of inventory planning is here - are you ready to embrace it?
Source: "VN2 Inventory Planning Competition: Winners Explain Their Solutions" - Nicolas Vandeput, YouTube, Dec 15, 2025 - https://www.youtube.com/watch?v=pypzcvwmApA
