If you've ever watched your carefully planned sprint forecast collapse because the world changed overnight, you know the pain bol.com's forecasting team felt in early 2020. The Dutch-Belgian e-commerce giant—serving over 12 million customers—faced a nightmare scenario: their demand models, built on stable historical patterns, went haywire as COVID-19 reshaped shopping behavior week by week.
What happened next is a masterclass in adapting forecasting systems under pressure.
The Problem: When Your Models Break
Before the pandemic, bol.com relied on traditional time-series modeling to predict demand across total and product levels. Linear models, gradient boosting algorithms—the usual toolkit for organizations that take forecasting seriously. These approaches worked fine when customer behavior followed recognizable patterns.
Then COVID hit. Online shopping spiked for essential and non-essential products alike. Restriction levels swung up and down. Historical sales data became nearly useless for forward-looking estimates. For months, the team couldn't produce forecasts anyone could act on.
The core issue? Their models treated the pandemic as noise rather than signal. There was no way to encode the reality that consumer behavior was being reshaped by restriction levels, public health messaging, and collective anxiety—all factors that shifted the entire baseline of what "normal" looked like.
The Solution: A Severity Index That Tells the Story
Bol.com's team attacked this by building something new: a COVID-19 severity index. This isn't a generic metric. The team studied government restrictions and analyzed how each intervention actually moved their sales data, then built an index calibrated to their specific business dynamics.
The index runs from 0 to 12. Zero represents no pandemic-related restrictions. Twelve represents the strictest measures they encountered. Crucially, some restrictions carried more weight than others in their models. The temporary closure of physical shops, for instance, had a larger effect on online sales than publicly available stringency indices would suggest.
By translating press conference announcements into future-looking index values, they gave their models something meaningful to work with. The severity index became a feature—something that could slot into any of their existing modeling approaches, whether linear regression or gradient boosting.
The Results: 40% Improvement and Stakeholder Trust
The numbers were striking. Adding the severity index delivered up to 40% relative improvement in mean absolute error during the validation period. But beyond the accuracy metrics, something equally important happened: stakeholder trust returned.
When forecasts align with what teams experience on the ground, they get used. Bol.com's approach succeeded partly because the severity index was interpretable—stakeholders could understand what was driving the numbers. They could question the assumptions, push back on scenarios, and ultimately make better decisions because the models weren't black boxes.
For long-term planning, the team went further. They enabled scenario forecasting by extending the severity index based on different possible pandemic trajectories. Stakeholders could request forecasts built on "what if restriction levels stay where they were in April 2020?" or "what if we return to summer 2019 conditions?" The process became collaborative rather than opaque.
Who Should Care
| Role | Why This Matters | Primary Benefit |
|---|---|---|
| Product Managers | Forecasts inform sprint planning and capacity decisions | More reliable backlog sizing under volatile conditions |
| Data Teams | Models need to encode structural shifts, not just trend lines | A replicable framework for building disruption-aware features |
| Operations Leads | Supply chain and logistics depend on accurate demand signals | Scenario planning replaces single-point predictions |
| Executives | Strategic decisions require stakeholder trust in data | Interpretable metrics build organizational confidence |
What to Watch Next
Demand volatility isn't going away. Post-pandemic supply chains, geopolitical disruptions, and AI-driven market shifts mean forecasting will stay hard. The lesson from bol.com isn't that you need a severity index—it's that you need a way to encode structural changes into your planning systems.
The trade-off is upfront investment: building domain-calibrated features takes time and collaboration. The reward is forecasts that hold when conditions shift, not just when they stay stable.
If your team's forecasts feel disconnected from reality, that's your signal. Build something that translates the world your team operates in into terms your models can use.
Key takeaways
- Encoding structural change as a feature beats hoping your models adapt automatically
- Interpretable metrics build stakeholder trust and drive adoption
- Scenario forecasting gives teams agency rather than false precision
- The principle extends beyond pandemics: any major disruption can become a calibrated feature
Frequently asked questions
What is a COVID-19 severity index in demand forecasting?
It's a numerical scale that translates pandemic restriction levels into values that forecasting models can use as input, calibrated to how those restrictions actually affect sales patterns for a specific business.
How did bol.com improve their pandemic forecasts?
By building a custom severity index that encoded restriction levels, they improved forecast accuracy by up to 40% compared to models that treated pandemic effects as unexplained noise.
Can this approach work for non-pandemic disruptions?
Yes. The principle—creating a feature that represents the structural change affecting your business—applies to any major shift like regulatory changes, market entry of competitors, or economic downturns.
What role did stakeholders play in bol.com's improved forecasting?
Stakeholders collaborated closely with the forecasting team to determine future severity index values, bringing domain expertise that pure data analysis couldn't capture.
How did scenario forecasting help bol.com's planning?
By extending the severity index across different hypothetical pandemic trajectories, stakeholders could evaluate forecasts under multiple futures rather than relying on single-point predictions that proved wrong when conditions shifted.