← Back to Case Studies
RetailNordMartForecastingRetailInventory OptimizationMachine Learning

Demand forecasting system that reduced inventory costs by 28%

Key results

Forecast error reduction

38% → 21% MAPE (44% improvement)

Stockout reduction

−67% during promotional periods

Excess inventory reduction

−29% carrying costs

Annual cost improvement

€2.8M estimated benefit

System adoption

100% of planners within 3 months

Problem

NordMart operated 47 retail locations across three countries, with an SKU catalog of approximately 12,000 active products. Their demand planning process relied on simple moving averages supplemented by manual planner adjustments, producing forecast errors that averaged 38% MAPE across the product catalog.

The consequences were significant: persistent stockouts on fast-moving SKUs during promotional periods (estimated at €2.3M in lost sales annually) and systematic overstock on slow-moving categories (carrying costs of €1.8M annually). Total inventory-related cost exposure was estimated at €4.1M per year.

The planning team spent approximately 60% of their time on manual data reconciliation and exception handling, leaving little capacity for strategic decisions.

Approach

The project began with a comprehensive audit of the existing data infrastructure. Key issues identified included: inconsistent promotional tagging, gaps in point-of-sale data for certain store formats, and the absence of external signals that had measurable impact on demand patterns.

The forecasting system was rebuilt around a hierarchical ensemble model combining:

  • SARIMA components to capture seasonal and trend structure at the category level
  • Gradient boosting models trained on engineered features including promotional flags, price elasticity proxies, competitive signals and weather variables
  • A hierarchical reconciliation layer to ensure consistency between store, regional and national level forecasts

Particular attention was paid to uncertainty quantification: the system produces prediction intervals at the 80th and 95th confidence levels, enabling planners to set safety stock levels analytically rather than by rule of thumb.

The system was deployed into NordMart's existing ERP through an API layer, with a planning interface designed specifically for the demand planning team.

Solution & Delivery

The delivered system included a full automated forecasting pipeline running daily refreshes, a web-based planning interface showing forecasts with uncertainty ranges for each store-SKU combination, automatic exception flagging for unusual patterns or data quality issues, and a model performance dashboard tracking accuracy metrics over time.

Training sessions were conducted with the planning team to build confidence in the system's outputs and establish protocols for when and how to apply manual overrides.