Skip to content
BLACKLAKE
REF-DEMA
Project artifact

ML-powered demand forecasting

A retailer's manual forecasting process caused chronic overstock and stockouts. Buyers spent hours in spreadsheets. We built an ML forecasting system that reduced inventory waste while improving availability.

LatencyReliability
Machine LearningForecastingRetail
Industry
Retail & e-commerce
Timeline
5 months
Executive skim
Three measured signals
Jump to outcomes
Availability improvement
+8%
Fewer stockouts on fast-moving items
Overstock reduction
15%
Less capital tied up in slow-moving inventory
Forecast generation
Automated daily
Replaced hours of manual spreadsheet work
System sketch

Context

A mid-size retailer forecasted demand through spreadsheets, leading to systematic overstock on slow-movers and stockouts on fast-movers.

Constraint

Forecasts had to integrate with existing ERP systems, be explainable to non-technical buyers, and improve both overstock and availability metrics.

Intervention

Built a time-series forecasting system using gradient-boosted models. Added a feature store for consistent training and serving. Created an explainability layer showing drivers behind each forecast.

Key decisions

  • Time-series models with seasonal adjustment
  • Feature store for training/serving consistency
  • ERP integration for seamless adoption
  • Explainability dashboard for buyers
  • Automated retraining pipeline
  • A/B testing infrastructure

Outcomes

Overstock reduced 15% in first quarter. Stockouts dropped, improving availability by 8%. Buyers adopted the system without manual overrides.

Why it matters

Better forecasts directly reduce working capital and improve sales—measurable P&L impact from applied ML.

Implementation

Practical technology choices that matched the constraints.

PythonXGBoostFeastAirflowBigQueryLookerFastAPI

Discuss a similar system

If this resembles your constraints, share a short description of what you run today and what needs to change.