Case study · Retail
Forecasts at SKU and store level, every week
Retail replenishment is only as good as the forecast behind it. This is how we built machine learning forecasts at SKU and store level that account for seasonality, promotions and prices, and feed them into replenishment with accuracy tracked against the current method.
Project overview
Machine learning forecasting for replenishment.
- Industry
- Retail
- Client
- Confidential
- Solution
- Demand forecasting
- Platform
- Azure ML + Fabric
- Timeline
- 10–16 weeks
- Team
- 3 people
- Granularity
- SKU × store × week
- Drivers
- Seasonality, promotions, price
- Output
- Replenishment system
Expected business outcomes
What the client set out to achieve
Every model is backtested against the current forecast before it replaces it.
-
10–30%
Lower forecast error
Compared with the current method on the same historical periods.
Target range from backtesting; varies by category
-
5–15%
Less excess stock
Better forecasts reduce the buffer needed for the same availability.
Target range after replenishment uses the forecast
-
Weekly
Automatic retraining
Models retrain on new sales and are monitored for drift.
With alerts when accuracy drops
Business challenges
Where the client started
Forecasts are averages with manual adjustments.
Method
Moving averages or last year plus a percentage.
Promotions
Uplift estimated by buyers from memory.
Stores
Forecasts at chain level, split to stores by a fixed ratio.
Stock
Overstock in some stores, gaps in others.
Trust
Buyers override forecasts because they don't trust them.
Proposed solution
What the ML build covers
Honest backtesting is as important as the model.
-
Feature pipeline
Sales, prices, promotions, calendar and stock-outs prepared weekly.
-
Forecast models
Gradient-boosted models across SKU-store series, with hierarchies.
-
Backtesting
Rolling backtests against the current method by category.
-
Integration
Forecasts delivered to the replenishment or ERP system.
-
Explainability
Drivers shown per forecast so buyers understand changes.
-
Monitoring
Accuracy tracked weekly with drift alerts.
Project requirements
What the client asked us for
The request was better forecasts. The proof is a backtest on the retailer's own history, category by category, before anything changes in replenishment.
Stock-outs are corrected in the training data, otherwise the model learns that demand is low when stock was simply missing.
What the project had to deliver
- 01SKU-store forecasts
- 02Promotions and prices included
- 03Proven by backtesting
- 04Explainable to buyers
Key challenges & considerations
What made it harder, and what we did
-
Stock-outs hiding true demand
What we did
Sales history was corrected for stock-outs before training, so the model learned demand rather than availability.
-
Short history for new products
What we did
New products borrowed the patterns of similar items.
-
Promotion data recorded inconsistently
What we did
Promotion history was cleaned and tagged with buyers.
-
Replenishment systems that accept limited input
What we did
Forecasts were delivered in the format the replenishment system accepted.
What we watched for
- Correct for stock-outs
- Always compare with the current method
- Roll out by category
- Explain forecasts to buyers
Development process
Backtest before you trust
Models earn their place category by category.
-
Data review
Sales, promotions, prices and stock-outs assessed.
-
Baseline
Current method measured as the baseline.
-
Models
Models built and backtested by category.
-
Pilot
Forecasts used for replenishment in pilot categories.
-
Scale
Rolled out where it beats the baseline, with monitoring.
Technology stack
Practical machine learning on Azure
Proven forecasting methods, not experiments.
Machine learning
- Python
- scikit-learn / LightGBM
- Azure Machine Learning
Data
- Microsoft Fabric or Azure SQL
- Feature pipelines
Serving
- Batch scoring
- REST API
- ERP/CRM integration
Monitoring
- Accuracy tracking
- Drift alerts
- Retraining schedule
Quality assurance & deployment
Measured against a test set before launch
AI quality was measured, not judged from a demo. A test set of real examples with agreed answers decided whether the system was ready, and the same set was re-run after every change.
- A test set of real examples with agreed answers
- Accuracy thresholds agreed with the business before launch
- Safety checks: refusals, permissions and prompt-injection tests
- A person reviewing low-confidence output
- Cost and response time measured at real volumes
Deployment
- 01PilotForecasts used for replenishment in pilot categories.
- 02ScaleRolled out where it beats the baseline, with monitoring.
- 03Pilot, then staged rolloutA small group of real users first, with results reviewed weekly before widening access.
- 04Monitoring and warrantyAccuracy, drift, usage and cost tracked after launch, and six months of free bug fixing.
Final deliverables
Forecasts buyers can rely on
Replenishment runs on weekly SKU-store forecasts that account for promotions and seasonality, in the categories where they proved better.
Buyers can see why a forecast changed, so overrides become the exception.
- Weekly SKU-store forecasts
- Backtested accuracy
- Explained drivers
- Monitored for drift
Handed over to the client
- 01Production AI serviceDeployed in the client's own cloud tenant
- 02Source code and modelsIn a repository the client can access
- 03Test set and resultsThe evidence behind the launch decision
- 04MonitoringAccuracy, usage and cost
- 05Six-month warrantyFree bug fixing after delivery
Success metrics
How success is measured
| Measure | Outcome | How it is measured |
|---|---|---|
| Lower forecast error | 10–30% | Target range from backtesting; varies by category |
| Less excess stock | 5–15% | Target range after replenishment uses the forecast |
| Automatic retraining | Weekly | With alerts when accuracy drops |
Verified reviews
What clients say about our custom software work
Verified reviews from our clients on the same kind of work, published on Clutch.
-
Custom Software Development
Web-Based Coupon App Development
The work produced by the team was of exceptional quality.
Jon Thies
CTO, Model Rocket Feb 2022 – Mar 2023
-
Custom Software Development
Mobile & Web Platform Development & Design
Silver Scintilla was responsive, respectful, and enjoyable to work with, quickly addressing any questions or requests.
DeShawn Brown
CEO & Founder, Lithios Jun 2021 – Nov 2022
-
Custom Software Development
Web-Based HRM System Development
Silver Scintilla provided the best solution with scalable functionality, personalized features, and robust security.
Suheb Khan
Director, Collaborative Insight Technologies Jul 2020 – Jun 2022
Free discovery call
Have an idea? Let's turn it into AI-powered software.
Book a free discovery call with our experts. Share your idea and we will help you shape the scope, timeline and budget, under NDA.
- Free consultation
- NDA before we talk
- Transparent estimate