Automated Price Prediction for a Large Retail Catalogue
Clearlead rebuilt a global metals retailer's price prediction system from the ground up, delivering a unified model and automated MLOps pipeline that now provides accurate weekly price recommendations across 30,000+ products in over 130 locations across the US, Canada, and the UK.
40% lower error
median prediction error reduced from 10% to 6%
90% of sales
covered by model predictions, up from 38% of the catalogue
130+ locations
across the US, Canada, and UK, updated on a weekly automated schedule
The client
- A global retailer
- Manufacturing, Retail
- United States, Canada, and United Kingdom
Overview
The client is a Canadian-headquartered metals retailer with over 130 franchise locations across the United States, Canada, and the United Kingdom. Their catalogue spans more than 30,000 unique product combinations, with prices that need to reflect current base metal costs, regional logistics, and currency movements.
Clearlead assessed the existing prediction system, identified why it wasn't achieving the coverage and accuracy the business needed, and rebuilt the approach: a unified model trained across the full catalogue, automatic thresholding for product categories where reliable predictions cannot be made, a production AWS SageMaker MLOps pipeline with weekly automated retraining, and an extension to the UK following successful North American deployment.
The system now provides accurate weekly price recommendations across the full product range, with median prediction error well within the target threshold.
Methods
- Machine learning
- Price prediction
- MLOps
- Automated retraining
- Feature engineering
Engagement type
- Design and build
The situation
The client operates a franchise retail network with over 130 locations across the United States, Canada, and the United Kingdom, selling a wide range of metals in small quantities. Their product catalogue spans more than 30,000 unique combinations of material, grade, shape, and unit, and the right price for any given product is not straightforward to determine.
Base metal costs fluctuate, logistics costs vary by region, and currency movements across three countries add another layer of complexity. Many products in the catalogue are purchased infrequently, and in some stores there is little or no recent purchase history for certain product types. When a product has not been sold recently in a given location, the store owner has no reliable in-store reference price to draw on. The model provides a suggested price accounting for current base metal costs, regional patterns, and market movements, and proved more accurate than simply reusing the most recent available sale price even for products that had been purchased recently. Franchise owners also retain the ability to adjust prices for their own commercial reasons, so the system is designed to provide a reliable suggested price rather than a fixed one.
The client engaged Clearlead to assess the existing prediction approach, identify why coverage was so limited, and design and build a replacement capable of serving the full catalogue reliably.
The engagement
Clearlead began with a focused one-week discovery investigation before recommending any changes. The investigation found that the existing segmented approach, which trained separate models for each product segment, couldn't generalise to the full catalogue at this scale: most segments lacked sufficient transaction volume to produce reliable predictions, and the segments that did perform well couldn't draw on pricing patterns from related products. A unified modelling approach, one model trained across all materials, grades, and shapes, was recommended and approved before any development work began.
-
Discovery investigation: a structured one-week assessment of the existing system, establishing why the segmented approach couldn't achieve the coverage and accuracy needed at this scale and complexity. The existing system could only generate predictions for approximately 38% of the product catalogue, with entire material categories producing no predictions at all.
-
Unified model design and feature engineering: a single model trained across the full product catalogue, incorporating rolling temporal price features, product characteristic features, and price factors encoding historical patterns by material, grade, shape, and region. For product categories where the model cannot predict within an acceptable accuracy threshold, predictions are automatically suppressed rather than surfaced with unreliable estimates.
-
Production MLOps pipeline on AWS SageMaker: automated weekly retraining, version-controlled model artefacts, and predictions delivered directly into the client's operational systems with no manual steps in the routine run. Monitoring across three AWS Managed Grafana dashboards covering prediction performance (with breakdown by store, region, and product type), data quality, and pipeline execution health, with automated alerting on degradation.
-
UK extension: following successful deployment and validation across North American locations, the pipeline was extended to UK stores. Data from the UK operation is normalised to work within the unified model, with regional data processed within the EU AWS region to satisfy data sovereignty requirements before predictions are routed back to UK systems.
Outcomes
- Accurate weekly price recommendations across more than 30,000 products in over 130 locations, with median prediction error of approximately 6% against a target of within 10%.
- Approximately 90% of the retailer's sales volume is now covered by model predictions, up from 38% of the product catalogue under the previous approach. Product categories where the model cannot meet the accuracy threshold are automatically excluded from recommendations.
- End-to-end automation: raw purchase data moves to updated price recommendations delivered into operational systems on a weekly schedule, without manual steps.
- A single unified model serving the US, Canada, and UK, with UK data processed within the EU AWS region and credentials managed in AWS Secrets Manager throughout.
Similar applications
The core techniques in this engagement apply wherever time series prediction must work reliably across products, locations, or entities with uneven data.
Demand and inventory forecasting
Products or SKUs where many items have thin or irregular purchase history in any given location, but shared patterns across the wider dataset can support reliable predictions.
Commodity-driven cost and price modelling
Businesses whose costs or prices are tied to volatile external inputs such as raw material prices, energy costs, or currency movements, and need those signals reflected in regularly updated operational outputs.
Prediction across distributed networks
Multi-site retailers, franchise networks, or distributed service businesses that need consistent, automated recommendations across many locations without manual processes at each site.
Unified modelling for thin-segment data
Any domain where separate models per product, customer, or geography fail because individual segments lack sufficient data, but a unified model trained across the full dataset can generalise where the segmented ones cannot.
Production ML for recurring decisions
Prediction tasks that feed a recurring business decision and need regular retraining as conditions shift, with outputs delivered automatically into operational systems.
Multi-region deployments with data localisation
ML systems spanning multiple jurisdictions where data sovereignty requirements mean regional data must be processed within specific regions before contributing to a shared model.
Discuss a similar engagement
Book a free 30-minute call. We will answer questions directly and say no when something is not a fit.