← Back to use cases
Coolblue improves inventory-demand forecasting with AI-generated models
Added 2026-10-14 · Reviewed 2026-10-15 · Coolblue
Read the source ↓What the company did
Google Cloud publishes Coolblue’s account of using AlphaEvolve to improve its 28-day demand-forecasting pipeline. The AI proposes and evaluates feature engineering, target preprocessing and regression-model combinations, with an evaluation metric covering both the first seven days and the full horizon while penalizing underforecasting.
How it went
Coolblue’s data scientist reports more than a 5% reduction in forecast error (WMAPE) against the existing production forecast after about 200 iterations. This is a customer-reported forecast improvement, not an independently reproduced result or a quantified reduction in stockouts.
Worth knowing
- The resulting forecast supports bulk purchasing and stock planning; the metric’s penalty reflects the cost of insufficient availability.
- Practical takeaway: Compare errors by SKU and horizon, including underforecasts, before changing purchases. Check whether aggregate gains hide shortages on important products.
Sources & provenance
Solve harder problems with AlphaEvolve, now available to everyone on Google Cloud ↗
Google Cloud · Published 2026-07-10 · Reviewed 2026-10-15
“optimize their 28-day demand forecasting pipeline”
Supporting evidence and review
Human review required
- Independent review advice: compare forecast errors and stockouts by SKU and horizon before applying purchasing changes.
- Keep purchasing approval and service-level constraints explicit.
Evidence & limitations
- Company/provider-reported deployment; no independent performance evaluation by Freight Fieldnotes.
- The source combines a product release with named customer production evidence.
- Reported forecast improvements are customer claims, not independently reproduced here.