Skip to content
E-CommerceDemand ForecastingCASE_10

Replacing a black-box vendor forecast with an owned, explainable system

A regional UK grocery retailer (engagement 2024, ~340 stores) had been paying ~$600k/year for a third-party demand-forecasting vendor whose forecast accuracy had degraded ~18% over two years. The vendor attributed it to "market volatility" and provided no model explainability. Over-forecasting was causing ~$2.1M/year in waste from perishable stock. A LightGBM ensemble trained on 4 years of sales history, enriched with weather, local events, promotional calendar, and competitor pricing, plus SHAP-based explainability on every forecast so the commercial team could interrogate the model rather than treat it as a black box, cut MAPE 31% vs. the prior vendor. Annual waste reduced ~$2.1M. Vendor contract terminated; the client's in-house data science team now owns, retrains, and monitors the model independently.

−31%
Forecast error (MAPE)
100%
Team owns the model
−$2.1M
Annual waste reduction

The Challenge

The retailer had been paying $600K/year for a third-party demand forecasting vendor. The forecast accuracy had degraded 18% over 2 years. The vendor attributed it to "market volatility", but provided no model explainability and no way to verify. Over-forecasting was causing $2.1M/year in waste from perishable stock. The internal data science team wanted to bring forecasting in-house but didn't have the architecture expertise to design and validate a production forecasting system.

Our Approach

Discovery & Blueprint scoped a gradient-boosting ensemble (LightGBM) trained on 4 years of sales history, enriched with external features: weather, local events, promotional calendar, and competitor pricing signals. Crucially, we used SHAP values for explainability, every forecast shows which features drove it. This let the commercial team interrogate and trust the model rather than treating it as a black box.

Outcome

Forecast MAPE reduced 31% vs. the previous vendor. Annual waste from over-forecasting reduced by $2.1M. The internal data science team now owns, retrains, and monitors the model independently. Vendor contract terminated.

What We Learned

01

Black-box vendor models create dependency, explainability is a strategic requirement, not a nice-to-have.

02

External signals (weather, events, promotions) consistently outperform internal-data-only forecasts for grocery.

03

The goal of a good handoff is that you're no longer needed.

Stages Engaged
AI Readiness Audit
Discovery & Blueprint
Concept Validation
Production Build
Total Duration
6 months total
Artifacts Delivered
PRD
Forecasting Architecture
Feature Engineering Spec
WBS
Model Monitoring Runbook
Start with a Feasibility Call

2 hours. No cost. We'll tell you honestly whether AI makes sense for your case.

Book a call