ML Demand Forecasting Eliminating $4M in Annual Stockouts
A national retail group with 340 stores was losing $4M a year to stockouts. Their spreadsheet-based demand planning couldn't factor in weather, trends, or competitor signals. We built ML forecasting models trained on five years of sales data plus external signals. Forecast accuracy went from 61% to 94%. Predictions now feed directly into ERP replenishment workflows.
The retailer runs 340 stores across 8 regions with 15,000 SKUs across apparel, homewares, and electronics. A central team of 6 planners managed demand forecasting with Excel models. Those models worked well enough for stable products. But they consistently missed on seasonal items, new launches, and anything sensitive to external factors like weather.
Stockouts clustered in three categories. Seasonal products saw demand spikes that the models consistently underestimated. New launches had no historical data to lean on. Regionally variable products got drowned out by national averages. The planning team knew the models weren't good enough. They just didn't have the tools to fix them.
61% forecast accuracy on seasonal products — below the industry benchmark of 75%
$4M annual revenue loss from stockouts, concentrated in 12% of SKUs
Planning team spending 70% of time managing exceptions rather than improving forecasts
No incorporation of external signals: weather, social trends, competitor pricing
Replenishment orders generated manually from spreadsheets — no automation
We built a three-model forecasting system. A base ARIMA model handles stable SKUs. A gradient boosting model handles volatile SKUs and incorporates external signals. A cold-start model uses category analogs for new product launches. All three feed an ensemble layer that picks the best forecast per SKU. Predictions go straight to the client's SAP ERP via API, triggering automated replenishment orders.
5-year historical sales data audit and cleaning — identified and corrected systematic data quality issues in 3 regional data warehouses
External signal integration: weather API, Google Trends, competitor pricing scraper, regional event calendar
Three-model architecture: ARIMA (stable), gradient boosting (volatile + external signals), cold-start analog model (new SKUs)
Ensemble layer that weights model outputs based on SKU category, velocity, and recency
SAP ERP integration — automated replenishment order generation from forecast outputs
Planner dashboard showing forecast vs actuals, model confidence, and override interface for exceptional items
Twelve months after deployment, forecast accuracy improved from 61% to 94% across the covered SKUs. Stockout revenue loss dropped by $4M. The planning team's exception management time fell from 70% to 25% of capacity. They now focus on strategic range planning instead of chasing errors.
Forecast accuracy: 61% → 94% across 15,000 SKUs
$4M annual stockout revenue loss eliminated
Planner exception time: 70% → 25% of capacity
New product launch accuracy improved 2.8× through cold-start analog modeling
Automated replenishment covering 87% of SKUs — manual intervention only for strategic items
“Our supply chain is now genuinely predictive. We stopped chasing stockouts and started preventing them. The ROI was clear within the first quarter.”
Rebecca Thorn
VP Supply Chain, National Retail Group
Client identity is withheld under NDA. The figures reported here were verified against the client's own internal reporting at project close.
External signals gave the biggest accuracy boost on seasonal products. But that only worked after fixing the underlying data quality issues.
The cold-start model for new products turned out to be more valuable than anyone expected. New launch stockouts were the most painful and visible problem.
Giving planners an override interface with a clear rationale for each forecast accelerated adoption significantly.
Key Results
Technology Stack
Facing a Similar Challenge?
Let's discuss your specific context and what results are realistic for your organisation.
Get in touch