An ML-powered demand forecasting engine that achieved 91% prediction accuracy, eliminating overstock issues and automating purchase orders for a national retail chain.
The Challenge
PredictEdge's buying team used Excel spreadsheets and gut instinct to forecast demand across 45,000 SKUs and 180 stores. Overstock sat at 20%, tying up $14M in working capital, while stockout rates of 12% cost an estimated $8M in lost sales annually. Seasonal planning was particularly problematic, with the team consistently over-ordering for holidays and under-ordering for unexpected trends. The 8-person buying team spent 80% of their time on manual forecasting rather than strategic vendor management.
- 20% overstock rate tying up $14M in working capital across warehouses
- 12% stockout rate resulting in $8M in estimated lost annual sales
- Manual Excel-based forecasting consuming 80% of the buying team's time
- Seasonal demand prediction accuracy of only 62%, leading to costly markdowns
- No ability to react to real-time demand signals like weather, trends, or competitor activity
Our Solution
Advenno built an ML-powered demand forecasting platform that predicts demand at the SKU-store-day level using an ensemble of models incorporating historical sales, seasonality, weather, local events, promotional calendars, and competitor pricing. The system automatically generates optimized purchase orders and integrates with existing ERP and warehouse management systems.
- Ensemble ML models combining gradient boosting, LSTM networks, and prophet for SKU-level daily forecasting
- 91% forecast accuracy incorporating weather, events, promotions, and trend signals
- Automated purchase order generation with vendor lead time and MOQ optimization
- Real-time demand sensing that adjusts forecasts based on current-week sales velocity
- Executive dashboard with inventory health scoring, markdown risk alerts, and buying recommendations
Our Approach
Data Collection & Cleansing
Aggregated 4 years of sales data across all stores and SKUs, cleaned anomalies from COVID disruptions and system outages, and enriched with external datasets including weather, events, and economic indicators.
Feature Engineering
Developed 200+ predictive features including lag variables, rolling statistics, holiday indicators, weather interactions, cannibalization effects, and promotional lift factors.
Model Development & Ensemble
Trained and evaluated multiple model architectures, ultimately building an ensemble that combines gradient boosting for stable items, LSTM for trending items, and Prophet for highly seasonal items.
ERP & WMS Integration
Built bi-directional integrations with SAP ERP and Manhattan WMS to ingest real-time inventory positions and push automated purchase orders directly to vendor portals.
Buying Team Training & Adoption
Trained the buying team on interpreting forecasts, managing exceptions, and using the override system, running the AI alongside manual forecasts for 6 weeks to build confidence.
The Results
Forecast accuracy improved from 62% to 91% within 3 months. Overstock dropped from 20% to 7%, freeing up $9.1M in working capital. Stockout rates fell from 12% to 3.8%, recovering an estimated $5.6M in previously lost sales. The buying team redirected 80% of their time from spreadsheets to strategic vendor negotiations and assortment planning. Markdown losses decreased by 44% as the system prevented over-ordering of slow-moving items.
Technologies Used
Our buyers used to spend all day in spreadsheets trying to predict demand. Now the AI does it better than any human could, and our team focuses on building vendor relationships and curating assortments. It's a complete transformation.